File size: 145,533 Bytes
6fa4bc9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973 2974 2975 2976 2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014 3015 3016 3017 3018 3019 3020 3021 3022 3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049 3050 3051 3052 3053 3054 3055 3056 3057 3058 3059 3060 3061 3062 3063 3064 3065 3066 3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103 3104 3105 3106 3107 3108 3109 3110 3111 3112 3113 3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128 3129 3130 3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159 3160 3161 3162 3163 3164 3165 | {
"paper_id": "2022",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T01:11:46.022556Z"
},
"title": "A Holistic Assessment of the Carbon Footprint of Noor, a Very Large Arabic Language Model",
"authors": [
{
"first": "Imad",
"middle": [],
"last": "Lakim",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TII",
"location": {
"settlement": "Abu Dhabi"
}
},
"email": "imad.lakim@tii.ae"
},
{
"first": "Ebtesam",
"middle": [],
"last": "Almazrouei",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TII",
"location": {
"settlement": "Abu Dhabi"
}
},
"email": "ebtesam.almazrouei@tii.ae"
},
{
"first": "Ibrahim",
"middle": [
"Abu"
],
"last": "Alhaol",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TII",
"location": {
"settlement": "Abu Dhabi"
}
},
"email": "ibrahim.abualhaol@tii.ae"
},
{
"first": "Merouane",
"middle": [],
"last": "Debbah",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "TII",
"location": {
"settlement": "Abu Dhabi"
}
},
"email": "merouane.debbah@tii.ae"
},
{
"first": "Julien",
"middle": [],
"last": "Launay",
"suffix": "",
"affiliation": {},
"email": "julien@lighton.io"
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "As ever larger language models grow more ubiquitous, it is crucial to consider their environmental impact. Characterised by extreme size and resource use, recent generations of models have been criticised for their voracious appetite for compute, and thus significant carbon footprint. Although reporting of carbon impact has grown more common in machine learning papers, this reporting is usually limited to compute resources used strictly for training. In this work, we propose a holistic assessment of the footprint of an extremescale language model, Noor. Noor is an ongoing project aiming to develop the largest multi-task Arabic language models-with up to 13B parameters-leveraging zero-shot generalisation to enable a wide range of downstream tasks via natural language instructions. We assess the total carbon bill of the entire project: starting with data collection and storage costs, including research and development budgets, pretraining costs, future serving estimates, and other exogenous costs necessary for this international cooperation. Notably, we find that inference costs and exogenous factors can have a significant impact on total budget. Finally, we discuss pathways to reduce the carbon footprint of extreme-scale models.",
"pdf_parse": {
"paper_id": "2022",
"_pdf_hash": "",
"abstract": [
{
"text": "As ever larger language models grow more ubiquitous, it is crucial to consider their environmental impact. Characterised by extreme size and resource use, recent generations of models have been criticised for their voracious appetite for compute, and thus significant carbon footprint. Although reporting of carbon impact has grown more common in machine learning papers, this reporting is usually limited to compute resources used strictly for training. In this work, we propose a holistic assessment of the footprint of an extremescale language model, Noor. Noor is an ongoing project aiming to develop the largest multi-task Arabic language models-with up to 13B parameters-leveraging zero-shot generalisation to enable a wide range of downstream tasks via natural language instructions. We assess the total carbon bill of the entire project: starting with data collection and storage costs, including research and development budgets, pretraining costs, future serving estimates, and other exogenous costs necessary for this international cooperation. Notably, we find that inference costs and exogenous factors can have a significant impact on total budget. Finally, we discuss pathways to reduce the carbon footprint of extreme-scale models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Recent progress in natural language processing (NLP) has been driven by the emergence of socalled foundation models (Bommasani et al., 2021) . This paradigm shift is characterised by a homogenisation of modelling methods-crystallising around the Transformer architecture (Vaswani et al., 2017) and by emergent capabilities (e.g. zero-shot generalisation) predominantly arising from sheer scale alone . NLP models are now experiencing a 3-4 months doubling time in size, as outlined by Figure 1 . Most recent large language models such as MT-NLG 530B (Smith et al., 2022 ), Gopher 280B (Rae et al., 2021 , or Jurassic-1 178B (Lieber et al., 2021) , all report training budgets in the thousands of PF-days 1 range. Because AI accelerators performance per watt has plateaued compared to deep learning budgets (Reuther et al., 2021; Sevilla et al., 2022) , practitioners have had to scale-out training over an increasingly large number of accelerators (Narayanan et al., 2021) . Accordingly, the energy cost of training state-of-theart models has grown significantly: increase in compute is no longer fuelled by improvements in hardware efficiency, but in hardware scale.",
"cite_spans": [
{
"start": 116,
"end": 140,
"text": "(Bommasani et al., 2021)",
"ref_id": "BIBREF6"
},
{
"start": 271,
"end": 293,
"text": "(Vaswani et al., 2017)",
"ref_id": "BIBREF39"
},
{
"start": 550,
"end": 569,
"text": "(Smith et al., 2022",
"ref_id": "BIBREF33"
},
{
"start": 570,
"end": 602,
"text": "), Gopher 280B (Rae et al., 2021",
"ref_id": null
},
{
"start": 624,
"end": 645,
"text": "(Lieber et al., 2021)",
"ref_id": "BIBREF17"
},
{
"start": 806,
"end": 828,
"text": "(Reuther et al., 2021;",
"ref_id": "BIBREF28"
},
{
"start": 829,
"end": 850,
"text": "Sevilla et al., 2022)",
"ref_id": null
},
{
"start": 948,
"end": 972,
"text": "(Narayanan et al., 2021)",
"ref_id": "BIBREF21"
}
],
"ref_spans": [
{
"start": 485,
"end": 493,
"text": "Figure 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Although this increase in size and compute budget is backed by empirical scaling laws drawing a clear link between compute spent and model performance , the societal benefits of larger models have been questioned (Toma\u0161ev et al., 2020; Bender et al., 2021) . Specifically to environmental concerns, in a time of climate crisis when carbon emissions must be drastically cut (Masson-Delmotte et al., 2018) , one may question whether these large compute budgets are justified. A crucial step towards answering this question is an in-depth evaluation of the footprint of large models.",
"cite_spans": [
{
"start": 213,
"end": 235,
"text": "(Toma\u0161ev et al., 2020;",
"ref_id": null
},
{
"start": 236,
"end": 256,
"text": "Bender et al., 2021)",
"ref_id": "BIBREF5"
},
{
"start": 373,
"end": 403,
"text": "(Masson-Delmotte et al., 2018)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Existing assessments of the environmental impacts of large models are usually focused on hyperparameter tuning and pretraining costs (Strubell et al., 2019; Patterson et al., 2021) . This trend is reflected by the growing number of tools available to help practitioners quantify the impact of machine learning computations (Bannour et al., 2021) . If some studies have also endeavoured to quantify select aspects of the machine learning pipeline (e.g. conference attendance (Skiles et al., 2021) , hardware lifecycle (Gupta et al., 2021) , etc.), end-to-end evaluations of machine learning projects life cycle emissions remain rare (Wu et al., 2022) . : Over the last four years, the size of stateof-the-art language models has doubled every 3-4 months. Note that this trend has been slowing down, due to scale-out limitations.",
"cite_spans": [
{
"start": 133,
"end": 156,
"text": "(Strubell et al., 2019;",
"ref_id": "BIBREF34"
},
{
"start": 157,
"end": 180,
"text": "Patterson et al., 2021)",
"ref_id": "BIBREF23"
},
{
"start": 323,
"end": 345,
"text": "(Bannour et al., 2021)",
"ref_id": "BIBREF4"
},
{
"start": 474,
"end": 495,
"text": "(Skiles et al., 2021)",
"ref_id": "BIBREF32"
},
{
"start": 517,
"end": 537,
"text": "(Gupta et al., 2021)",
"ref_id": "BIBREF12"
},
{
"start": 632,
"end": 649,
"text": "(Wu et al., 2022)",
"ref_id": null
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "To fill this gap, we produce an end-to-end assessment of the carbon footprint of Noor, a project seeking to train a very large Arabic language model. Our contributions are the following:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Holistic assessment. We evaluate the total carbon bill of the entire project: starting with data collection, curation, and storage, including research and development and hyper-parameters tuning budgets, pretraining costs, future serving estimates, and other exogenous impacts sparked by this international cooperation (e.g. flights, personnel, etc.)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Beyond pretraining. We identify pretraining compute as driving more than half of the emissions of the project. However, all combined, other R&D, storage, and personnel counts still amount for 35% of the carbon footprint. We also identify downstream use in the wild as potentially significant. This leads us to recommend for the end-to-end footprint to be systematically assessed on a per-project basis. Notably, in scenarios with a low-impact training electric mix, costs beyond pretraining may become the main sources of emissions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "Pathways to lower footprints. Finally, we discuss ways to reduce the environmental footprints of projects involving large models, and put in perspective the footprint of similar projects.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Introduction",
"sec_num": "1"
},
{
"text": "In light of ever increasing computational budgets (Sevilla et al., 2022) and of the need to cut on emissions to abate global warming (Masson-Delmotte et al., 2018) , the environmental impact of deep learning has drawn significant interest. Strubell et al., 2019 notably highlighted the potential high environmental costs of deep learning. However, its headline figures were produced in the specific context of neural architecture search, a relatively rare practice for extreme-scale models nowadays. Lacoste et al., 2019; Lottick et al., 2019; Schwartz et al., 2020 subsequently called for AI research to be more aware of its environmental cost. An increasing number of tools, such as codecarbon (Schmidt et al., 2021) , have been developed to help with tracking the impact of deep learning experiments (Bannour et al., 2021) . All of these lines of research share similar recommendations: the carbon footprint of deep learning is a direct consequence of the electricity mix and efficiency of the data center, suggesting that picking an appropriate provider is the most straightforward way to reduce environmental impact.",
"cite_spans": [
{
"start": 50,
"end": 72,
"text": "(Sevilla et al., 2022)",
"ref_id": null
},
{
"start": 133,
"end": 163,
"text": "(Masson-Delmotte et al., 2018)",
"ref_id": null
},
{
"start": 240,
"end": 261,
"text": "Strubell et al., 2019",
"ref_id": "BIBREF34"
},
{
"start": 500,
"end": 521,
"text": "Lacoste et al., 2019;",
"ref_id": "BIBREF16"
},
{
"start": 522,
"end": 543,
"text": "Lottick et al., 2019;",
"ref_id": "BIBREF18"
},
{
"start": 544,
"end": 565,
"text": "Schwartz et al., 2020",
"ref_id": null
},
{
"start": 696,
"end": 718,
"text": "(Schmidt et al., 2021)",
"ref_id": null
},
{
"start": 803,
"end": 825,
"text": "(Bannour et al., 2021)",
"ref_id": "BIBREF4"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "Specifically to extreme-scale models, Patterson et al., 2021 estimated the energy consumption of five large NLP models, including GPT-3. They identified that a judicious choice of neural architecture, datacenter and accelerator can help reduce considerably carbon budgets. Thompson et al., 2020 identified a clear relationship between large models performance and their carbon impact, building upon work on neural scaling laws . Taddeo et al., 2021 estimated the cost of training GPT-3 in different data centers across the worldwide, highlighting again the high dependency on the local energy mix and specific infrastructure.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "Two recent studies have provided insights into the end-to-end carbon footprint of deployed models in the industry. Wu et al., 2022 studied the impact of the increasingly large recommender systems leveraged at Meta, while Patterson et al., 2022 provided an assessment of the costs (including inference) of large models at Google. They expect the carbon footprint of training to plateau in coming years, and then to shrink-owing to more efficient high performance computing platforms. They also assert that current studies are overestimating the real environmental costs of large models, in light of the wide availability of \"clean\" compute platforms.",
"cite_spans": [
{
"start": 221,
"end": 243,
"text": "Patterson et al., 2022",
"ref_id": "BIBREF22"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "In the field of astrophysics, Aujoux et al., 2021 did an extensive study to estimate the carbon footprint of the Giant Array for Neutrino Detection (GRAND) project, a multi-decade worldwide experiment. Inspired by their holistic methodology, we seek to establish the first end-to-end assessment of an extreme-scale NLP project.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Related work",
"sec_num": "2"
},
{
"text": "The current state-of-the-art generative language model in Modern Standard Arabic is AraGPT (Antoun et al., 2021), a 1.5B parameters model. The Noor project seeks to expand upon this model, introducing a 1.5B, 2.7B, 6.7B, and 13B Arabic models, trained a custom curated dataset of 150B tokens, inspired by The Pile (Gao et al., 2020) . These larger scales are expected to make the model able to tackle novel tasks through zero-shot generalization, as exhibited by GPT-3 (Brown et al., 2020) or GPT-J (Wang and Komatsuzaki, 2021) .",
"cite_spans": [
{
"start": 314,
"end": 332,
"text": "(Gao et al., 2020)",
"ref_id": "BIBREF11"
},
{
"start": 463,
"end": 489,
"text": "GPT-3 (Brown et al., 2020)",
"ref_id": null
},
{
"start": 499,
"end": 527,
"text": "(Wang and Komatsuzaki, 2021)",
"ref_id": "BIBREF40"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Noor Project",
"sec_num": "3"
},
{
"text": "Noor is an on-going international cooperation between the Technology Innovation Institute in the United Arab Emirates and LightOn in France. The Noor project can be split in four parts:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Noor Project",
"sec_num": "3"
},
{
"text": "\u2022 Data curation. A custom curated dataset of 150B tokens has been assembled for Noor. This dataset has been scrapped from diversified sources, and also includes data from Common Crawl. We filter this data with an LM-based quality-scoring system inspired by CCNet (Wenzek et al., 2019) .",
"cite_spans": [
{
"start": 263,
"end": 284,
"text": "(Wenzek et al., 2019)",
"ref_id": "BIBREF41"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "The Noor Project",
"sec_num": "3"
},
{
"text": "\u2022 R&D experiments. To validate tokenization, dataset, architecture, and establish scaling laws, we trained a number of R&D models (100M-1.5B parameters on 10-30B tokens).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Noor Project",
"sec_num": "3"
},
{
"text": "\u2022 Main training. We train a suite of four models of 1.5B, 2.7B, 6.7B, and 13B parameters.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Noor Project",
"sec_num": "3"
},
{
"text": "\u2022 Model use. Prospectively, we include some estimations of the future inference cost of these models as they are put in use.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "The Noor Project",
"sec_num": "3"
},
{
"text": "Before beginning our assessment, we propose to identify some of the key influencing factors on the potential carbon footprint of large models, focusing first on factors directly related to the models themselves and not to the project producing them.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "Model size. The number of floating operations per forward pass is directly proportional to the size of the network. A common approximation for the total compute budget C required for training a Transformer model with N parameters on D tokens is C = 6N D . As the optimal dataset size only grows sublinearly with model size for autoregressive modelling , compute budget will scale more or less linearly with model size. The larger the number of operations, the more energy is needed to train the model. For inference, the cost for each token is reduced to a third compared to training, and environmental impact will be driven by the total number of words/tokens processed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "Hardware characteristics. The throughput (in FLOPs) that can be tackled by the hardware will drive the total time required to perform the task. More efficient hardware will have more throughput per Watt. We note however that most available chips suitable for large model training (e.g., NVIDIA GPUs, Google TPUs, etc.) exhibit similar efficiency characteristics (Reuther et al., 2021) .",
"cite_spans": [
{
"start": 362,
"end": 384,
"text": "(Reuther et al., 2021)",
"ref_id": "BIBREF28"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "Modelling decisions. We identified above two key factors: number of tokens processed (for training or inference), and hardware throughput. We note that both of these are also strongly impacted by modelling decisions. A more fertile tokenizer will use less tokens for the same text, leading to faster processing. Similarly, small changes in model architecture (e.g., choosing hidden sizes in accordance with wave/tile quantization) and in implementation (e.g., 3D parallelism) can drastically increase throughput, and reduce total training time.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "Data center efficiency. The energy consumed does not serve only to power up the servers, but also to cool down the data center itself and to respond to other electrical needs. The Power Usage Effectiveness (PUE) is used to assess the overall efficiency of a data center. It measures the quotient of the total energy requirement and the final energy used by the servers. The PUE will be influenced by the data center architecture. Worldwide average is around 1.8, but Google for instance reports an average PUE of 1.11. Waste heat in data centers can also be reused for collective water heating, driving down the PUE, as in the Jean Zay HPC.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "Electricity mix. The breakdown of the energy sources powering a data center is a crucial factor, and depends primarily on the region. The electricity mix determines the carbon emissions per kWh of electricity. Today, the world average of carbon emission by kwh of electricity generated is 475 gCO2e/kWh, and an increasing number of data centers from cloud providers are using 100% renewable or nuclear energy to power their hardware. Taking Google Cloud as an example again, their",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "Montreal facility reports 27gCO2e/kWH, twenty times lower than the world average. Beyond factors related to the models themselves, we seek in this study to take into account a number of other costs: storage, preprocessing, and transfer costs for the dataset, personnel costs such as travel and individual laptops, etc. We note however one limitation from our study: we do not take into account the lifecycle of the hardware used. Unfortunately, numbers are scarcely available, and not made public by the main manufacturers.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "5 Carbon footprint of the Noor project",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Factors influencing the carbon footprint of large models",
"sec_num": "4"
},
{
"text": "We begin by accounting for the electricity consumption of all aspects of the project. The impact of this consumption will be highly dependent on the carbon intensity of the electricity mix used. Nonelectric sources (e.g., international flights) will be added to the carbon budget in a second phase.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Electricity consumption",
"sec_num": "5.1"
},
{
"text": "The energy consumption of data depends on both the energy required for powering the disks to store the data, and the energy consumed when moving the data from one server to another. We average storage costs over the 6 months of the project.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "Storage. Although disk wattage is generally reported on per-disk level, Posani et al., 2019 estimates the power per TB of data using aggregated technical specifications. The paper reports that the average peak consumption of cloud storage is around 11.3W/TB. It means an energy consumption of 99 kWh/TB a year. This estimation considers a PUE of 1.6 and a redundancy factor of 2 since managed services will also have a back-up.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "The breakdown of our data storage is as follows:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "\u2022 Curated data. Including both raw and processed data, we have accumulated around 2TB of curated data. This is stored for the 6 months of the project, resulting in 99kWh used.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "\u2022 Bulk data. We use Common Crawl (CC) for acquiring large amounts of web data. Each CC dump is on average around 10TB, and we discard it immediately after processing it. On average, it takes 24 hours to fully process a dump: we used 21 dumps from CC, meaning we stored 210TB of data for 24hours, equivalent to 57 kWh of energy consumption. After processing the dumps, we got on average 1.2TB of data per dump, thus 25TB in total. Considering that this data will be stored for 6 months, we end up with 1.3 MWh of energy consumption for the bulk data. Note that we keep the processed data in all languages (not just Modern Standard Arabic).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "\u2022 Models. The weights of the Noor models (1.3B, 2.7B, 6.7B and 13B) are respectively 2.6GB, 5.4G, 13.4GB, and 26GB in halfprecision. This corresponds to training checkpoints (including the full-precision optimizer) of 20.8GB, 43.2GB, 107.2GB, and 208GB. We save such checkpoints every 10B tokens.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "In total, we end-up with 5.7TB of model weights and intermediary checkpoints for future analysis and interpretability work, consuming 0.3MWh in total.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "Transfers. Posani et al., 2019 provided an estimate of 23.9 kJ per GB (6.38 kWh per TB) transferred, using the formula of Baliga et al., 2011 and the same hypothesis as Aslan et al., 2017 (800km average distance between core nodes). The 210TB of CC data are downloaded on the preprocessing servers once; the 25TB of processed data are moved once to our archival machines, and another time to the HPC used for training; the curated data is downloaded once, moved to the archival machines, and then moved to the HPC; the 5.7TB of models are moved once from our HPC, and then to our inference servers for final models or to workstations for intermediary checkpoints. Consequently, we estimate the transfer energy bill at 1.8 MWh.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "Total. Thus, the total energy consumption of data is estimated to be about 3.5 MWh, dominated by the multilingual Common Crawl data. We note that as ideal dataset size increases sublinearly with model size (Kaplan et al., 2020), we expect checkpoints and model transfers to eventually dominate the costs of storage and transfer for larger models. Note that we neglect costs linked to a potential public release of the models, as it is difficult to predict traffic. As a rough estimation, 10,000 downloads of the 13B model would represent 260TB of traffic, and 1.66MWh consumed.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data storage and transfers",
"sec_num": "5.1.1"
},
{
"text": "We take all text data through a pipeline inspired by CCNet (Wenzek et al., 2019) for preprocessing. This pipeline takes care of deduplication, language identification, and finally quality filtering with a ures , we estimate the average power consumption of each node at 350W; hence, the power of the cluster is 5.6kW. We processed 21 dumps of Com-monCrawl, plus our curated data, for a total of 381 wall-clock hours. Accordingly, assuming a PUE of 1.1 as reported by Google, the total energy consumed by data preprocessing is 2.35MWh.",
"cite_spans": [
{
"start": 59,
"end": 80,
"text": "(Wenzek et al., 2019)",
"ref_id": "BIBREF41"
}
],
"ref_spans": [
{
"start": 205,
"end": 209,
"text": "ures",
"ref_id": null
}
],
"eq_spans": [],
"section": "Data processing",
"sec_num": "5.1.2"
},
{
"text": "Note that for CommonCrawl data, this results in data processed for every language supported (176 for identification, 48 for quality filtering). Accordingly, this cost could be amortised over future projects. For high-resource languages, this also results in very large amounts of data: processing more dumps would not be necessary, even to train a 1 trillion parameters model.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Data processing",
"sec_num": "5.1.2"
},
{
"text": "We carried experiments to validate tokenization methods, dataset composition, tune hyperparameters, and establish scaling laws. This early research and development work was performed on MeluXina, a high-performance super-computer located in Luxembourg. We used a total of 16,800 A100hours in this phase. Each node used in MeluXina has 4 A100 SXM 40GB with a TDP of 400W, and two AMD EPYC 7763 CPUs with a TDP of 280W. They report a PUE of 1.35. Thus, we estimate the consumption of this R&D phase to be of 10.7MWh.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Research and development",
"sec_num": "5.1.3"
},
{
"text": "We expect the budget of this phase to roughly scale with model size. Indeed, debugging potential issues (e.g., numerical instabilities , etc.) for the final larger model will cost significantly more.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Research and development",
"sec_num": "5.1.3"
},
{
"text": "Using the C = 6N D approximation, it is possible to calculate in advance the training budget required for a specific model. We observe an ef-fective throughput with our Megatron+DeepSpeed codebase of around 100 TFLOPs 2 across models, in line with the state-of-the-art. We train four main models (1.5B, 2.7B, 6.7B, 13B) on 150B tokens.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Main training",
"sec_num": "5.1.4"
},
{
"text": "We train the smaller model on MeluXina, but the other three on our own HPC cluster. Each node contains 8 A100 80GB and 2 AMD EPYC 7763 CPUs. The PUE of our data center is 1.5, 20% more efficient than the world average. Table 1 outlines the costs of the main training. The total electric energy consumed to train the Noor suite of models is thus 41.6 MWh, 55% of it spent on the largest 13B model.",
"cite_spans": [],
"ref_spans": [
{
"start": 219,
"end": 226,
"text": "Table 1",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "Main training",
"sec_num": "5.1.4"
},
{
"text": "As the models of Noor have yet to be deployed, this is only a prospective estimate. Inference costs in general are difficult to estimate in advance, even more so for open source models which will be deployed to platforms with varying characteristics. We provide an estimate of the energy consumption during inference per generated token.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference",
"sec_num": "5.1.5"
},
{
"text": "We thereafter denote as processed tokens the tokens in the original prompt sent to the model, and as generated tokens the tokens generated by the model using the prompt. To simplify calculations, we make the following assumptions from our experience with another large-scale API: (1) an A100 is used, which is sufficient for Noor-13B, but could be reduced to a more efficient T4 for Noor-1.5B/2.7B;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference",
"sec_num": "5.1.5"
},
{
"text": "(2) inference time per generated token is constant, whichever the number of processed tokens (per our benchmarks, thanks to caching, this is true up to 512 processed tokens roughly); (3) batch size is assumed to be 1, as batching is more challenging and less consistent for inference workloads.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference",
"sec_num": "5.1.5"
},
{
"text": "Under these hypothesises, an A100 can generate up to 72,000 tokens per hour.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Inference",
"sec_num": "5.1.5"
},
{
"text": "Accordingly, we estimate that 26 Joules are required per token generated (400W Figure 2 : Breakdown of the electricity consumption (total 59.14 MWh) of the Noor project. Data preprocessing is included in R&D, amounting for 20% of it. We also note that R&D and dataset costs could be amortised through other projects or larger models.",
"cite_spans": [],
"ref_spans": [
{
"start": 79,
"end": 87,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Inference",
"sec_num": "5.1.5"
},
{
"text": "for the GPU, 70W for the CPU, and 1.1 PUE on Google Cloud imply 517Wh of energy consumption for 72,000 tokens. Converted to Joule, it results in 26 Joules per token.) Accordingly, 3 billion tokens would have to be generated for inference costs to catch up with training costs. At some point during its beta, GPT-3 was reported to generate 4.5 billion words per day (Pilipiszyn, 2021) .",
"cite_spans": [
{
"start": 365,
"end": 383,
"text": "(Pilipiszyn, 2021)",
"ref_id": "BIBREF24"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Inference",
"sec_num": "5.1.5"
},
{
"text": "Beyond costs related to data, R&D, training, and inference, one may wonder if direct electricity use from scientists involved in the project is significant. Assuming that the average laptop consumes 70W, plus 30W for an external screen, six research scientists dedicating 100% of their time during 6 months for this project, 8 hours per day, will use up 0.604MWh. We could also include costs of e-mail exchanges and video-conferences specifically, but these were found to be negligible in Aujoux et al., 2021 . We round up the marginal costs to 1MWh, and note that this is but a rough estimate.",
"cite_spans": [
{
"start": 489,
"end": 508,
"text": "Aujoux et al., 2021",
"ref_id": "BIBREF2"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Additional costs",
"sec_num": "5.1.6"
},
{
"text": "We showed that the total electricity consumption of the Noor project is not only about training the final models, as outlined in Figure 2 . Nearly a third of the energy consumed (30%) went to tasks outside of main models pretraining.",
"cite_spans": [],
"ref_spans": [
{
"start": 129,
"end": 137,
"text": "Figure 2",
"ref_id": null
}
],
"eq_spans": [],
"section": "Summary",
"sec_num": "5.1.7"
},
{
"text": "Because of larger uncertainties, we keep the serv-ing/inference assessment out of the previous budget. However, especially in the context of openly available models, the inference budget can rapidly catch up with the total budget outlined in 2.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Summary",
"sec_num": "5.1.7"
},
{
"text": "Now, from the electricity consumption, and using information on the local carbon intensity, we will derive the full footprint of the Noor project. We will also add energy use coming from non-electric sources (e.g., flights). As the carbon intensity of the electricity mix varies significantly across regions, we outlined below the locations of interest:",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "\u2022 Storage. We used Amazon S3 in Bahrain;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "\u2022 R&D. We used a GCP CPU cluster located in Netherlands, and MeluXina in Luxembourg;",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "\u2022 Main training. The smaller 1.3B model was trained on MeluXina, and the remaining models were trained on our dedicated HPC platform in the United Arab Emirates (UAE);",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "\u2022 Other. Six full-time scientists were involved, half in France and half in the UAE. Table 2 shows the resulting carbon footprint for each of the development stages of Noor project. This highlights the importance of location for carbon footprint: notably, all calculations on performed on the relatively low-carbon MeluXina HPC end-up having very limited costs, even compared to small items like storage in Bahrain.",
"cite_spans": [],
"ref_spans": [
{
"start": 85,
"end": 92,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "In addition to these development costs, we consider the carbon footprint of three round-trip flights of four scientists between Paris and Abu Dhabi. These trips were taken to run training workshops, brainstorming sessions, and discussions related to the project. We use the carbon emissions simulator of the International Civil Aviation Organization. One round-trip emits 527 kgCO2e per person, totalling 6.4 tons of emissions over all trips.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "Finally, Figure 3 displays the total distribution of the carbon footprint of the project. As shown in the figure, factors like flights may be usually neglected, but have a significant contribution in the total carbon footprint. Specifically, as conference returns in-person, this is a systematic impact that exists on most papers. In the case of Noor, the few flights operated account for 18% of the total carbon emission of the whole project. Interestingly, we note that with increasingly clean electricity and efficient data centers, the exogenous costs linked to flights and personnel are bound to increase in proportional impact.",
"cite_spans": [],
"ref_spans": [
{
"start": 9,
"end": 17,
"text": "Figure 3",
"ref_id": "FIGREF1"
}
],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "Inference. Forecasting the carbon footprint of inference is harder for open models: as they may be downloaded and deployed by anyone, it is impossible to predict the carbon intensity of the electricity they will use. We study two scenarios: an intermediate one, based on the world average emission per kWh (475 gCO2e/kWh) and a best-case one, based on the low-impact French mix (56 gCO2e/kWh). These two scenarios correspond to around 300,000 tokens generated per kgCO2e, or to 2,500,000 tokens generated per kgCO2e in the best-case. Going back to the 4.5 billion words per day of GPT-3, this amounts to 30 tons of CO2e per day and 3.5 tons. ",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Carbon footprint",
"sec_num": "5.2"
},
{
"text": "From our experience with Noor, we highlight some recommendations for future projects to minimise their carbon footprint.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Best practices and recommendations",
"sec_num": "6"
},
{
"text": "A first angle of attack is to make the machine learning techniques used more efficient.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Modelling & engineering",
"sec_num": "6.1"
},
{
"text": "\u2022 Efficient architectures. Mixture-of-experts (MoE) models split the large fully-connected layers of a Transformer into distinct experts (Fedus et al., 2021) . Although larger, MoE Transformers can bring significant energy savings during training and inference (Du et al., 2021) , as the experts are only sparsely activated. Recent work demonstrate that they may even scale favorably compared to dense models (Clark et al., 2022) . More broadly, even small changes (e.g. better embeddings, activation functions) may have a non-negligible impact on the overall carbon footprint.",
"cite_spans": [
{
"start": 137,
"end": 157,
"text": "(Fedus et al., 2021)",
"ref_id": "BIBREF10"
},
{
"start": 261,
"end": 278,
"text": "(Du et al., 2021)",
"ref_id": null
},
{
"start": 409,
"end": 429,
"text": "(Clark et al., 2022)",
"ref_id": "BIBREF8"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modelling & engineering",
"sec_num": "6.1"
},
{
"text": "\u2022 Efficient inference. As we have shown, inference costs can rapidly catch up with training costs: it is also interesting to make the model leaner for inference. Quantization (Yang et al., 2019) reduces numerical precision at inference time and accelerates inference, but it has seen limited adoption with large models. Distillation (i.e., training a smaller model from the outputs of a larger one) is a promising direction, already demonstrated for Transformers applied to vision (Touvron et al., 2021) .",
"cite_spans": [
{
"start": 175,
"end": 194,
"text": "(Yang et al., 2019)",
"ref_id": "BIBREF43"
},
{
"start": 481,
"end": 503,
"text": "(Touvron et al., 2021)",
"ref_id": "BIBREF38"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Modelling & engineering",
"sec_num": "6.1"
},
{
"text": "\u2022 Efficient implementations. Crucially, distributed training implementations must be as efficient as possible, to amortise the large idle consumption of the hardware -MeluXina reports for instance idle power of around 150W per GPU when accounting for CPU cores, infrastructure, etc. This includes taking into account fine-grained effects depending on architectures, such as wave and tile quantization, to achieve the best throughput possible.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Modelling & engineering",
"sec_num": "6.1"
},
{
"text": "A second angle of attack is to focus on the hardware used to train these models.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Hardware",
"sec_num": "6.2"
},
{
"text": "\u2022 Data center choice. A data center with a PUE of 1.1 will decrease energy consumption by 39% compared to the world average of 1.8. Low PUE platforms should be preferred.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Hardware",
"sec_num": "6.2"
},
{
"text": "\u2022 Local carbon intensity. As highlighted by Table 2 , the carbon intensity of the electricity mix significantly impacts the final footprint. Locating training in an area with a clean mix is an easy step to take that can drastically cut the footprint of a project. This is especially easy to do on online cloud platforms, which have many areas of availability.",
"cite_spans": [],
"ref_spans": [
{
"start": 44,
"end": 51,
"text": "Table 2",
"ref_id": "TABREF1"
}
],
"eq_spans": [],
"section": "Hardware",
"sec_num": "6.2"
},
{
"text": "\u2022 Efficient inference. Carefully selecting a proper AI accelerator for managed inference workloads can limit the footprint of model use. For instance, for smaller models (<3B), it may be possible to use T4s rather than A100s, which are 20% more energy efficient per FLOP than A100s. Finally, specialised accelerators are also starting to become available (Reuther et al., 2020) . We note that this may however require specific developments.",
"cite_spans": [
{
"start": 355,
"end": 377,
"text": "(Reuther et al., 2020)",
"ref_id": "BIBREF27"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "Hardware",
"sec_num": "6.2"
},
{
"text": "Finally, it is important to not underestimate costs beyond machine learning workloads.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Other practices",
"sec_num": "6.3"
},
{
"text": "\u2022 Minimising exogenous impact. Although we found the final footprint to be dominated by the main training runs, we still note the significant impact of the international flights taken during this cooperation (20% of the final footprint). Minimising such high-intensity cost center is important.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Other practices",
"sec_num": "6.3"
},
{
"text": "\u2022 Costs reporting and offset. The full cost of model development is rarely, if ever, reported in the literature. We highly recommend the AI community to start reporting the full energy consumption and the CO2e of their projects. This reporting can also be used as the basis for offsetting carbon emissions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Other practices",
"sec_num": "6.3"
},
{
"text": "We undertook an end-to-end assessment of the carbon footprint associated with the development of an extreme-scale language model. We took into account data collection and storage, research and development, pretraining, and included estimates for future serving and inference. We also added personnel costs, such as international flights to run training workshops and brainstorming sessions. In total, we estimate the development of the suite of the four Noor models to have emitted 36.5 tons of CO2, 65% of which for training the models, 18% for the international flights, 12% for data storage, and 4% for small-scale research and development experiments. To put this in perspective, the average carbon footprint per individual in the US is around 20 tons, so our project generated a little over two years of individual US emissions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and conclusion",
"sec_num": "7"
},
{
"text": "We find that the main driver of this carbon footprint is the carbon intensity of the mix used for model training. Appropriately selecting the location of calculations can significantly reduce the environmental impact of a project. For instance, in this project, running all computations in France would have reduced the total footprint to 14.9 tCO2e, 42% of which from the international flights. As the impact of the computations themselves become smaller, it is important for practitionners to more carefully weigh in exogenous contributions.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and conclusion",
"sec_num": "7"
},
{
"text": "All-in-all, with careful considerations around data center choice, it is possible to run extremescale NLP projects with a low carbon impact.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and conclusion",
"sec_num": "7"
},
{
"text": "Finally, we also identified that large-scale inference could also rapidly outtake pretraining costs in terms of carbon impact. Inference, if not centrally managed, is harder to control: with a publicly available model, it will happen on hardware decided by the end user. We thus think its equally important for practitioners to alert users regarding best efficient inference practices, and regarding best practices to limit the environmental cost of computations (e.g. choosing an efficient data center, running inference in a country with a low-impact mix, etc.)",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Discussion and conclusion",
"sec_num": "7"
},
{
"text": "A PF-day is 1 PFLOPs (10 A100) sustained for a day.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
},
{
"text": "These are effective FLOPs for training the model, not hardware FLOPs. Hardware FLOPs are closer to 150 TFLOPs.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "",
"sec_num": null
}
],
"back_matter": [],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Aragpt2: Pre-trained transformer for arabic language generation",
"authors": [
{
"first": "Wissam",
"middle": [],
"last": "Antoun",
"suffix": ""
},
{
"first": "Fady",
"middle": [],
"last": "Baly",
"suffix": ""
},
{
"first": "Hazem",
"middle": [],
"last": "Hajj",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Wissam Antoun, Fady Baly, and Hazem Hajj. 2021. Aragpt2: Pre-trained transformer for arabic lan- guage generation.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Electricity intensity of internet data transmission: Untangling the estimates: Electricity intensity of data transmission",
"authors": [
{
"first": "Joshua",
"middle": [],
"last": "Aslan",
"suffix": ""
},
{
"first": "Kieren",
"middle": [],
"last": "Mayers",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "Koomey",
"suffix": ""
},
{
"first": "Chris",
"middle": [],
"last": "France",
"suffix": ""
}
],
"year": 2017,
"venue": "Journal of Industrial Ecology",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1111/jiec.12630"
]
},
"num": null,
"urls": [],
"raw_text": "Joshua Aslan, Kieren Mayers, Jonathan Koomey, and Chris France. 2017. Electricity intensity of internet data transmission: Untangling the estimates: Elec- tricity intensity of data transmission. Journal of In- dustrial Ecology, 22.",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "Estimating the carbon footprint of the grand project, a multi-decade astrophysics experiment",
"authors": [
{
"first": "Clarisse",
"middle": [],
"last": "Aujoux",
"suffix": ""
},
{
"first": "Kumiko",
"middle": [],
"last": "Kotera",
"suffix": ""
},
{
"first": "Odile",
"middle": [],
"last": "Blanchard",
"suffix": ""
}
],
"year": 2021,
"venue": "Astroparticle Physics",
"volume": "131",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.1016/j.astropartphys.2021.102587"
]
},
"num": null,
"urls": [],
"raw_text": "Clarisse Aujoux, Kumiko Kotera, and Odile Blanchard. 2021. Estimating the carbon footprint of the grand project, a multi-decade astrophysics experiment. As- troparticle Physics, 131:102587.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Green cloud computing: Balancing energy in processing, storage, and transport. Proceedings of the IEEE",
"authors": [
{
"first": "Jayant",
"middle": [],
"last": "Baliga",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Ayre",
"suffix": ""
},
{
"first": "Kerry",
"middle": [],
"last": "Hinton",
"suffix": ""
},
{
"first": "Rodney",
"middle": [],
"last": "Tucker",
"suffix": ""
}
],
"year": 2011,
"venue": "",
"volume": "99",
"issue": "",
"pages": "149--167",
"other_ids": {
"DOI": [
"10.1109/JPROC.2010.2060451"
]
},
"num": null,
"urls": [],
"raw_text": "Jayant Baliga, Robert Ayre, Kerry Hinton, and Rodney Tucker. 2011. Green cloud computing: Balancing energy in processing, storage, and transport. Pro- ceedings of the IEEE, 99:149 -167.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Evaluating the carbon footprint of nlp methods: a survey and analysis of existing tools",
"authors": [
{
"first": "Nesrine",
"middle": [],
"last": "Bannour",
"suffix": ""
},
{
"first": "Sahar",
"middle": [],
"last": "Ghannay",
"suffix": ""
},
{
"first": "Aur\u00e9lie",
"middle": [],
"last": "N\u00e9v\u00e9ol",
"suffix": ""
},
{
"first": "Anne-Laure",
"middle": [],
"last": "Ligozat",
"suffix": ""
}
],
"year": 2021,
"venue": "EMNLP, Workshop SustaiNLP",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nesrine Bannour, Sahar Ghannay, Aur\u00e9lie N\u00e9v\u00e9ol, and Anne-Laure Ligozat. 2021. Evaluating the carbon footprint of nlp methods: a survey and analysis of existing tools. In EMNLP, Workshop SustaiNLP.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "On the dangers of stochastic parrots: Can language models be too big?",
"authors": [
{
"first": "M",
"middle": [],
"last": "Emily",
"suffix": ""
},
{
"first": "Timnit",
"middle": [],
"last": "Bender",
"suffix": ""
},
{
"first": "Angelina",
"middle": [],
"last": "Gebru",
"suffix": ""
},
{
"first": "Shmargaret",
"middle": [],
"last": "Mcmillan-Major",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Shmitchell",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency",
"volume": "",
"issue": "",
"pages": "610--623",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Emily M Bender, Timnit Gebru, Angelina McMillan- Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Confer- ence on Fairness, Accountability, and Transparency, pages 610-623.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "On the opportunities and risks of foundation models",
"authors": [
{
"first": "Rishi",
"middle": [],
"last": "Bommasani",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Drew",
"suffix": ""
},
{
"first": "Ehsan",
"middle": [],
"last": "Hudson",
"suffix": ""
},
{
"first": "Russ",
"middle": [],
"last": "Adeli",
"suffix": ""
},
{
"first": "Simran",
"middle": [],
"last": "Altman",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Arora",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Sydney Von Arx",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Michael",
"suffix": ""
},
{
"first": "Jeannette",
"middle": [],
"last": "Bernstein",
"suffix": ""
},
{
"first": "Antoine",
"middle": [],
"last": "Bohg",
"suffix": ""
},
{
"first": "Emma",
"middle": [],
"last": "Bosselut",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Brunskill",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2108.07258"
]
},
"num": null,
"urls": [],
"raw_text": "Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. 2021. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Unified scaling laws for routed language models",
"authors": [
{
"first": "Aidan",
"middle": [],
"last": "Clark",
"suffix": ""
},
{
"first": "Diego",
"middle": [],
"last": "De Las Casas",
"suffix": ""
},
{
"first": "Aurelia",
"middle": [],
"last": "Guy",
"suffix": ""
},
{
"first": "Arthur",
"middle": [],
"last": "Mensch",
"suffix": ""
},
{
"first": "Michela",
"middle": [],
"last": "Paganini",
"suffix": ""
},
{
"first": "Jordan",
"middle": [],
"last": "Hoffmann",
"suffix": ""
},
{
"first": "Bogdan",
"middle": [],
"last": "Damoc",
"suffix": ""
},
{
"first": "Blake",
"middle": [],
"last": "Hechtman",
"suffix": ""
},
{
"first": "Trevor",
"middle": [],
"last": "Cai",
"suffix": ""
},
{
"first": "Sebastian",
"middle": [],
"last": "Borgeaud",
"suffix": ""
}
],
"year": 2022,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2202.01169"
]
},
"num": null,
"urls": [],
"raw_text": "Aidan Clark, Diego de las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bog- dan Damoc, Blake Hechtman, Trevor Cai, Se- bastian Borgeaud, et al. 2022. Unified scaling laws for routed language models. arXiv preprint arXiv:2202.01169.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity",
"authors": [
{
"first": "William",
"middle": [],
"last": "Fedus",
"suffix": ""
},
{
"first": "Barret",
"middle": [],
"last": "Zoph",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2101.03961"
]
},
"num": null,
"urls": [],
"raw_text": "William Fedus, Barret Zoph, and Noam Shazeer. 2021. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. arXiv preprint arXiv:2101.03961.",
"links": null
},
"BIBREF11": {
"ref_id": "b11",
"title": "The Pile: An 800gb dataset of diverse text for language modeling",
"authors": [
{
"first": "Leo",
"middle": [],
"last": "Gao",
"suffix": ""
},
{
"first": "Stella",
"middle": [],
"last": "Biderman",
"suffix": ""
},
{
"first": "Sid",
"middle": [],
"last": "Black",
"suffix": ""
},
{
"first": "Laurence",
"middle": [],
"last": "Golding",
"suffix": ""
},
{
"first": "Travis",
"middle": [],
"last": "Hoppe",
"suffix": ""
},
{
"first": "Charles",
"middle": [],
"last": "Foster",
"suffix": ""
},
{
"first": "Jason",
"middle": [],
"last": "Phang",
"suffix": ""
},
{
"first": "Horace",
"middle": [],
"last": "He",
"suffix": ""
},
{
"first": "Anish",
"middle": [],
"last": "Thite",
"suffix": ""
},
{
"first": "Noa",
"middle": [],
"last": "Nabeshima",
"suffix": ""
},
{
"first": "Shawn",
"middle": [],
"last": "Presser",
"suffix": ""
},
{
"first": "Connor",
"middle": [],
"last": "Leahy",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2101.00027"
]
},
"num": null,
"urls": [],
"raw_text": "Leo Gao, Stella Biderman, Sid Black, Laurence Gold- ing, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, and Connor Leahy. 2020. The Pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027.",
"links": null
},
"BIBREF12": {
"ref_id": "b12",
"title": "Chasing carbon: The elusive environmental footprint of computing",
"authors": [
{
"first": "Udit",
"middle": [],
"last": "Gupta",
"suffix": ""
},
{
"first": "Young",
"middle": [
"Geun"
],
"last": "Kim",
"suffix": ""
},
{
"first": "Sylvia",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Jordan",
"middle": [],
"last": "Tse",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Hsien-Hsin",
"suffix": ""
},
{
"first": "Gu-Yeon",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Wei",
"suffix": ""
},
{
"first": "Carole-Jean",
"middle": [],
"last": "Brooks",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wu",
"suffix": ""
}
],
"year": 2021,
"venue": "2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA)",
"volume": "",
"issue": "",
"pages": "854--867",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Udit Gupta, Young Geun Kim, Sylvia Lee, Jor- dan Tse, Hsien-Hsin S Lee, Gu-Yeon Wei, David Brooks, and Carole-Jean Wu. 2021. Chasing car- bon: The elusive environmental footprint of com- puting. In 2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA), pages 854-867. IEEE.",
"links": null
},
"BIBREF13": {
"ref_id": "b13",
"title": "Scaling laws for autoregressive generative modeling",
"authors": [
{
"first": "Tom",
"middle": [],
"last": "Henighan",
"suffix": ""
},
{
"first": "Jared",
"middle": [],
"last": "Kaplan",
"suffix": ""
},
{
"first": "Mor",
"middle": [],
"last": "Katz",
"suffix": ""
},
{
"first": "Mark",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "Christopher",
"middle": [],
"last": "Hesse",
"suffix": ""
},
{
"first": "Jacob",
"middle": [],
"last": "Jackson",
"suffix": ""
},
{
"first": "Heewoo",
"middle": [],
"last": "Jun",
"suffix": ""
},
{
"first": "B",
"middle": [],
"last": "Tom",
"suffix": ""
},
{
"first": "Prafulla",
"middle": [],
"last": "Brown",
"suffix": ""
},
{
"first": "Scott",
"middle": [],
"last": "Dhariwal",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Gray",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2010.14701"
]
},
"num": null,
"urls": [],
"raw_text": "Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. 2020. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701.",
"links": null
},
"BIBREF15": {
"ref_id": "b15",
"title": "What changes can large-scale language models bring? intensive study on hyperclova: Billions-scale korean generative pretrained transformers",
"authors": [
{
"first": "Boseop",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Hyoungseok",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Sang-Woo",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Gichang",
"middle": [],
"last": "Lee",
"suffix": ""
},
{
"first": "Donghyun",
"middle": [],
"last": "Kwak",
"suffix": ""
},
{
"first": "Sunghyun",
"middle": [],
"last": "Dong Hyeon Jeon",
"suffix": ""
},
{
"first": "Sungju",
"middle": [],
"last": "Park",
"suffix": ""
},
{
"first": "Seonhoon",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "Dongpil",
"middle": [],
"last": "Kim",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Seo",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2109.04650"
]
},
"num": null,
"urls": [],
"raw_text": "Boseop Kim, HyoungSeok Kim, Sang-Woo Lee, Gichang Lee, Donghyun Kwak, Dong Hyeon Jeon, Sunghyun Park, Sungju Kim, Seonhoon Kim, Dong- pil Seo, et al. 2021. What changes can large-scale language models bring? intensive study on hyper- clova: Billions-scale korean generative pretrained transformers. arXiv preprint arXiv:2109.04650.",
"links": null
},
"BIBREF16": {
"ref_id": "b16",
"title": "Quantifying the carbon emissions of machine learning",
"authors": [
{
"first": "Alexandre",
"middle": [],
"last": "Lacoste",
"suffix": ""
},
{
"first": "Alexandra",
"middle": [],
"last": "Luccioni",
"suffix": ""
},
{
"first": "Victor",
"middle": [],
"last": "Schmidt",
"suffix": ""
},
{
"first": "Thomas",
"middle": [],
"last": "Dandres",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:1910.09700"
]
},
"num": null,
"urls": [],
"raw_text": "Alexandre Lacoste, Alexandra Luccioni, Victor Schmidt, and Thomas Dandres. 2019. Quantifying the carbon emissions of machine learning. arXiv preprint arXiv:1910.09700.",
"links": null
},
"BIBREF17": {
"ref_id": "b17",
"title": "Jurassic-1: Technical details and evaluation",
"authors": [
{
"first": "Opher",
"middle": [],
"last": "Lieber",
"suffix": ""
},
{
"first": "Or",
"middle": [],
"last": "Sharir",
"suffix": ""
},
{
"first": "Barak",
"middle": [],
"last": "Lenz",
"suffix": ""
},
{
"first": "Yoav",
"middle": [],
"last": "Shoham",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Opher Lieber, Or Sharir, Barak Lenz, and Yoav Shoham. 2021. Jurassic-1: Technical details and evaluation. White Paper. AI21 Labs.",
"links": null
},
"BIBREF18": {
"ref_id": "b18",
"title": "Energy usage reports: Environmental awareness as part of algorithmic accountability",
"authors": [
{
"first": "Kadan",
"middle": [],
"last": "Lottick",
"suffix": ""
},
{
"first": "Silvia",
"middle": [],
"last": "Susai",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Sorelle",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [
"P"
],
"last": "Friedler",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Wilson",
"suffix": ""
}
],
"year": 2019,
"venue": "Workshop on Tackling Climate Change with Machine Learning at NeurIPS",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Kadan Lottick, Silvia Susai, Sorelle A. Friedler, and Jonathan P. Wilson. 2019. Energy usage reports: Environmental awareness as part of algorithmic ac- countability. Workshop on Tackling Climate Change with Machine Learning at NeurIPS 2019.",
"links": null
},
"BIBREF20": {
"ref_id": "b20",
"title": "Global warming of 1.5 c. An IPCC Special Report on the impacts of global warming of",
"authors": [
{
"first": "",
"middle": [],
"last": "Pidcock",
"suffix": ""
}
],
"year": 2018,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "R Pidcock, et al. 2018. Global warming of 1.5 c. An IPCC Special Report on the impacts of global warm- ing of, 1(5).",
"links": null
},
"BIBREF21": {
"ref_id": "b21",
"title": "Efficient large-scale language model training on gpu clusters using megatron-lm",
"authors": [
{
"first": "Deepak",
"middle": [],
"last": "Narayanan",
"suffix": ""
},
{
"first": "Mohammad",
"middle": [],
"last": "Shoeybi",
"suffix": ""
},
{
"first": "Jared",
"middle": [],
"last": "Casper",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Legresley",
"suffix": ""
},
{
"first": "Mostofa",
"middle": [],
"last": "Patwary",
"suffix": ""
},
{
"first": "Vijay",
"middle": [],
"last": "Korthikanti",
"suffix": ""
},
{
"first": "Dmitri",
"middle": [],
"last": "Vainbrand",
"suffix": ""
},
{
"first": "Prethvi",
"middle": [],
"last": "Kashinkunti",
"suffix": ""
},
{
"first": "Julie",
"middle": [],
"last": "Bernauer",
"suffix": ""
},
{
"first": "Bryan",
"middle": [],
"last": "Catanzaro",
"suffix": ""
}
],
"year": 2021,
"venue": "Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis",
"volume": "",
"issue": "",
"pages": "1--15",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan Catanzaro, et al. 2021. Efficient large-scale language model training on gpu clusters using megatron-lm. In Proceedings of the International Conference for High Perfor- mance Computing, Networking, Storage and Anal- ysis, pages 1-15.",
"links": null
},
"BIBREF22": {
"ref_id": "b22",
"title": "The Carbon Footprint of Machine Learning Training Will Plateau",
"authors": [
{
"first": "David",
"middle": [],
"last": "Patterson",
"suffix": ""
},
{
"first": "Joseph",
"middle": [],
"last": "Gonzalez",
"suffix": ""
},
{
"first": "Urs",
"middle": [],
"last": "H\u00f6lzle",
"suffix": ""
},
{
"first": "Hung",
"middle": [],
"last": "Quoc",
"suffix": ""
},
{
"first": "Chen",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Lluis-Miquel",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Munguia",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "Rothchild",
"suffix": ""
},
{
"first": "Maud",
"middle": [],
"last": "So",
"suffix": ""
},
{
"first": "Jeffrey",
"middle": [],
"last": "Texier",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Dean",
"suffix": ""
}
],
"year": 2022,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.36227/techrxiv.19139645.v2"
]
},
"num": null,
"urls": [],
"raw_text": "David Patterson, Joseph Gonzalez, Urs H\u00f6lzle, Quoc Hung Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jef- frey Dean. 2022. The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink.",
"links": null
},
"BIBREF23": {
"ref_id": "b23",
"title": "Carbon emissions and large neural network training",
"authors": [
{
"first": "David",
"middle": [],
"last": "Patterson",
"suffix": ""
},
{
"first": "Joseph",
"middle": [],
"last": "Gonzalez",
"suffix": ""
},
{
"first": "Quoc",
"middle": [],
"last": "Le",
"suffix": ""
},
{
"first": "Chen",
"middle": [],
"last": "Liang",
"suffix": ""
},
{
"first": "Lluis-Miquel",
"middle": [],
"last": "Munguia",
"suffix": ""
},
{
"first": "Daniel",
"middle": [],
"last": "Rothchild",
"suffix": ""
},
{
"first": "David",
"middle": [],
"last": "So",
"suffix": ""
},
{
"first": "Maud",
"middle": [],
"last": "Texier",
"suffix": ""
},
{
"first": "Jeff",
"middle": [],
"last": "Dean",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "David Patterson, Joseph Gonzalez, Quoc Le, Chen Liang, Lluis-Miquel Munguia, Daniel Rothchild, David So, Maud Texier, and Jeff Dean. 2021. Car- bon emissions and large neural network training.",
"links": null
},
"BIBREF24": {
"ref_id": "b24",
"title": "Gpt-3 powers the next generation of apps",
"authors": [
{
"first": "Ashley",
"middle": [],
"last": "Pilipiszyn",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ashley Pilipiszyn. 2021. Gpt-3 powers the next gener- ation of apps.",
"links": null
},
"BIBREF25": {
"ref_id": "b25",
"title": "The carbon footprint of distributed cloud storage",
"authors": [
{
"first": "Lorenzo",
"middle": [],
"last": "Posani",
"suffix": ""
},
{
"first": "Alessio",
"middle": [],
"last": "Paccoia",
"suffix": ""
},
{
"first": "Marco",
"middle": [],
"last": "Moschettini",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Lorenzo Posani, Alessio Paccoia, and Marco Moschet- tini. 2019. The carbon footprint of distributed cloud storage.",
"links": null
},
"BIBREF26": {
"ref_id": "b26",
"title": "Scaling language models: Methods, analysis & insights from training gopher",
"authors": [
{
"first": "Sebastian",
"middle": [],
"last": "Jack W Rae",
"suffix": ""
},
{
"first": "Trevor",
"middle": [],
"last": "Borgeaud",
"suffix": ""
},
{
"first": "Katie",
"middle": [],
"last": "Cai",
"suffix": ""
},
{
"first": "Jordan",
"middle": [],
"last": "Millican",
"suffix": ""
},
{
"first": "Francis",
"middle": [],
"last": "Hoffmann",
"suffix": ""
},
{
"first": "John",
"middle": [],
"last": "Song",
"suffix": ""
},
{
"first": "Sarah",
"middle": [],
"last": "Aslanides",
"suffix": ""
},
{
"first": "Roman",
"middle": [],
"last": "Henderson",
"suffix": ""
},
{
"first": "Susannah",
"middle": [],
"last": "Ring",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Young",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2112.11446"
]
},
"num": null,
"urls": [],
"raw_text": "Jack W Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song, John Aslanides, Sarah Henderson, Roman Ring, Susan- nah Young, et al. 2021. Scaling language models: Methods, analysis & insights from training gopher. arXiv preprint arXiv:2112.11446.",
"links": null
},
"BIBREF27": {
"ref_id": "b27",
"title": "Survey of machine learning accelerators",
"authors": [
{
"first": "Albert",
"middle": [],
"last": "Reuther",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "Michaleas",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "Vijay",
"middle": [],
"last": "Gadepally",
"suffix": ""
},
{
"first": "Siddharth",
"middle": [],
"last": "Samsi",
"suffix": ""
},
{
"first": "Jeremy",
"middle": [],
"last": "Kepner",
"suffix": ""
}
],
"year": 2020,
"venue": "2020 IEEE high performance extreme computing conference (HPEC)",
"volume": "",
"issue": "",
"pages": "1--12",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Albert Reuther, Peter Michaleas, Michael Jones, Vi- jay Gadepally, Siddharth Samsi, and Jeremy Kep- ner. 2020. Survey of machine learning accelerators. In 2020 IEEE high performance extreme computing conference (HPEC), pages 1-12. IEEE.",
"links": null
},
"BIBREF28": {
"ref_id": "b28",
"title": "Ai accelerator survey and trends",
"authors": [
{
"first": "Albert",
"middle": [],
"last": "Reuther",
"suffix": ""
},
{
"first": "Peter",
"middle": [],
"last": "Michaleas",
"suffix": ""
},
{
"first": "Michael",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "Vijay",
"middle": [],
"last": "Gadepally",
"suffix": ""
},
{
"first": "Siddharth",
"middle": [],
"last": "Samsi",
"suffix": ""
},
{
"first": "Jeremy",
"middle": [],
"last": "Kepner",
"suffix": ""
}
],
"year": 2021,
"venue": "2021 IEEE High Performance Extreme Computing Conference (HPEC)",
"volume": "",
"issue": "",
"pages": "1--9",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Albert Reuther, Peter Michaleas, Michael Jones, Vi- jay Gadepally, Siddharth Samsi, and Jeremy Kepner. 2021. Ai accelerator survey and trends. In 2021 IEEE High Performance Extreme Computing Con- ference (HPEC), pages 1-9. IEEE.",
"links": null
},
"BIBREF29": {
"ref_id": "b29",
"title": "Sorelle Friedler, and Sasha Luccioni. 2021. CodeCarbon: Estimate and Track Carbon Emissions from Machine Learning Computing",
"authors": [
{
"first": "Victor",
"middle": [],
"last": "Schmidt",
"suffix": ""
},
{
"first": "Kamal",
"middle": [],
"last": "Goyal",
"suffix": ""
},
{
"first": "Aditya",
"middle": [],
"last": "Joshi",
"suffix": ""
},
{
"first": "Boris",
"middle": [],
"last": "Feld",
"suffix": ""
},
{
"first": "Liam",
"middle": [],
"last": "Conell",
"suffix": ""
},
{
"first": "Nikolas",
"middle": [],
"last": "Laskaris",
"suffix": ""
},
{
"first": "Doug",
"middle": [],
"last": "Blank",
"suffix": ""
},
{
"first": "Jonathan",
"middle": [],
"last": "Wilson",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"DOI": [
"10.5281/zenodo.4658424"
]
},
"num": null,
"urls": [],
"raw_text": "Victor Schmidt, Kamal Goyal, Aditya Joshi, Boris Feld, Liam Conell, Nikolas Laskaris, Doug Blank, Jonathan Wilson, Sorelle Friedler, and Sasha Luc- cioni. 2021. CodeCarbon: Estimate and Track Car- bon Emissions from Machine Learning Computing.",
"links": null
},
"BIBREF31": {
"ref_id": "b31",
"title": "Tamay Besiroglu, Marius Hobbhahn, and Pablo Villalobos. 2022. Compute trends across three eras of machine learning",
"authors": [
{
"first": "Jaime",
"middle": [],
"last": "Sevilla",
"suffix": ""
},
{
"first": "Lennart",
"middle": [],
"last": "Heim",
"suffix": ""
},
{
"first": "Anson",
"middle": [],
"last": "Ho",
"suffix": ""
}
],
"year": null,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2202.05924"
]
},
"num": null,
"urls": [],
"raw_text": "Jaime Sevilla, Lennart Heim, Anson Ho, Tamay Be- siroglu, Marius Hobbhahn, and Pablo Villalobos. 2022. Compute trends across three eras of machine learning. arXiv preprint arXiv:2202.05924.",
"links": null
},
"BIBREF32": {
"ref_id": "b32",
"title": "Conference demographics and footprint changed by virtual platforms",
"authors": [
{
"first": "Matthew",
"middle": [],
"last": "Skiles",
"suffix": ""
},
{
"first": "Euijin",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Orad",
"middle": [],
"last": "Reshef",
"suffix": ""
},
{
"first": "Diego",
"middle": [
"Robalino"
],
"last": "Mu\u00f1oz",
"suffix": ""
},
{
"first": "Diana",
"middle": [],
"last": "Cintron",
"suffix": ""
},
{
"first": "Mary",
"middle": [
"Laura"
],
"last": "Lind",
"suffix": ""
},
{
"first": "Alexander",
"middle": [],
"last": "Rush",
"suffix": ""
},
{
"first": "Patricia",
"middle": [
"Perez"
],
"last": "Calleja",
"suffix": ""
},
{
"first": "Robert",
"middle": [],
"last": "Nerenberg",
"suffix": ""
},
{
"first": "Andrea",
"middle": [],
"last": "Armani",
"suffix": ""
}
],
"year": 2021,
"venue": "Nature Sustainability",
"volume": "",
"issue": "",
"pages": "1--8",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Matthew Skiles, Euijin Yang, Orad Reshef, Diego Robalino Mu\u00f1oz, Diana Cintron, Mary Laura Lind, Alexander Rush, Patricia Perez Calleja, Robert Nerenberg, Andrea Armani, et al. 2021. Conference demographics and footprint changed by virtual platforms. Nature Sustainability, pages 1-8.",
"links": null
},
"BIBREF33": {
"ref_id": "b33",
"title": "Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model",
"authors": [
{
"first": "Shaden",
"middle": [],
"last": "Smith",
"suffix": ""
},
{
"first": "Mostofa",
"middle": [],
"last": "Patwary",
"suffix": ""
},
{
"first": "Brandon",
"middle": [],
"last": "Norick",
"suffix": ""
},
{
"first": "Patrick",
"middle": [],
"last": "Legresley",
"suffix": ""
},
{
"first": "Samyam",
"middle": [],
"last": "Rajbhandari",
"suffix": ""
},
{
"first": "Jared",
"middle": [],
"last": "Casper",
"suffix": ""
},
{
"first": "Zhun",
"middle": [],
"last": "Liu",
"suffix": ""
},
{
"first": "Shrimai",
"middle": [],
"last": "Prabhumoye",
"suffix": ""
},
{
"first": "George",
"middle": [],
"last": "Zerveas",
"suffix": ""
},
{
"first": "Vijay",
"middle": [],
"last": "Korthikanti",
"suffix": ""
}
],
"year": 2022,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {
"arXiv": [
"arXiv:2201.11990"
]
},
"num": null,
"urls": [],
"raw_text": "Shaden Smith, Mostofa Patwary, Brandon Norick, Patrick LeGresley, Samyam Rajbhandari, Jared Casper, Zhun Liu, Shrimai Prabhumoye, George Zerveas, Vijay Korthikanti, et al. 2022. Using deepspeed and megatron to train megatron-turing nlg 530b, a large-scale generative language model. arXiv preprint arXiv:2201.11990.",
"links": null
},
"BIBREF34": {
"ref_id": "b34",
"title": "Energy and policy considerations for deep learning in nlp",
"authors": [
{
"first": "Emma",
"middle": [],
"last": "Strubell",
"suffix": ""
},
{
"first": "Ananya",
"middle": [],
"last": "Ganesh",
"suffix": ""
},
{
"first": "Andrew",
"middle": [],
"last": "Mccallum",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Emma Strubell, Ananya Ganesh, and Andrew McCal- lum. 2019. Energy and policy considerations for deep learning in nlp.",
"links": null
},
"BIBREF35": {
"ref_id": "b35",
"title": "Artificial intelligence and the climate emergency: Opportunities, challenges, and recommendations",
"authors": [
{
"first": "Mariarosaria",
"middle": [],
"last": "Taddeo",
"suffix": ""
},
{
"first": "Andreas",
"middle": [],
"last": "Tsamados",
"suffix": ""
},
{
"first": "Josh",
"middle": [],
"last": "Cowls",
"suffix": ""
},
{
"first": "Luciano",
"middle": [],
"last": "Floridi",
"suffix": ""
}
],
"year": 2021,
"venue": "One Earth",
"volume": "4",
"issue": "",
"pages": "776--779",
"other_ids": {
"DOI": [
"10.1016/j.oneear.2021.05.018"
]
},
"num": null,
"urls": [],
"raw_text": "Mariarosaria Taddeo, Andreas Tsamados, Josh Cowls, and Luciano Floridi. 2021. Artificial intelligence and the climate emergency: Opportunities, chal- lenges, and recommendations. One Earth, 4:776- 779.",
"links": null
},
"BIBREF36": {
"ref_id": "b36",
"title": "The computational limits of deep learning",
"authors": [
{
"first": "C",
"middle": [],
"last": "Neil",
"suffix": ""
},
{
"first": "Kristjan",
"middle": [],
"last": "Thompson",
"suffix": ""
},
{
"first": "Keeheon",
"middle": [],
"last": "Greenewald",
"suffix": ""
},
{
"first": "Gabriel",
"middle": [
"F"
],
"last": "Lee",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Manso",
"suffix": ""
}
],
"year": 2020,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Neil C. Thompson, Kristjan Greenewald, Keeheon Lee, and Gabriel F. Manso. 2020. The computational lim- its of deep learning.",
"links": null
},
"BIBREF37": {
"ref_id": "b37",
"title": "Fanny Cachat van der Haert, Frank Mugisha, et al. 2020. Ai for social good: unlocking the opportunity for positive impact",
"authors": [
{
"first": "Nenad",
"middle": [],
"last": "Toma\u0161ev",
"suffix": ""
},
{
"first": "Julien",
"middle": [],
"last": "Cornebise",
"suffix": ""
},
{
"first": "Frank",
"middle": [],
"last": "Hutter",
"suffix": ""
},
{
"first": "Shakir",
"middle": [],
"last": "Mohamed",
"suffix": ""
},
{
"first": "Angela",
"middle": [],
"last": "Picciariello",
"suffix": ""
}
],
"year": null,
"venue": "Nature Communications",
"volume": "11",
"issue": "1",
"pages": "1--6",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nenad Toma\u0161ev, Julien Cornebise, Frank Hutter, Shakir Mohamed, Angela Picciariello, Bec Con- nelly, Danielle Belgrave, Daphne Ezer, Fanny Cachat van der Haert, Frank Mugisha, et al. 2020. Ai for social good: unlocking the opportunity for positive impact. Nature Communications, 11(1):1- 6.",
"links": null
},
"BIBREF38": {
"ref_id": "b38",
"title": "Training data-efficient image transformers & distillation through attention",
"authors": [
{
"first": "Hugo",
"middle": [],
"last": "Touvron",
"suffix": ""
},
{
"first": "Matthieu",
"middle": [],
"last": "Cord",
"suffix": ""
},
{
"first": "Matthijs",
"middle": [],
"last": "Douze",
"suffix": ""
},
{
"first": "Francisco",
"middle": [],
"last": "Massa",
"suffix": ""
},
{
"first": "Alexandre",
"middle": [],
"last": "Sablayrolles",
"suffix": ""
},
{
"first": "Herv\u00e9",
"middle": [],
"last": "J\u00e9gou",
"suffix": ""
}
],
"year": 2021,
"venue": "International Conference on Machine Learning",
"volume": "",
"issue": "",
"pages": "10347--10357",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Hugo Touvron, Matthieu Cord, Matthijs Douze, Fran- cisco Massa, Alexandre Sablayrolles, and Herv\u00e9 J\u00e9gou. 2021. Training data-efficient image trans- formers & distillation through attention. In Inter- national Conference on Machine Learning, pages 10347-10357. PMLR.",
"links": null
},
"BIBREF39": {
"ref_id": "b39",
"title": "Attention is all you need. Advances in neural information processing systems",
"authors": [
{
"first": "Ashish",
"middle": [],
"last": "Vaswani",
"suffix": ""
},
{
"first": "Noam",
"middle": [],
"last": "Shazeer",
"suffix": ""
},
{
"first": "Niki",
"middle": [],
"last": "Parmar",
"suffix": ""
},
{
"first": "Jakob",
"middle": [],
"last": "Uszkoreit",
"suffix": ""
},
{
"first": "Llion",
"middle": [],
"last": "Jones",
"suffix": ""
},
{
"first": "Aidan",
"middle": [
"N"
],
"last": "Gomez",
"suffix": ""
},
{
"first": "\u0141ukasz",
"middle": [],
"last": "Kaiser",
"suffix": ""
},
{
"first": "Illia",
"middle": [],
"last": "Polosukhin",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information process- ing systems, 30.",
"links": null
},
"BIBREF40": {
"ref_id": "b40",
"title": "GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model",
"authors": [
{
"first": "Ben",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Aran",
"middle": [],
"last": "Komatsuzaki",
"suffix": ""
}
],
"year": 2021,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Ben Wang and Aran Komatsuzaki. 2021. GPT-J- 6B: A 6 Billion Parameter Autoregressive Language Model.",
"links": null
},
"BIBREF41": {
"ref_id": "b41",
"title": "Ccnet: Extracting high quality monolingual datasets from web crawl data",
"authors": [
{
"first": "Guillaume",
"middle": [],
"last": "Wenzek",
"suffix": ""
},
{
"first": "Marie-Anne",
"middle": [],
"last": "Lachaux",
"suffix": ""
},
{
"first": "Alexis",
"middle": [],
"last": "Conneau",
"suffix": ""
},
{
"first": "Vishrav",
"middle": [],
"last": "Chaudhary",
"suffix": ""
},
{
"first": "Francisco",
"middle": [],
"last": "Guzm\u00e1n",
"suffix": ""
},
{
"first": "Armand",
"middle": [],
"last": "Joulin",
"suffix": ""
},
{
"first": "Edouard",
"middle": [],
"last": "Grave",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Guillaume Wenzek, Marie-Anne Lachaux, Alexis Con- neau, Vishrav Chaudhary, Francisco Guzm\u00e1n, Ar- mand Joulin, and Edouard Grave. 2019. Ccnet: Ex- tracting high quality monolingual datasets from web crawl data.",
"links": null
},
"BIBREF43": {
"ref_id": "b43",
"title": "Quantization networks",
"authors": [
{
"first": "Jiwei",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Xu",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "Jun",
"middle": [],
"last": "Xing",
"suffix": ""
},
{
"first": "Xinmei",
"middle": [],
"last": "Tian",
"suffix": ""
},
{
"first": "Houqiang",
"middle": [],
"last": "Li",
"suffix": ""
},
{
"first": "Bing",
"middle": [],
"last": "Deng",
"suffix": ""
},
{
"first": "Jianqiang",
"middle": [],
"last": "Huang",
"suffix": ""
},
{
"first": "Xiansheng",
"middle": [],
"last": "Hua",
"suffix": ""
}
],
"year": 2019,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Jiwei Yang, Xu Shen, Jun Xing, Xinmei Tian, Houqiang Li, Bing Deng, Jianqiang Huang, and Xi- ansheng Hua. 2019. Quantization networks.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "Figure 1: Over the last four years, the size of stateof-the-art language models has doubled every 3-4 months. Note that this trend has been slowing down, due to scale-out limitations.",
"num": null,
"uris": null,
"type_str": "figure"
},
"FIGREF1": {
"text": "Breakdown of the carbon footprint (total 36.5t tC02e) of the Noor project. This breakdown is highly dependent on the localisation of the workloads and the local carbon intensity of the electricity mix.",
"num": null,
"uris": null,
"type_str": "figure"
},
"TABREF0": {
"text": "Training compute budget and energy used for training the Noor models. Assuming a pretraining dataset of 150B tokens and a throughput of 100 TFLOPs per A100.",
"num": null,
"html": null,
"content": "<table><tr><td colspan=\"3\">Model Budget [PF-days] Budget [A100-hours]</td><td>HPC</td><td>Consumption [MWh]</td></tr><tr><td>1.3B</td><td>13.5</td><td>3300</td><td>MeluXina</td><td>2.1</td></tr><tr><td>2.7B</td><td>28.1</td><td>6800</td><td>Noor-HPC</td><td>4.8</td></tr><tr><td>6.7B</td><td>69.8</td><td>17000</td><td>Noor-HPC</td><td>11.8</td></tr><tr><td>13B</td><td>135</td><td>33000</td><td>Noor-HPC</td><td>22.9</td></tr><tr><td colspan=\"3\">reference language model trained on Wikipedia.</td><td/><td/></tr><tr><td colspan=\"3\">Processing with our pipeline occurs on a CPU clus-</td><td/><td/></tr><tr><td colspan=\"2\">ter with 768 cores, split over 16 nodes.</td><td/><td/><td/></tr><tr><td colspan=\"3\">Using average high-performance CPUs TDP fig-</td><td/><td/></tr></table>",
"type_str": "table"
},
"TABREF1": {
"text": "Carbon footprint of each phase of the Noor project. Use MWh Footprint [tCO2e]",
"num": null,
"html": null,
"content": "<table><tr><td colspan=\"3\">Phase Mix [gCO2e/kWh] Storage Amazon S3 Provider Location Bahrain 1188 GCP Netherlands 410 R&D MeluXina Luxembourg 60</td><td>3.5 2.35 10.7</td><td>4.2 0.96 0.65</td></tr><tr><td colspan=\"2\">Training MeluXina Luxembourg Noor-HPC UAE</td><td>60 600</td><td>2.1 39.5</td><td>0.13 23.7</td></tr><tr><td>Others</td><td>France UAE</td><td>56 600</td><td>0.33 0.66</td><td>0.02 0.4</td></tr></table>",
"type_str": "table"
}
}
}
} |