Datasets:
willtheorangeguy
/
2017-WAN-Show-Transcripts

Dataset card Data Studio Files
xet
 Community
text
stringlengths
11
441
[1602.00 --> 1604.30] So 8-core memory up to 16 DIMMs per CPU.
[1607.06 --> 1608.34] They've got...
[1608.34 --> 1613.54] Epic provides up to 128 PCI Express 3.0 lanes.
[1613.86 --> 1617.06] Allowing you to connect more GPUs directly to the CPU.
[1617.06 --> 1620.88] And that's really important for machine learning type applications.
[1620.88 --> 1626.10] We've got our unboxing of SFUs, the fastest supercomputer in Canada.
[1626.36 --> 1630.26] So I went down there and they've got a ton of GPU nodes.
[1630.70 --> 1637.02] And they would love to have the GPUs connected directly to the PCIe lanes on the CPU.
[1637.12 --> 1639.06] In fact, we were talking about that for...
[1639.06 --> 1643.66] It's in my notes, but I don't remember exactly what workload specifically it was.
[1643.66 --> 1643.90] Okay.
[1643.90 --> 1648.22] Where inter-GPU communication is incredibly important.
[1648.52 --> 1650.98] And the more hubs that you have to pass through, the worse.
[1651.58 --> 1652.82] And then I even talked to...
[1652.82 --> 1655.48] It's so cool talking to these high-level architect guys.
[1655.56 --> 1656.86] I'm like, okay, well, what about NVLink?
[1656.90 --> 1660.84] And he's like, yeah, the cost just wasn't right for what we're doing.
[1660.96 --> 1665.96] And it's all these kinds of trade-offs where normally as a consumer, you hear about this crap.
[1665.96 --> 1673.82] And you've got NVIDIA on a stage with freaking wood screws and whatever it is that they're talking about going, this is the future of blah, blah, blah.
[1674.00 --> 1678.50] But then, you know, to us, this is all totally unattainable stuff.
[1678.50 --> 1695.62] Whereas you're talking about people who are working on data centers that are in the value of $10 to $20 million, where they're actually taking like a $2 million solution and a $3 million solution and evaluating it the same way that we would like two different laptops.
[1697.00 --> 1697.44] Yeah.
[1697.44 --> 1699.88] But they're just working on a different scale.
[1700.00 --> 1701.62] It's all the same conversations.
[1701.82 --> 1705.80] It's like, do I want to pay this much more for this much more performance increase?
[1705.90 --> 1708.80] And they just, for their design, it didn't end up making sense.
[1709.12 --> 1711.14] And it was going to add some cost.
[1711.18 --> 1712.50] And I don't know the exact numbers.
[1712.58 --> 1716.08] So there's no like NDA information that I could hope to give you.
[1716.14 --> 1717.54] I just know that they didn't use it.
[1717.66 --> 1721.72] And they opted for these, they opted for certain GPU nodes.
[1721.72 --> 1730.32] Some of the GPU nodes for certain workloads are connected directly to the CPUs, and some of them are not, depending on what they're needed for.
[1730.38 --> 1731.48] And these ones are higher cost.
[1731.56 --> 1732.38] I was just like, oh, cool.
[1732.88 --> 1733.76] That's pretty fancy.
[1733.78 --> 1734.78] You guys are so baller.
[1734.80 --> 1735.28] I love you.
[1736.86 --> 1740.44] So Epic is expected to launch before the end of June.
[1740.48 --> 1746.82] You can tell AMD is in full-on crisis mode about their stock dropping, what was it, two weeks ago?
[1747.38 --> 1748.00] Oh, my goodness.
[1748.10 --> 1749.38] We have so much stuff coming.
[1749.38 --> 1750.44] We have stuff.
[1750.56 --> 1751.62] Please just wait.
[1752.08 --> 1754.24] We have so much stuff.
[1755.48 --> 1758.30] Even Radeon Technologies Group was showing off stuff.
[1758.38 --> 1761.94] So this was originally posted by PCGamer324 on the forum.
[1762.04 --> 1767.36] The original article here is from, oh, three, did I say three?
[1767.58 --> 1768.82] Okay, no, Digital Trends.
[1768.84 --> 1769.26] Here we go.
[1769.96 --> 1776.34] AMD's Radeon Vega Frontier Edition is the most powerful graphics card yet, according to AMD.
[1779.28 --> 1780.56] All right.
[1780.56 --> 1780.62] All right.
[1780.62 --> 1787.20] So the numbers put forth by AMD are as follows.
[1787.20 --> 1791.22] It'll have 64 next-generation compute units.
[1791.22 --> 1799.90] So that's 13 teraflops of peak single precision 32-bit compute performance, which should put it in the neighborhood of NVIDIA's top cards.
[1799.90 --> 1801.48] I think their top card is 12.
[1801.48 --> 1807.34] And that's a quadro P6000.
[1807.34 --> 1812.28] Don't quote me on that, but I'm trying to look this up as quickly as I can.
[1814.42 --> 1815.50] Quadro P6000.
[1815.60 --> 1815.74] Yeah.
[1815.74 --> 1822.04] 12 teraflops for their top-of-the-line quadro, which is about a $5,000 graphics card.
[1822.60 --> 1823.92] What else do we know about it?
[1823.92 --> 1828.24] I wasn't sure about all the stuff you said because I was checking in with something.
[1828.24 --> 1834.68] 16 gigs of HBM2 memory, so that's pretty impressive on a super-wide 2048-bit bus.
[1834.68 --> 1836.20] Have you shown them the shroud?
[1837.10 --> 1839.52] The shroud looks freaking sick.
[1839.60 --> 1839.88] Yeah.
[1841.24 --> 1843.42] AMD finally gets it on the shroud.
[1843.52 --> 1848.30] Although, if it's like plastic and kind of cheesy looking, then I really don't know if that's going to be that impressive.
[1848.46 --> 1855.14] To be fair, the last one that came out that was plastic and cheesy looking, once it was in a case, it was actually a little hard to tell.
[1855.90 --> 1856.20] Right.
[1856.20 --> 1856.86] Okay, fair enough.
[1856.86 --> 1859.62] It's still not amazing, but it wasn't that bad.
[1860.74 --> 1877.60] So they're saying that Vega's memory architecture can access terabytes of memory, and we've actually seen them kind of move beyond the memory that's HBM2 or GDDR memory that is soldered to the board.
[1877.60 --> 1887.32] And what was that professional card that was using SSDs, like it was using NAND in order to have – I forget how much memory it was on it.
[1887.66 --> 1888.14] I don't remember.
[1888.14 --> 1889.02] I'm sorry, you guys.
[1889.08 --> 1895.36] I wish I had a more exact set of talking points for this, but they were showing it off at CES.
[1895.36 --> 1905.58] And the idea was that they were using NAND flash in order to dramatically increase the amount of – basically the size of the project that you could work on,
[1905.64 --> 1912.84] because it's still faster since you've got a nice lightning-fast interface than trying to go out to like a scratch disk or something.
[1913.04 --> 1922.46] So apparently AMD is claiming that they've done several benchmarks with it, and the Vega Frontier Edition is apparently 70% faster in SOLIDWORKS?
[1922.46 --> 1924.46] Okay.
[1925.30 --> 1927.06] Which is like pretty insane.
[1927.36 --> 1935.32] Another program that I haven't personally heard of called Katya – I'm hoping I'm saying that correctly – it's 27% faster in.
[1936.28 --> 1936.76] Okay.
[1937.04 --> 1937.96] Which is pretty wild.
[1938.76 --> 1942.38] I mean, AMD has been known from time to time.
[1944.86 --> 1950.26] Sorry, I didn't mean to get an ice cube in my mouth, and now – okay, it's gone now.
[1950.28 --> 1950.58] Very nice.
[1950.58 --> 1954.52] So AMD has been known to cherry-pick benchmarks from time to time.
[1956.62 --> 1959.16] But still, it's promising.
[1959.42 --> 1962.64] And like that happened last time – you just reorganized it and control Z that.
[1963.28 --> 1966.06] That happened last time with their CPUs to a certain degree.
[1966.24 --> 1970.10] But then a big part of what they're claiming did end up being true.
[1970.14 --> 1970.54] That's right.
[1970.54 --> 1977.72] And what we do know, based on that it's not getting completely stomped on, is that we're going to have competition.
[1978.04 --> 1978.24] Yeah.
[1978.38 --> 1980.54] At least until NVIDIA trickles down Volta.
[1981.08 --> 1982.36] And like – yeah, fair.
[1982.66 --> 1987.56] But like with their CPU, what they were claiming did completely end up being true.