{ "assumptions": "H100 SXM BF16 dense peak 989 TFLOP/s planning scenario; MFU .2/.35/.5 not measured; price $2/$3/$5 per GPU-hour are editable hypothetical inputs, not vendor quotes; excludes retries/people/tax/storage/power", "architectures": { "B_custom_120B_class": { "layers": 49, "hidden_size": 3072, "expert_intermediate": 2048, "experts": 128, "selected": 4, "shared": 1, "heads": 24, "kv_heads": 8, "head_dim": 128, "vocab": 131072, "attention_parameters": 1233125376, "expert_parameters": 119304880128, "router_parameters": 19267584, "norm_parameters": 304128, "tied_embedding_parameters": 402653184, "total_parameters": 120960230400, "active_parameters_upper_convention": 6279570432, "active_convention": "Counts full tied embedding/output matrix, all attention, router and selected/shared FFNs; actual sparse embedding lookup uses fewer elements", "kv_cache_GB_batch1": { "32768": 6.576668672, "65536": 13.153337344, "131072": 26.306674688, "262144": 52.613349376 } }, "C_custom_larger_MoE": { "layers": 80, "hidden_size": 4096, "expert_intermediate": 1536, "experts": 160, "selected": 4, "shared": 1, "heads": 32, "kv_heads": 8, "head_dim": 128, "vocab": 131072, "attention_parameters": 3355443200, "expert_parameters": 243101859840, "router_parameters": 52428800, "norm_parameters": 659456, "tied_embedding_parameters": 536870912, "total_parameters": 247047262208, "active_parameters_upper_convention": 11495149568, "active_convention": "Counts full tied embedding/output matrix, all attention, router and selected/shared FFNs; actual sparse embedding lookup uses fewer elements", "kv_cache_GB_batch1": { "32768": 10.73741824, "65536": 21.47483648, "131072": 42.94967296, "262144": 85.89934592 } } }, "estimates": { "A_dense_120B": { "low_cost": { "total_parameters": 120000000000.0, "active_parameters": 120000000000.0, "tokens": 2400000000000.0, "gpus": 1024, "peak_tflops": 989, "mfu": 0.5, "gpu_hour_usd": 2, "training_flops": 1.728e+24, "hours": 947.9271991911022, "gpu_hours": 970677.4519716887, "compute_usd": 1941354.9039433773, "weight_GB": { "bf16": 240.0, "fp8": 120.0, "int8": 120.0, "int4_ideal": 60.0 }, "adam_training_state_GB": 1920.0, "checkpoint_weights_master_moments_GB": 1680.0, "token_storage_uint32_GB": 9600.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "expected": { "total_parameters": 120000000000.0, "active_parameters": 120000000000.0, "tokens": 2400000000000.0, "gpus": 1024, "peak_tflops": 989, "mfu": 0.35, "gpu_hour_usd": 3, "training_flops": 1.728e+24, "hours": 1354.181713130146, "gpu_hours": 1386682.0742452694, "compute_usd": 4160046.222735808, "weight_GB": { "bf16": 240.0, "fp8": 120.0, "int8": 120.0, "int4_ideal": 60.0 }, "adam_training_state_GB": 1920.0, "checkpoint_weights_master_moments_GB": 1680.0, "token_storage_uint32_GB": 9600.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "high_cost": { "total_parameters": 120000000000.0, "active_parameters": 120000000000.0, "tokens": 2400000000000.0, "gpus": 1024, "peak_tflops": 989, "mfu": 0.2, "gpu_hour_usd": 5, "training_flops": 1.728e+24, "hours": 2369.8179979777556, "gpu_hours": 2426693.6299292217, "compute_usd": 12133468.149646109, "weight_GB": { "bf16": 240.0, "fp8": 120.0, "int8": 120.0, "int4_ideal": 60.0 }, "adam_training_state_GB": 1920.0, "checkpoint_weights_master_moments_GB": 1680.0, "token_storage_uint32_GB": 9600.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" } }, "B_custom_120B_class": { "low_cost": { "total_parameters": 120960230400, "active_parameters": 6279570432, "tokens": 2400000000000.0, "gpus": 256, "peak_tflops": 989, "mfu": 0.5, "gpu_hour_usd": 2, "training_flops": 9.04258142208e+22, "hours": 198.41918705763396, "gpu_hours": 50795.311886754294, "compute_usd": 101590.62377350859, "weight_GB": { "bf16": 241.9204608, "fp8": 120.9602304, "int8": 120.9602304, "int4_ideal": 60.4801152 }, "adam_training_state_GB": 1935.3636864, "checkpoint_weights_master_moments_GB": 1693.4432256, "token_storage_uint32_GB": 9600.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "expected": { "total_parameters": 120960230400, "active_parameters": 6279570432, "tokens": 2400000000000.0, "gpus": 256, "peak_tflops": 989, "mfu": 0.35, "gpu_hour_usd": 3, "training_flops": 9.04258142208e+22, "hours": 283.4559815109057, "gpu_hours": 72564.73126679186, "compute_usd": 217694.19380037556, "weight_GB": { "bf16": 241.9204608, "fp8": 120.9602304, "int8": 120.9602304, "int4_ideal": 60.4801152 }, "adam_training_state_GB": 1935.3636864, "checkpoint_weights_master_moments_GB": 1693.4432256, "token_storage_uint32_GB": 9600.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "high_cost": { "total_parameters": 120960230400, "active_parameters": 6279570432, "tokens": 2400000000000.0, "gpus": 256, "peak_tflops": 989, "mfu": 0.2, "gpu_hour_usd": 5, "training_flops": 9.04258142208e+22, "hours": 496.0479676440849, "gpu_hours": 126988.27971688574, "compute_usd": 634941.3985844287, "weight_GB": { "bf16": 241.9204608, "fp8": 120.9602304, "int8": 120.9602304, "int4_ideal": 60.4801152 }, "adam_training_state_GB": 1935.3636864, "checkpoint_weights_master_moments_GB": 1693.4432256, "token_storage_uint32_GB": 9600.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" } }, "C_custom_larger_MoE": { "low_cost": { "total_parameters": 247047262208, "active_parameters": 11495149568, "tokens": 3000000000000.0, "gpus": 512, "peak_tflops": 989, "mfu": 0.5, "gpu_hour_usd": 2, "training_flops": 2.06912692224e+23, "hours": 227.01176946410516, "gpu_hours": 116230.02596562184, "compute_usd": 232460.05193124368, "weight_GB": { "bf16": 494.094524416, "fp8": 247.047262208, "int8": 247.047262208, "int4_ideal": 123.523631104 }, "adam_training_state_GB": 3952.756195328, "checkpoint_weights_master_moments_GB": 3458.661670912, "token_storage_uint32_GB": 12000.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "expected": { "total_parameters": 247047262208, "active_parameters": 11495149568, "tokens": 3000000000000.0, "gpus": 512, "peak_tflops": 989, "mfu": 0.35, "gpu_hour_usd": 3, "training_flops": 2.06912692224e+23, "hours": 324.3025278058645, "gpu_hours": 166042.89423660262, "compute_usd": 498128.6827098079, "weight_GB": { "bf16": 494.094524416, "fp8": 247.047262208, "int8": 247.047262208, "int4_ideal": 123.523631104 }, "adam_training_state_GB": 3952.756195328, "checkpoint_weights_master_moments_GB": 3458.661670912, "token_storage_uint32_GB": 12000.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "high_cost": { "total_parameters": 247047262208, "active_parameters": 11495149568, "tokens": 3000000000000.0, "gpus": 512, "peak_tflops": 989, "mfu": 0.2, "gpu_hour_usd": 5, "training_flops": 2.06912692224e+23, "hours": 567.5294236602629, "gpu_hours": 290575.0649140546, "compute_usd": 1452875.324570273, "weight_GB": { "bf16": 494.094524416, "fp8": 247.047262208, "int8": 247.047262208, "int4_ideal": 123.523631104 }, "adam_training_state_GB": 3952.756195328, "checkpoint_weights_master_moments_GB": 3458.661670912, "token_storage_uint32_GB": 12000.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" } }, "D_30B_continue": { "low_cost": { "total_parameters": 30500000000.0, "active_parameters": 3300000000.0, "tokens": 10000000000.0, "gpus": 8, "peak_tflops": 989, "mfu": 0.5, "gpu_hour_usd": 2, "training_flops": 1.98e+20, "hours": 13.902932254802831, "gpu_hours": 111.22345803842265, "compute_usd": 222.4469160768453, "weight_GB": { "bf16": 61.0, "fp8": 30.5, "int8": 30.5, "int4_ideal": 15.25 }, "adam_training_state_GB": 488.0, "checkpoint_weights_master_moments_GB": 427.0, "token_storage_uint32_GB": 40.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "expected": { "total_parameters": 30500000000.0, "active_parameters": 3300000000.0, "tokens": 10000000000.0, "gpus": 8, "peak_tflops": 989, "mfu": 0.35, "gpu_hour_usd": 3, "training_flops": 1.98e+20, "hours": 19.861331792575474, "gpu_hours": 158.8906543406038, "compute_usd": 476.6719630218114, "weight_GB": { "bf16": 61.0, "fp8": 30.5, "int8": 30.5, "int4_ideal": 15.25 }, "adam_training_state_GB": 488.0, "checkpoint_weights_master_moments_GB": 427.0, "token_storage_uint32_GB": 40.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "high_cost": { "total_parameters": 30500000000.0, "active_parameters": 3300000000.0, "tokens": 10000000000.0, "gpus": 8, "peak_tflops": 989, "mfu": 0.2, "gpu_hour_usd": 5, "training_flops": 1.98e+20, "hours": 34.75733063700708, "gpu_hours": 278.05864509605664, "compute_usd": 1390.2932254802831, "weight_GB": { "bf16": 61.0, "fp8": 30.5, "int8": 30.5, "int4_ideal": 15.25 }, "adam_training_state_GB": 488.0, "checkpoint_weights_master_moments_GB": 427.0, "token_storage_uint32_GB": 40.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" } }, "E_4B_continue": { "low_cost": { "total_parameters": 4000000000.0, "active_parameters": 4000000000.0, "tokens": 1000000000.0, "gpus": 8, "peak_tflops": 989, "mfu": 0.5, "gpu_hour_usd": 2, "training_flops": 2.4e+19, "hours": 1.6852039096730704, "gpu_hours": 13.481631277384563, "compute_usd": 26.963262554769127, "weight_GB": { "bf16": 8.0, "fp8": 4.0, "int8": 4.0, "int4_ideal": 2.0 }, "adam_training_state_GB": 64.0, "checkpoint_weights_master_moments_GB": 56.0, "token_storage_uint32_GB": 4.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "expected": { "total_parameters": 4000000000.0, "active_parameters": 4000000000.0, "tokens": 1000000000.0, "gpus": 8, "peak_tflops": 989, "mfu": 0.35, "gpu_hour_usd": 3, "training_flops": 2.4e+19, "hours": 2.407434156675815, "gpu_hours": 19.25947325340652, "compute_usd": 57.778419760219556, "weight_GB": { "bf16": 8.0, "fp8": 4.0, "int8": 4.0, "int4_ideal": 2.0 }, "adam_training_state_GB": 64.0, "checkpoint_weights_master_moments_GB": 56.0, "token_storage_uint32_GB": 4.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" }, "high_cost": { "total_parameters": 4000000000.0, "active_parameters": 4000000000.0, "tokens": 1000000000.0, "gpus": 8, "peak_tflops": 989, "mfu": 0.2, "gpu_hour_usd": 5, "training_flops": 2.4e+19, "hours": 4.213009774182676, "gpu_hours": 33.704078193461406, "compute_usd": 168.52039096730704, "weight_GB": { "bf16": 8.0, "fp8": 4.0, "int8": 4.0, "int4_ideal": 2.0 }, "adam_training_state_GB": 64.0, "checkpoint_weights_master_moments_GB": 56.0, "token_storage_uint32_GB": 4.0, "limitations": "6NT omits attention quadratic term, routing, communication and recomputation; quantization excludes scales/metadata; costs illustrative" } } }, "topologies": [ { "world": 8, "tp": 1, "pp": 1, "cp": 1, "dp": 8, "ep": 1, "expert_data_parallel": 8, "note": "EP is a subgroup of DP here, not another world-size multiplier; framework-specific mapping must be validated" }, { "world": 64, "tp": 2, "pp": 2, "cp": 1, "dp": 16, "ep": 8, "expert_data_parallel": 2, "note": "EP is a subgroup of DP here, not another world-size multiplier; framework-specific mapping must be validated" }, { "world": 1024, "tp": 8, "pp": 8, "cp": 2, "dp": 8, "ep": 1, "expert_data_parallel": 8, "note": "EP is a subgroup of DP here, not another world-size multiplier; framework-specific mapping must be validated" } ], "storage_formula": "3 retained checkpoints * checkpoint_GB + 2 replicated copies of token shards + raw corpus + intermediate dedup data; add validation/holdouts and free-space headroom" }