{ "execution_mode": "CPU algorithm with controlled inputs; any latency/capacity coefficients are assumptions", "gpu_catalog": [ { "name": "example-small", "memory_gb": 48.0, "bandwidth_gbps": 864.0, "bf16_tflops": 362.0, "hourly_cost_usd": 1.55, "ridge_point_flops_per_byte": 418.98 }, { "name": "example-fast", "memory_gb": 80.0, "bandwidth_gbps": 3350.0, "bf16_tflops": 989.0, "hourly_cost_usd": 3.2, "ridge_point_flops_per_byte": 295.22 }, { "name": "example-large-memory", "memory_gb": 141.0, "bandwidth_gbps": 4800.0, "bf16_tflops": 989.0, "hourly_cost_usd": 3.8, "ridge_point_flops_per_byte": 206.04 }, { "name": "example-wide-bandwidth", "memory_gb": 192.0, "bandwidth_gbps": 8000.0, "bf16_tflops": 2250.0, "hourly_cost_usd": 6.2, "ridge_point_flops_per_byte": 281.25 } ], "llama_8b_batch8": [ { "gpu": "example-small", "memory_gb": 48.0, "required_memory_gb": 20.42, "fits_single_gpu": true, "ridge_point_flops_per_byte": 418.98, "prefill": { "phase": "prefill", "flops": 262144000000000.0, "hbm_bytes": 18147483648.0, "operational_intensity": 14445.2, "ridge_point": 418.98, "regime": "compute-bound", "roofline_tflops_bound": 362.0, "latency_ms": 724.155 }, "decode": { "phase": "decode_step", "flops": 128000000000.0, "hbm_bytes": 18281701376.0, "operational_intensity": 7.0, "ridge_point": 418.98, "regime": "bandwidth-bound", "roofline_tflops_bound": 6.05, "latency_ms": 21.159 }, "ttft_ms": 724.15, "tpot_ms": 21.159, "decode_tokens_per_sec": 378.09, "end_to_end_tokens_per_sec": 334.66, "tokens_per_dollar": 777267.7 }, { "gpu": "example-fast", "memory_gb": 80.0, "required_memory_gb": 20.42, "fits_single_gpu": true, "ridge_point_flops_per_byte": 295.22, "prefill": { "phase": "prefill", "flops": 262144000000000.0, "hbm_bytes": 18147483648.0, "operational_intensity": 14445.2, "ridge_point": 295.22, "regime": "compute-bound", "roofline_tflops_bound": 989.0, "latency_ms": 265.06 }, "decode": { "phase": "decode_step", "flops": 128000000000.0, "hbm_bytes": 18281701376.0, "operational_intensity": 7.0, "ridge_point": 295.22, "regime": "bandwidth-bound", "roofline_tflops_bound": 23.46, "latency_ms": 5.457 }, "ttft_ms": 265.06, "tpot_ms": 5.457, "decode_tokens_per_sec": 1466.01, "end_to_end_tokens_per_sec": 1236.27, "tokens_per_dollar": 1390804.63 }, { "gpu": "example-large-memory", "memory_gb": 141.0, "required_memory_gb": 20.42, "fits_single_gpu": true, "ridge_point_flops_per_byte": 206.04, "prefill": { "phase": "prefill", "flops": 262144000000000.0, "hbm_bytes": 18147483648.0, "operational_intensity": 14445.2, "ridge_point": 206.04, "regime": "compute-bound", "roofline_tflops_bound": 989.0, "latency_ms": 265.06 }, "decode": { "phase": "decode_step", "flops": 128000000000.0, "hbm_bytes": 18281701376.0, "operational_intensity": 7.0, "ridge_point": 206.04, "regime": "bandwidth-bound", "roofline_tflops_bound": 33.61, "latency_ms": 3.809 }, "ttft_ms": 265.06, "tpot_ms": 3.809, "decode_tokens_per_sec": 2100.29, "end_to_end_tokens_per_sec": 1656.48, "tokens_per_dollar": 1569298.89 }, { "gpu": "example-wide-bandwidth", "memory_gb": 192.0, "required_memory_gb": 20.42, "fits_single_gpu": true, "ridge_point_flops_per_byte": 281.25, "prefill": { "phase": "prefill", "flops": 262144000000000.0, "hbm_bytes": 18147483648.0, "operational_intensity": 14445.2, "ridge_point": 281.25, "regime": "compute-bound", "roofline_tflops_bound": 2250.0, "latency_ms": 116.508 }, "decode": { "phase": "decode_step", "flops": 128000000000.0, "hbm_bytes": 18281701376.0, "operational_intensity": 7.0, "ridge_point": 281.25, "regime": "bandwidth-bound", "roofline_tflops_bound": 56.01, "latency_ms": 2.285 }, "ttft_ms": 116.51, "tpot_ms": 2.285, "decode_tokens_per_sec": 3501.09, "end_to_end_tokens_per_sec": 2929.13, "tokens_per_dollar": 1700786.9 } ], "qwen_32b_long_context": [ { "gpu": "example-small", "memory_gb": 48.0, "required_memory_gb": 85.33, "fits_single_gpu": false, "ridge_point_flops_per_byte": 418.98, "prefill": { "phase": "prefill", "flops": 4194304000000000.0, "hbm_bytes": 81179869184.0, "operational_intensity": 51666.8, "ridge_point": 418.98, "regime": "compute-bound", "roofline_tflops_bound": 362.0, "latency_ms": 11586.475 }, "decode": { "phase": "decode_step", "flops": 1024000000000.0, "hbm_bytes": 82253611008.0, "operational_intensity": 12.45, "ridge_point": 418.98, "regime": "bandwidth-bound", "roofline_tflops_bound": 10.76, "latency_ms": 95.201 }, "ttft_ms": 11586.48, "tpot_ms": 95.201, "decode_tokens_per_sec": 168.07, "end_to_end_tokens_per_sec": 136.0, "tokens_per_dollar": 315876.78 }, { "gpu": "example-fast", "memory_gb": 80.0, "required_memory_gb": 85.33, "fits_single_gpu": false, "ridge_point_flops_per_byte": 295.22, "prefill": { "phase": "prefill", "flops": 4194304000000000.0, "hbm_bytes": 81179869184.0, "operational_intensity": 51666.8, "ridge_point": 295.22, "regime": "compute-bound", "roofline_tflops_bound": 989.0, "latency_ms": 4240.954 }, "decode": { "phase": "decode_step", "flops": 1024000000000.0, "hbm_bytes": 82253611008.0, "operational_intensity": 12.45, "ridge_point": 295.22, "regime": "bandwidth-bound", "roofline_tflops_bound": 41.71, "latency_ms": 24.553 }, "ttft_ms": 4240.95, "tpot_ms": 24.553, "decode_tokens_per_sec": 651.65, "end_to_end_tokens_per_sec": 487.98, "tokens_per_dollar": 548978.69 }, { "gpu": "example-large-memory", "memory_gb": 141.0, "required_memory_gb": 85.33, "fits_single_gpu": true, "ridge_point_flops_per_byte": 206.04, "prefill": { "phase": "prefill", "flops": 4194304000000000.0, "hbm_bytes": 81179869184.0, "operational_intensity": 51666.8, "ridge_point": 206.04, "regime": "compute-bound", "roofline_tflops_bound": 989.0, "latency_ms": 4240.954 }, "decode": { "phase": "decode_step", "flops": 1024000000000.0, "hbm_bytes": 82253611008.0, "operational_intensity": 12.45, "ridge_point": 206.04, "regime": "bandwidth-bound", "roofline_tflops_bound": 59.76, "latency_ms": 17.136 }, "ttft_ms": 4240.95, "tpot_ms": 17.136, "decode_tokens_per_sec": 933.71, "end_to_end_tokens_per_sec": 630.28, "tokens_per_dollar": 597104.98 }, { "gpu": "example-wide-bandwidth", "memory_gb": 192.0, "required_memory_gb": 85.33, "fits_single_gpu": true, "ridge_point_flops_per_byte": 281.25, "prefill": { "phase": "prefill", "flops": 4194304000000000.0, "hbm_bytes": 81179869184.0, "operational_intensity": 51666.8, "ridge_point": 281.25, "regime": "compute-bound", "roofline_tflops_bound": 2250.0, "latency_ms": 1864.135 }, "decode": { "phase": "decode_step", "flops": 1024000000000.0, "hbm_bytes": 82253611008.0, "operational_intensity": 12.45, "ridge_point": 281.25, "regime": "bandwidth-bound", "roofline_tflops_bound": 99.59, "latency_ms": 10.282 }, "ttft_ms": 1864.13, "tpot_ms": 10.282, "decode_tokens_per_sec": 1556.12, "end_to_end_tokens_per_sec": 1150.85, "tokens_per_dollar": 668233.6 } ], "h100_batch_sweep": [ { "batch_size": 1, "prefill_intensity": 1015.48, "prefill_regime": "compute-bound", "decode_intensity": 0.99, "decode_regime": "bandwidth-bound", "decode_roofline_tflops_bound": 3.32, "tpot_ms": 4.819, "decode_tokens_per_sec": 207.51 }, { "batch_size": 8, "prefill_intensity": 7676.82, "prefill_regime": "compute-bound", "decode_intensity": 7.47, "decode_regime": "bandwidth-bound", "decode_roofline_tflops_bound": 25.02, "tpot_ms": 5.117, "decode_tokens_per_sec": 1563.42 }, { "batch_size": 32, "prefill_intensity": 25833.4, "prefill_regime": "compute-bound", "decode_intensity": 24.9, "decode_regime": "bandwidth-bound", "decode_roofline_tflops_bound": 83.41, "tpot_ms": 6.138, "decode_tokens_per_sec": 5213.42 }, { "batch_size": 128, "prefill_intensity": 63205.55, "prefill_regime": "compute-bound", "decode_intensity": 59.79, "decode_regime": "bandwidth-bound", "decode_roofline_tflops_bound": 200.29, "tpot_ms": 10.225, "decode_tokens_per_sec": 12518.34 }, { "batch_size": 256, "prefill_intensity": 83286.85, "prefill_regime": "compute-bound", "decode_intensity": 78.01, "decode_regime": "bandwidth-bound", "decode_roofline_tflops_bound": 261.33, "tpot_ms": 15.674, "decode_tokens_per_sec": 16332.78 } ] }