| { |
| "NVIDIA_T4": { |
| "name": "NVIDIA T4", |
| "cuda_cores": 2560, |
| "tensor_cores": 320, |
| "memory_gb": 16, |
| "memory_bandwidth_gbps": 320, |
| "base_clock_mhz": 585, |
| "boost_clock_mhz": 1590, |
| "sm_count": 40, |
| "fp32_tflops": 8.1, |
| "fp16_tflops": 65, |
| "tdp_watts": 70, |
| "compute_capability": 7.5, |
| "l2_cache_mb": 4 |
| }, |
| "NVIDIA_V100": { |
| "name": "NVIDIA V100", |
| "cuda_cores": 5120, |
| "tensor_cores": 640, |
| "memory_gb": 32, |
| "memory_bandwidth_gbps": 900, |
| "base_clock_mhz": 1230, |
| "boost_clock_mhz": 1530, |
| "sm_count": 80, |
| "fp32_tflops": 15.7, |
| "fp16_tflops": 125, |
| "tdp_watts": 300, |
| "compute_capability": 7.0, |
| "l2_cache_mb": 6 |
| }, |
| "NVIDIA_A10G": { |
| "name": "NVIDIA A10G", |
| "cuda_cores": 9216, |
| "tensor_cores": 288, |
| "memory_gb": 24, |
| "memory_bandwidth_gbps": 600, |
| "base_clock_mhz": 885, |
| "boost_clock_mhz": 1695, |
| "sm_count": 80, |
| "fp32_tflops": 31.2, |
| "fp16_tflops": 62.5, |
| "tdp_watts": 150, |
| "compute_capability": 8.6, |
| "l2_cache_mb": 6 |
| }, |
| "NVIDIA_A100_40GB": { |
| "name": "NVIDIA A100 40GB", |
| "cuda_cores": 6912, |
| "tensor_cores": 432, |
| "memory_gb": 40, |
| "memory_bandwidth_gbps": 1555, |
| "base_clock_mhz": 765, |
| "boost_clock_mhz": 1410, |
| "sm_count": 108, |
| "fp32_tflops": 19.5, |
| "fp16_tflops": 312, |
| "tdp_watts": 400, |
| "compute_capability": 8.0, |
| "l2_cache_mb": 40 |
| }, |
| "NVIDIA_A100_80GB": { |
| "name": "NVIDIA A100 80GB", |
| "cuda_cores": 6912, |
| "tensor_cores": 432, |
| "memory_gb": 80, |
| "memory_bandwidth_gbps": 2039, |
| "base_clock_mhz": 765, |
| "boost_clock_mhz": 1410, |
| "sm_count": 108, |
| "fp32_tflops": 19.5, |
| "fp16_tflops": 312, |
| "tdp_watts": 400, |
| "compute_capability": 8.0, |
| "l2_cache_mb": 40 |
| }, |
| "NVIDIA_L4": { |
| "name": "NVIDIA L4", |
| "cuda_cores": 7424, |
| "tensor_cores": 232, |
| "memory_gb": 24, |
| "memory_bandwidth_gbps": 300, |
| "base_clock_mhz": 795, |
| "boost_clock_mhz": 2040, |
| "sm_count": 58, |
| "fp32_tflops": 30.3, |
| "fp16_tflops": 121, |
| "tdp_watts": 72, |
| "compute_capability": 8.9, |
| "l2_cache_mb": 48 |
| }, |
| "NVIDIA_L40S": { |
| "name": "NVIDIA L40S", |
| "cuda_cores": 18176, |
| "tensor_cores": 568, |
| "memory_gb": 48, |
| "memory_bandwidth_gbps": 864, |
| "base_clock_mhz": 1110, |
| "boost_clock_mhz": 2520, |
| "sm_count": 142, |
| "fp32_tflops": 91.6, |
| "fp16_tflops": 183.2, |
| "tdp_watts": 350, |
| "compute_capability": 8.9, |
| "l2_cache_mb": 96 |
| }, |
| "NVIDIA_RTX_3090": { |
| "name": "NVIDIA RTX 3090", |
| "cuda_cores": 10496, |
| "tensor_cores": 328, |
| "memory_gb": 24, |
| "memory_bandwidth_gbps": 936, |
| "base_clock_mhz": 1395, |
| "boost_clock_mhz": 1695, |
| "sm_count": 82, |
| "fp32_tflops": 35.6, |
| "fp16_tflops": 71, |
| "tdp_watts": 350, |
| "compute_capability": 8.6, |
| "l2_cache_mb": 6 |
| }, |
| "NVIDIA_RTX_4090": { |
| "name": "NVIDIA RTX 4090", |
| "cuda_cores": 16384, |
| "tensor_cores": 512, |
| "memory_gb": 24, |
| "memory_bandwidth_gbps": 1008, |
| "base_clock_mhz": 2235, |
| "boost_clock_mhz": 2520, |
| "sm_count": 128, |
| "fp32_tflops": 82.6, |
| "fp16_tflops": 165.2, |
| "tdp_watts": 450, |
| "compute_capability": 8.9, |
| "l2_cache_mb": 72 |
| }, |
| "NVIDIA_H100_SXM": { |
| "name": "NVIDIA H100 SXM", |
| "cuda_cores": 16896, |
| "tensor_cores": 528, |
| "memory_gb": 80, |
| "memory_bandwidth_gbps": 3350, |
| "base_clock_mhz": 1095, |
| "boost_clock_mhz": 1830, |
| "sm_count": 132, |
| "fp32_tflops": 67, |
| "fp16_tflops": 989, |
| "tdp_watts": 700, |
| "compute_capability": 9.0, |
| "l2_cache_mb": 50 |
| }, |
| "NVIDIA_H100_PCIe": { |
| "name": "NVIDIA H100 PCIe", |
| "cuda_cores": 14592, |
| "tensor_cores": 456, |
| "memory_gb": 80, |
| "memory_bandwidth_gbps": 2039, |
| "base_clock_mhz": 1095, |
| "boost_clock_mhz": 1620, |
| "sm_count": 114, |
| "fp32_tflops": 48, |
| "fp16_tflops": 756, |
| "tdp_watts": 350, |
| "compute_capability": 9.0, |
| "l2_cache_mb": 50 |
| }, |
| "NVIDIA_RTX_A6000": { |
| "name": "NVIDIA RTX A6000", |
| "cuda_cores": 10752, |
| "tensor_cores": 336, |
| "memory_gb": 48, |
| "memory_bandwidth_gbps": 768, |
| "base_clock_mhz": 1410, |
| "boost_clock_mhz": 1860, |
| "sm_count": 84, |
| "fp32_tflops": 38.7, |
| "fp16_tflops": 77.4, |
| "tdp_watts": 300, |
| "compute_capability": 8.6, |
| "l2_cache_mb": 6 |
| } |
| } |
