code large_stringlengths 193 899 | workload_type large_stringclasses 15
values | gpu_name large_stringclasses 12
values | gpu_features large_stringclasses 12
values | runtime_ms float64 0.01 4.35k | flops float64 30 21,045B | memory_bytes int64 160 13.2B | arithmetic_intensity float64 0 6.37k | dtype_bytes int64 2 4 | workload_params large_stringlengths 11 79 |
|---|---|---|---|---|---|---|---|---|---|
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 80GB | {"gpu_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} | 0.038811 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L4 | {"gpu_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} | 0.048713 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L40S | {"gpu_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} | 0.039054 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.032743 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.022749 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 SXM | {"gpu_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} | 0.038869 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.03448 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 64) x (64, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 64, dtype=torch.float16, device='cuda')
B = torch.randn(64, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.028085 | 33,554,432 | 1,056,768 | 31.751938 | 2 | {"M": 4096, "N": 64, "K": 64, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA T4 | {"gpu_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} | 0.059232 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA V100 | {"gpu_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} | 0.078369 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A10G | {"gpu_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} | 0.07088 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 40GB | {"gpu_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} | 0.051834 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 80GB | {"gpu_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} | 0.058311 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L4 | {"gpu_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} | 0.03864 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L40S | {"gpu_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} | 0.053308 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.065234 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.028118 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 SXM | {"gpu_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} | 0.059857 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.0322 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float32, device='cuda')
B = torch.randn(256, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.033519 | 134,217,728 | 5,308,416 | 25.283951 | 4 | {"M": 4096, "N": 64, "K": 256, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA T4 | {"gpu_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} | 0.041607 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA V100 | {"gpu_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} | 0.022916 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A10G | {"gpu_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} | 0.031256 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 40GB | {"gpu_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} | 0.037938 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 80GB | {"gpu_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} | 0.050333 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L4 | {"gpu_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} | 0.043545 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L40S | {"gpu_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} | 0.048593 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.029482 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.027258 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 SXM | {"gpu_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} | 0.041963 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.046994 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 256) x (256, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 256, dtype=torch.float16, device='cuda')
B = torch.randn(256, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.044212 | 134,217,728 | 2,654,208 | 50.567901 | 2 | {"M": 4096, "N": 64, "K": 256, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA T4 | {"gpu_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} | 0.073476 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA V100 | {"gpu_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} | 0.069294 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A10G | {"gpu_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} | 0.072437 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 40GB | {"gpu_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} | 0.038728 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 80GB | {"gpu_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} | 0.04722 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L4 | {"gpu_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} | 0.083185 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L40S | {"gpu_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} | 0.057002 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.080085 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.034632 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 SXM | {"gpu_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} | 0.032849 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.059925 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float32, device='cuda')
B = torch.randn(512, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.067636 | 268,435,456 | 9,568,256 | 28.054795 | 4 | {"M": 4096, "N": 64, "K": 512, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA T4 | {"gpu_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} | 0.061368 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA V100 | {"gpu_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} | 0.03741 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A10G | {"gpu_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} | 0.072164 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 40GB | {"gpu_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} | 0.04272 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA A100 80GB | {"gpu_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} | 0.053457 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L4 | {"gpu_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} | 0.076606 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA L40S | {"gpu_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} | 0.044823 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.042589 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.032387 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 SXM | {"gpu_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} | 0.058305 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.040655 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 512) x (512, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 512, dtype=torch.float16, device='cuda')
B = torch.randn(512, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchronize()
| matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.036839 | 268,435,456 | 4,784,128 | 56.109589 | 2 | {"M": 4096, "N": 64, "K": 512, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA T4 | {"gpu_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} | 0.109957 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA V100 | {"gpu_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} | 0.086867 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A10G | {"gpu_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} | 0.073136 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 40GB | {"gpu_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} | 0.058968 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 80GB | {"gpu_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} | 0.095945 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L4 | {"gpu_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} | 0.125291 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L40S | {"gpu_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} | 0.080264 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.067095 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.048724 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA H100 SXM | {"gpu_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} | 0.059571 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.083233 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float32, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.058954 | 536,870,912 | 18,087,936 | 29.681159 | 4 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA T4 | {"gpu_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} | 0.086616 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA V100 | {"gpu_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} | 0.061877 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A10G | {"gpu_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} | 0.076673 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 40GB | {"gpu_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} | 0.056016 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 80GB | {"gpu_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} | 0.024338 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L4 | {"gpu_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} | 0.066459 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L40S | {"gpu_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} | 0.072484 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.059103 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.044138 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA H100 SXM | {"gpu_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} | 0.043584 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.025128 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 1024) x (1024, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 1024, dtype=torch.float16, device='cuda')
B = torch.randn(1024, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.077138 | 536,870,912 | 9,043,968 | 59.362319 | 2 | {"M": 4096, "N": 64, "K": 1024, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA T4 | {"gpu_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} | 0.288867 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA V100 | {"gpu_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} | 0.110034 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A10G | {"gpu_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} | 0.123621 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 40GB | {"gpu_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} | 0.104475 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 80GB | {"gpu_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} | 0.117777 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L4 | {"gpu_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} | 0.124706 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L40S | {"gpu_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} | 0.064899 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.095274 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX 4090 | {"gpu_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} | 0.045896 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA H100 SXM | {"gpu_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} | 0.081939 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA H100 PCIe | {"gpu_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} | 0.077821 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float32, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float32, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX A6000 | {"gpu_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} | 0.102025 | 1,073,741,824 | 35,127,296 | 30.567164 | 4 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float32"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA T4 | {"gpu_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} | 0.163052 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA V100 | {"gpu_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} | 0.065338 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A10G | {"gpu_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} | 0.080602 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 40GB | {"gpu_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} | 0.068114 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA A100 80GB | {"gpu_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} | 0.050604 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L4 | {"gpu_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} | 0.097304 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA L40S | {"gpu_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} | 0.093905 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
import torch
def matmul_kernel(A, B):
# Matrix multiplication: (4096, 2048) x (2048, 64) -> (4096, 64)
C = torch.matmul(A, B)
return C
A = torch.randn(4096, 2048, dtype=torch.float16, device='cuda')
B = torch.randn(2048, 64, dtype=torch.float16, device='cuda')
C = matmul_kernel(A, B)
torch.cuda.synchroniz... | matmul | NVIDIA RTX 3090 | {"gpu_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} | 0.069664 | 1,073,741,824 | 17,563,648 | 61.134328 | 2 | {"M": 4096, "N": 64, "K": 2048, "dtype": "float16"} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.