── pod # NVIDIA RTX A5000 | SMs=64 | power limit 230.00 W | max SM clock 2100 MHz | driver 595.91.07 | torch 2.14.1+cu130 FP32 matmul (IEEE, no TF32) 17.32 TFLOPS | 1328 MHz | 154.8 W | 79.6% of full rate @ that clock TF32 matmul 44.57 TFLOPS | 1395 MHz | 227.5 W | 97.5% of full rate @ that clock FP16 matmul, FP32 accum, n=8192 88.33 TFLOPS | 1380 MHz | 213.1 W | 97.7% of full rate @ that clock FP16 matmul, FP16 accum, n=8192 88.76 TFLOPS | 1395 MHz | 229.2 W | 97.1% of full rate @ that clock BF16 matmul (FP32 accum), n=8192 90.62 TFLOPS | 1418 MHz | 222.5 W | 97.6% of full rate @ that clock FP16 matmul, FP32 accum, n=16384 86.06 TFLOPS | 1335 MHz | 227.1 W | 98.4% of full rate @ that clock FP16 matmul, FP16 accum, n=16384 88.52 TFLOPS | 1380 MHz | 229.1 W | 97.9% of full rate @ that clock BF16 matmul (FP32 accum), n=16384 87.84 TFLOPS | 1395 MHz | 226.9 W | 96.1% of full rate @ that clock FP8: no hardware support (needs sm_89+), skipped INT8 int_mm (B col-major), n=8192 166.62 TOPS | 1365 MHz | 228.6 W | 93.1% of full rate @ that clock INT8 int_mm (B col-major), n=16384 164.27 TOPS | 1320 MHz | 227.4 W | 94.9% of full rate @ that clock FP64 matmul 0.42 TFLOPS | 1890 MHz | 137.5 W | 85.8% of full rate @ that clock DRAM copy bandwidth (1 GiB) 683.98 GB/s | 1785 MHz | 198.5 W | RESULT {"gpu": "NVIDIA RTX A5000", "sms": 64, "power_limit_w": 230.0, "max_sm_mhz": 2100.0, "driver": "595.91.07", "torch": "2.14.1+cu130", "results": [{"test": "FP32 matmul (IEEE, no TF32)", "key": "fp32", "value": 17.32, "unit": "TFLOPS", "sm_mhz": 1327.5, "power_w": 154.76999999999998, "pct_of_full_rate_at_clock": 79.6}, {"test": "TF32 matmul", "key": "tf32", "value": 44.57, "unit": "TFLOPS", "sm_mhz": 1395.0, "power_w": 227.47, "pct_of_full_rate_at_clock": 97.5}, {"test": "FP16 matmul, FP32 accum, n=8192", "key": "fp16", "value": 88.33, "unit": "TFLOPS", "sm_mhz": 1380.0, "power_w": 213.13, "pct_of_full_rate_at_clock": 97.7}, {"test": "FP16 matmul, FP16 accum, n=8192", "key": "fp16_acc16", "value": 88.76, "unit": "TFLOPS", "sm_mhz": 1395.0, "power_w": 229.2, "pct_of_full_rate_at_clock": 97.1}, {"test": "BF16 matmul (FP32 accum), n=8192", "key": "bf16", "value": 90.62, "unit": "TFLOPS", "sm_mhz": 1417.5, "power_w": 222.46, "pct_of_full_rate_at_clock": 97.6}, {"test": "FP16 matmul, FP32 accum, n=16384", "key": "fp16", "value": 86.06, "unit": "TFLOPS", "sm_mhz": 1335.0, "power_w": 227.05, "pct_of_full_rate_at_clock": 98.4}, {"test": "FP16 matmul, FP16 accum, n=16384", "key": "fp16_acc16", "value": 88.52, "unit": "TFLOPS", "sm_mhz": 1380.0, "power_w": 229.08, "pct_of_full_rate_at_clock": 97.9}, {"test": "BF16 matmul (FP32 accum), n=16384", "key": "bf16", "value": 87.84, "unit": "TFLOPS", "sm_mhz": 1395.0, "power_w": 226.92000000000002, "pct_of_full_rate_at_clock": 96.1}, {"test": "INT8 int_mm (B col-major), n=8192", "key": "int8", "value": 166.62, "unit": "TOPS", "sm_mhz": 1365.0, "power_w": 228.64, "pct_of_full_rate_at_clock": 93.1}, {"test": "INT8 int_mm (B col-major), n=16384", "key": "int8", "value": 164.27, "unit": "TOPS", "sm_mhz": 1320.0, "power_w": 227.41, "pct_of_full_rate_at_clock": 94.9}, {"test": "FP64 matmul", "key": "fp64", "value": 0.42, "unit": "TFLOPS", "sm_mhz": 1890.0, "power_w": 137.51, "pct_of_full_rate_at_clock": 85.8}, {"test": "DRAM copy bandwidth (1 GiB)", "key": "dram", "value": 683.98, "unit": "GB/s", "sm_mhz": 1785.0, "power_w": 198.45999999999998, "pct_of_full_rate_at_clock": null}]} ── pod # NVIDIA A40 | SMs=84 | power limit 300.00 W | max SM clock 1740 MHz | driver 595.91.07 | torch 2.14.1+cu130 FP32 matmul (IEEE, no TF32) 23.31 TFLOPS | 1395 MHz | 297.2 W | 77.7% of full rate @ that clock TF32 matmul 58.19 TFLOPS | 1440 MHz | 299.7 W | 94.0% of full rate @ that clock FP16 matmul, FP32 accum, n=8192 104.33 TFLOPS | 1282 MHz | 295.3 W | 94.6% of full rate @ that clock FP16 matmul, FP16 accum, n=8192 110.19 TFLOPS | 1388 MHz | 291.4 W | 92.3% of full rate @ that clock BF16 matmul (FP32 accum), n=8192 106.85 TFLOPS | 1298 MHz | 292.3 W | 95.7% of full rate @ that clock FP16 matmul, FP32 accum, n=16384 118.11 TFLOPS | 1380 MHz | 299.2 W | 99.5% of full rate @ that clock FP16 matmul, FP16 accum, n=16384 117.85 TFLOPS | 1410 MHz | 300.4 W | 97.2% of full rate @ that clock BF16 matmul (FP32 accum), n=16384 122.04 TFLOPS | 1440 MHz | 299.2 W | 98.5% of full rate @ that clock FP8: no hardware support (needs sm_89+), skipped INT8 int_mm (B col-major), n=8192 218.07 TOPS | 1440 MHz | 300.7 W | 88.0% of full rate @ that clock INT8 int_mm (B col-major), n=16384 227.06 TOPS | 1395 MHz | 300.1 W | 94.6% of full rate @ that clock FP64 matmul 0.50 TFLOPS | 1740 MHz | 134.2 W | 85.4% of full rate @ that clock DRAM copy bandwidth (1 GiB) 562.97 GB/s | 1740 MHz | 212.9 W | RESULT {"gpu": "NVIDIA A40", "sms": 84, "power_limit_w": 300.0, "max_sm_mhz": 1740.0, "driver": "595.91.07", "torch": "2.14.1+cu130", "results": [{"test": "FP32 matmul (IEEE, no TF32)", "key": "fp32", "value": 23.31, "unit": "TFLOPS", "sm_mhz": 1395.0, "power_w": 297.15999999999997, "pct_of_full_rate_at_clock": 77.7}, {"test": "TF32 matmul", "key": "tf32", "value": 58.19, "unit": "TFLOPS", "sm_mhz": 1440.0, "power_w": 299.71000000000004, "pct_of_full_rate_at_clock": 94.0}, {"test": "FP16 matmul, FP32 accum, n=8192", "key": "fp16", "value": 104.33, "unit": "TFLOPS", "sm_mhz": 1282.5, "power_w": 295.315, "pct_of_full_rate_at_clock": 94.6}, {"test": "FP16 matmul, FP16 accum, n=8192", "key": "fp16_acc16", "value": 110.19, "unit": "TFLOPS", "sm_mhz": 1387.5, "power_w": 291.375, "pct_of_full_rate_at_clock": 92.3}, {"test": "BF16 matmul (FP32 accum), n=8192", "key": "bf16", "value": 106.85, "unit": "TFLOPS", "sm_mhz": 1297.5, "power_w": 292.315, "pct_of_full_rate_at_clock": 95.7}, {"test": "FP16 matmul, FP32 accum, n=16384", "key": "fp16", "value": 118.11, "unit": "TFLOPS", "sm_mhz": 1380.0, "power_w": 299.2, "pct_of_full_rate_at_clock": 99.5}, {"test": "FP16 matmul, FP16 accum, n=16384", "key": "fp16_acc16", "value": 117.85, "unit": "TFLOPS", "sm_mhz": 1410.0, "power_w": 300.44, "pct_of_full_rate_at_clock": 97.2}, {"test": "BF16 matmul (FP32 accum), n=16384", "key": "bf16", "value": 122.04, "unit": "TFLOPS", "sm_mhz": 1440.0, "power_w": 299.2, "pct_of_full_rate_at_clock": 98.5}, {"test": "INT8 int_mm (B col-major), n=8192", "key": "int8", "value": 218.07, "unit": "TOPS", "sm_mhz": 1440.0, "power_w": 300.7, "pct_of_full_rate_at_clock": 88.0}, {"test": "INT8 int_mm (B col-major), n=16384", "key": "int8", "value": 227.06, "unit": "TOPS", "sm_mhz": 1395.0, "power_w": 300.15, "pct_of_full_rate_at_clock": 94.6}, {"test": "FP64 matmul", "key": "fp64", "value": 0.5, "unit": "TFLOPS", "sm_mhz": 1740.0, "power_w": 134.24, "pct_of_full_rate_at_clock": 85.4}, {"test": "DRAM copy bandwidth (1 GiB)", "key": "dram", "value": 562.97, "unit": "GB/s", "sm_mhz": 1740.0, "power_w": 212.93, "pct_of_full_rate_at_clock": null}]} 2 ok, 0 failed