Code for Tensor Core Micro Benchmarking for PIC2 MEEC
- Half Precision
- FMA benchmark
- Compute Roof -> FLOPs
- Alternate Formats
- Bf16
- Tf32
- Single Precision
- Double Precision
- Integer
| Inputs | Accumulators | Shape | FMA/clk/SM | FLOPs |
|---|---|---|---|---|
| FP16 | FP16 | m16n8k8 | 511.438 | |
| FP16 | FP16 | m16n8k16 | 511.820 | |
| FP16 | FP16 | m8n8k4ª | 1.0093 | |
| FP16 | FP32 | m16n8k8 | 255.675 | |
| FP16 | FP32 | m16n8k16 | 255.942 | |
| FP16 | FP32 | m8n8k4ª | 1.0441 |
Notes: ª - "mma.sync.m8n8k4 is optimized for target architecture sm_70 and may have substantially reduced performance on other target architectures." 1