add: gaudi3 benchmark v2 results
MAMF results for gaudi3.
| | bf16 | fp8 |
| --: | :-- | :-- |
| MAMF | 1243.0 TFLOPS @ 16384x4096x768 | 1289.5 TLOPS @ 16640x1536x3072 |
| spec | 1677 TFLOPS | 1677 TFLOPS |
| util | 74.1% | 76.9% |
- [zip file for full logs](https://github.com/user-attachments/files/22069863/gaudi3_mnk_0_20480_256.zip)
Note that results are obtained from MNK (0, 20480, 256) sweep, better results might exist.
Log contains some anomaly such as TFLOPS far over the spec, this appears to stem from occasional instability in the Synapse event API. However, for the MAMF figure I verified correctness by analyzing the device trace for the given workload, and it looked legitimate.
**_I've also updated the command for current clock and theoretical clock for intel gaudi software._**
With these commands, you can observe the clock rate dropping during computation. The drop is especially severe for BF16 matmul compared to FP8 matmul, which explains the relatively low TFLOPS utilization. This is reasonable, since BF16 computation performs much more work within the same clock latency, given that Gaudi3 has identical theoretical throughput for both BF16 and FP8.
Please refer to #115 for more details.
合并状态:已合并 合并于 2025-09-01 关闭于 2025-09-01 0 条评论