Yes. Intel's Gaudi documentation says the Gaudi software "is integrated with PyTorch and optimized for Intel Gaudi AI accelerator", and runs in the background so developers keep a familiar PyTorch workflow.
Running CUDA-targeted code
For models written for GPUs, Intel ships a toolkit for porting PyTorch models. It is a preinstalled Python package that maps calls from torch.cuda, GPU-related Torch parameters, Apex and pynvml to Gaudi equivalents, so a model runs "with fewer modifications."
Its documented limits:
- It does not translate calls to third-party programs such as
nvccornvidia-smi; those need manual changes. - It "does not optimize the performance of the model, so further modifications may be required."
- Apex is not preinstalled in Gaudi containers.
- Gaudi favours bfloat16 over float16.
Next
For the hardware, see what Intel Gaudi is. The full NVIDIA comparison is in Intel Gaudi 3 vs NVIDIA, and the field overview is NVIDIA's main competitors in AI.
