Intel Gaudi is Intel's AI accelerator line, aimed at workloads such as large language models, multimodal models and enterprise RAG. The latest generation, Gaudi 3, is offered as a PCIe Gen5 card.
What Intel says about Gaudi 3
On Intel's Gaudi page, Intel compares Gaudi 3 with Gaudi 2: up to 2x AI compute at FP8, up to 4x at BF16, and up to 2x network bandwidth. It highlights standard Ethernet networking and, in a footnote, lists 1200 GB/s of open-standard RoCE for Gaudi 3 against 900 GB/s of closed NVLink for an NVIDIA H100. Those are Intel's figures, so verify them on your workload.
Software comes as Intel Gaudi software with PyTorch integration, Hugging Face resources and tools for porting GPU-based models. See whether Gaudi supports PyTorch.
More depth
The full comparison is in Intel Gaudi 3 vs NVIDIA, and the rest of the field is in NVIDIA's main competitors in AI.
