Intel Gaudi 2 (Habana Labs)
Overview
Intel Gaudi 2, developed by Intel-acquired Habana Labs, is a dedicated accelerator for deep learning training and inference. Its biggest feature is the integrated 24x 100GbE RDMA NICs, eliminating the need for separate InfiniBand or Ethernet cards for large-scale distributed training clusters. Gaudi 2 performance rivals NVIDIA H100, at a better price.
Core Specifications
| Item | Spec |
|---|
| Architecture | Gaudi 2 (Habana custom) |
| Process | TSMC 7nm |
| Compute Engines | 24x TPC (Tensor Processor Core) |
| Matrix Engines | 2x MME (Matrix Multiplication Engine) |
| Memory | 96 GB HBM2e |
| Memory Bandwidth | 2.46 TB/s |
| FP32 | N/A (inference path) |
| BF16/FP16 | 432 TFLOPS |
| FP8 | 865 TFLOPS (AMD estimate) |
| INT8 | 865 TOPS |
| TDP | 600 W |
| Networking | 24x 100GbE RDMA (integrated NIC) |
| PCIe | Gen 4 x16 |
| Form Factor | OAM (HL-225H) / PCIe |
Key Features
- Integrated 24x 100GbE NICs: No need for separate InfiniBand
- TPC + MME heterogeneous compute: Flexible matrix and vector operations
- OpenEthernet standard: Avoids vendor lock-in
- MLPerf benchmark validated: Strong H100 alternative
Use Cases
- LLM training (tens to hundreds of billions of parameters)
- Large-scale distributed training clusters
- Multimodal AI