Skip to main content

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

ItemSpec
ArchitectureGaudi 2 (Habana custom)
ProcessTSMC 7nm
Compute Engines24x TPC (Tensor Processor Core)
Matrix Engines2x MME (Matrix Multiplication Engine)
Memory96 GB HBM2e
Memory Bandwidth2.46 TB/s
FP32N/A (inference path)
BF16/FP16432 TFLOPS
FP8865 TFLOPS (AMD estimate)
INT8865 TOPS
TDP600 W
Networking24x 100GbE RDMA (integrated NIC)
PCIeGen 4 x16
Form FactorOAM (HL-225H) / PCIe

Vendor Information

ItemDetails
ManufacturerIntel Corporation (Habana Labs)
Websitehttps://habana.ai
Product Pagehttps://habana.ai/products/gaudi2/
Driver/SDKhttps://developer.habana.ai/downloads/
SynapseAI SDKhttps://developer.habana.ai/

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