Skip to main content

Intel Gaudi 3 (Habana Labs)

Overview

Intel Gaudi 3 was announced in April 2024 (Intel Vision Conference) as a comprehensive upgrade over Gaudi 2. Built on a 5nm process, it features 128GB HBM2e memory, 1,835 TFLOPS BF16 / FP8 compute, and 24× 200GbE RDMA integrated NICs. According to Intel's official data, Gaudi 3 is 50% faster than H100 in LLM training, with 50% higher inference throughput and 40% better energy efficiency.

Core Specifications

ItemSpecification
ArchitectureGaudi 3 (Habana 5th Gen)
Process NodeTSMC 5nm
Compute Engines64× TPC (8 per group, 8 groups)
Matrix Engines8× MME (64,000 parallel ops per MME)
Memory128 GB HBM2e
Memory Bandwidth3.67 TB/s (+50% vs Gaudi 2)
On-die SRAM96 MB (12.8 TB/s bandwidth)
BF16 Matrix1,835 TFLOPS (4× Gaudi 2)
FP8 Matrix1,835 TFLOPS
BF16 Vector28.7 TFLOPS
TDP900 W (air-cooled) / 1,200 W (liquid)
Network24× 200GbE RoCE (integrated NIC)
PCIeGen 5 x16
Form FactorOAM 2.0 (HL-325L) / PCIe Add-in Card (HL-338)

Gaudi 2 vs Gaudi 3

MetricGaudi 2Gaudi 3Improvement
Process7nm5nm
Memory96 GB128 GB+33%
Bandwidth2.46 TB/s3.67 TB/s+49%
BF16 Compute432 TFLOPS1,835 TFLOPS4.2×
Network24× 100GbE24× 200GbE
TDP600 W900 W+50%
PCIeGen 4Gen 5

Vendor Information

ItemDetails
ManufacturerIntel Corporation (Habana Labs)
Official Websitehttps://habana.ai
Product Pagehttps://habana.ai/products/gaudi3/
Whitepaperhttps://www.intel.com/content/www/us/en/content-details/817486/intel-gaudi-3-ai-accelerator-white-paper.html
ReleaseApril 2024 (OEM Q2, PCIe card Q4)

Key Features

  • Integrated 24× 200GbE RoCE: 1,200 GB/s bidirectional bandwidth
  • Open Ethernet standard: Avoids vendor lock-in
  • MLPerf performance validated
  • PyTorch optimized: Only 3 lines of code change needed to migrate from H100
  • vs H200: Llama 2 7B/70B, Falcon 180B inference 30% faster

Use Cases

  • LLM training and inference
  • Multimodal AI
  • Enterprise GenAI (cost-sensitive)
  • Large-scale distributed training