Intel Unveils Gaudi 3 AI Accelerator with 128GB HBM2e and 1.8TB/s Networking
Intel announced Gaudi 3 AI accelerator featuring 64 Tensor Processor Cores and 24 integrated 200GbE networking ports for open AI training and inference.
Verified State Diff
Comparison Mode:
- Previous State
Gaudi 2 was constrained to 96GB memory and standard PCIe interconnect throughput.
+ Verified New State
Gaudi 3 delivers 128GB HBM2e at 3.7 TB/s with integrated 200GbE RDMA Ethernet fabric.
Impact & Verification Analysis
WHO IS AFFECTED
Enterprise AI clusters, private cloud infrastructure, and model training teams.
WHY IT MATTERS
Provides open Ethernet-based scaling for large AI clusters without proprietary NVLink switching.
Full Fact Overview
Gaudi 3 provides up to 50% faster time-to-train on Llama 2 70B compared to Nvidia H100 while offering substantial total-cost-of-ownership savings.
Multi-Source Evidence Chain (1)
TRACKED ENTITY
Explore all historical Intel changes