Intel has officially launched its Gaudi 3 accelerator, designed for AI workloads, aiming to challenge Nvidia’s popular H100 and H200 GPUs. While Gaudi 3 may not match Nvidia’s GPUs in raw speed, Intel is betting on its lower price point and reduced total cost of ownership (TCO) to attract customers.
The Gaudi 3 packs significant upgrades, featuring dual chipsets with 64 tensor processor cores (TPCs), 96MB of on-chip SRAM boasting a bandwidth of 19.2TB/s, and eight matrix multiplication engines (MMEs). Notably, it also integrates 128GB of HBM2E memory, offering a massive 3.67TB/s memory bandwidth, alongside 24 200 GbE network interfaces, making it a strong contender in AI and high-performance computing (HPC) tasks.
Compared to its predecessor, Gaudi 2, which offered 24 TPCs and 96GB of memory, Gaudi 3 represents a giant leap. However, Intel has streamlined its support to focus on FP8 matrix operations and BFloat16 vector computations, dropping FP32, TF32, and FP16 support.
With its mix of high-performance specs and cost-efficiency, Gaudi 3 is shaping up to be a competitive alternative in the AI hardware space, appealing to organizations focused on balancing performance with affordability.
All Comments (0)