Intel has unveiled its latest Gaudi 3 AI accelerator, boasting up to 128GB of HBM2e memory and a memory bandwidth of up to 3.7TB/s. How does it stack up against NVIDIA’s Hopper H100 80GB AI GPU? Well, the key difference lies in cost: Intel’s Gaudi 3 AI accelerator is priced at $16,000 per unit with 128GB of HBM2e memory, while NVIDIA’s Hopper H100 AI GPU with 80GB of HBM3 memory costs over $30,000. In bulk purchases, a single Gaudi 2 AI accelerator costs around $8,125, while the Gaudi 3 AI accelerator comes in at $15,650.
According to Tirias Research analyst Jim McGregor, Intel is aiming to capitalize on the lucrative AI chip market, where NVIDIA currently holds a dominant position with its Hopper H100 80GB AI GPU. NVIDIA’s CEO, Jensen Huang, has hinted at the release of the next-generation Blackwell B200 AI GPUs, which are expected to outperform both Intel’s Gaudi 3 and NVIDIA’s own H100 AI GPUs. With the Blackwell B200 AI GPUs set to enter the market, prices for the Hopper H100 AI GPUs may drop below the current average of $30,000 per unit.
As Intel faces stiff competition in the AI market, not only from NVIDIA but also from custom, in-house designed AI processors used by cloud service providers like Amazon Web Services and Google, the company is striving to gain market share with its Gaudi 3 AI accelerators. Despite its efforts, Intel remains the underdog in the AI market, facing challenges in competing with both established players and emerging technologies.
Taking a closer look at the technical specifications of the Gaudi 3 AI accelerator, each chip features 32 Tensor Cores for a total of 64 Tensor Cores. With 48MB of SRAM per chip and a total of 96MB of SRAM per package, the Gaudi 3 AI accelerator offers high bandwidth compatibility with its HBM memory. The 128GB of HBM2e memory delivers impressive memory bandwidth of up to 3.7TB/s.
In terms of design, the Gaudi 3 AI accelerator is available in both PCIe form factor (HL-388) and OAM form (HL-328/325L/335). The PCIe version utilizes the PCIe 5.0 interface with full PCIe 5.0 x16 lanes, boasting a power consumption of 450W to 600W. The OAM version, designed for air-cooled servers, features a TDP ranging from 450W to 900W, while the water-cooled version can reach up to 900W TDP.
In conclusion, Intel’s Gaudi 3 AI accelerator presents a competitive option in the AI chip market, offering high memory capacity and bandwidth at a relatively lower cost compared to NVIDIA’s Hopper H100 AI GPU. However, the company faces significant challenges in establishing itself as a major player in the AI sector, with tough competition from industry giants and custom-designed processors used by major cloud service providers.
Article Source
https://www.tweaktown.com/news/98977/intel-discounting-new-gaudi-3-ai-accelerator-16k-against-nvidia-h100-gpu-for-30k/index.html
