Intel Gaudi 3 AI accelerator is faster than NVIDIA's current-gen H100 AI GPU

Intel has had a lot to talk about lately; at its recent AI Everywhere event, the company unveiled its new 5th Gen Xeon "Emerald Rapids" CPU platform, the new Core Ultra branding and their Meteor Lake CPUs, as well as its next-gen AI accelerator: Gaudi 3.

4

But now we've got reports from Korean media that Intel's upcoming Gaudi 3 AI accelerator could compete with NVIDIA's current-gen H100 AI GPU and AMD's just-announced Instinct MI300X AI accelerator. Intel CEO Pat Gelsinger said: "Our Gaudi roadmap remains on track with Gaudi3 out of the fab, now in packaging and expected to launch next year. In 2025, Falcon Shores brings our GPU and Gaudi capabilities into a single product".

We know that the new Intel Gaudi 3 AI accelerator will have 1.5x more performance than the previous-gen Gaudi 2, as well as 4x more BFloat16 performance, twice the compute power, and a 50% bump in memory capacity from 96GB on Gaudi 2 to a much better 144GB on Gaudi 3. We should expect Intel to be using HBM3 or the faster HBM3e memory standard, but that still needs to be confirmed.

• Read more: Intel CEO swipes at NVIDIA on AI: 'entire industry is motivated to eliminate the CUDA market'

We do know that Intel will have its Gaudi 3 AI accelerator fabbed on TSMC's 5nm process node, similar to NVIDIA's current-gen H100 AI GPU -- with NVIDIA using a custom 4N process node from TSMC, while AMD's new Instinct MI300X is made up from both 5nm and 6nm process nodes by TSMC.

4

• Read more: AMD MI300X vs NVIDIA H100 battle, AMD says it has the 'performance advantage'
• Read more: AMD Instinct MI300X: AI accelerator with 192GB HBM3 at 5.3TB/sec bandwidth
• Read more: NVIDIA H200 AI GPU: up to 141GB of HBM3e with 4.8TB/sec bandwidth
• Read more: NVIDIA's next-gen GB200 'Blackwell' GPU listed on its 2024 data center roadmap
• Read more: NVIDIA AI GPU shipments expected to surge 150% year-over-year in 2024