Intel plans to market its own Gaudi accelerators very competitively compared to those from Nvidia. In this way, the company hopes to gain a significant share of the AI ​​​​scene.

When you say AI, you say Nvidia, and Intel regrets that. CEO Pat Gelsinger wants to end the dominance of Nvidia accelerators in AI training and inference at Computex. The company’s fun accelerators are supposed to ensure this. Or, more precisely: that’s what the prices of these accelerators have to achieve.

Gelsinger says a single Gaudi 3 accelerator sells for about $15,000, half of what Nvidia charges for its Hopper GPUs. Moreover, available Hopper GPUs are being snapped up faster than they roll off the production line, mostly by hyperscalers and a few other large companies.

Competing with H100

Intel is positioning its Gaudi 3 accelerator as competitive with the Nvidia H100. Gaudi would even outperform Hopper in several relevant benchmarks. In addition, Intel is relying on more open standards for the chip, such as Ethernet. Nvidia, in turn, combines its accelerators with its own interconnects in servers.

In terms of performance, Intel is one generation behind Nvidia. After all, the Nvidia Hopper H100 already has a successor in the form of the Nvidia Blackwell B100 (and B200 and GB200). These monstrous accelerators consume between 700 and 1,200 watts and are the most powerful of their kind. The GPUs will be available this year, but it is unclear in what quantities.

However, Intel doesn’t need to release the most powerful chips to be relevant. Gaudi 3 seems to be powerful even for the most demanding applications, with a cheaper price and presumably higher availability. There is potential there.