Intel has announced a new data centre graphics processing unit, or GPU, aimed at AI inference workloads, as well as a rack-scale reference design based on its earlier Gaudi 3 system, as it seeks a foothold in the growing AI infrastructure market.
Lip-Bu Tan, the new chief executive of the struggling chipmaker, said on Intel’s most recent earnings call that the company would focus its efforts on chips for AI inference and AI agents. Inference refers to running AI models rather than training them.
“We need to start by first understanding emerging and real AI workloads — then work backwards to design software, systems, and silicon to enable the best outcomes for those particular workloads,” Tan said, adding that Intel would work toward a “full-stack AI solution”.
Crescent Island
Intel said the new Crescent Island data centre GPU would be based on its Xe3P microarchitecture and would include up to 160GB of memory.
It said the chip would begin sampling with customers in the second half of next year.
Intel chief technology officer Sachin Katti said AI was shifting to real-time pervasive inference that would require heterogeneous systems capable of using chips from different vendors, depending on the task.
“Intel’s Xe architecture data center GPU will provide the efficient headroom customers need — and more value — as token volumes surge,” Katti said.
Intel has said it plans to shift to regular, annual GPU releases, similar to those from competitors such as market leader Nvidia.
The company’s Gaudi 3 rack-scale systems feature up to 64 GPUs per rack for real-time inferencing tasks.
Intel’s earlier Gaudi chip failed to achieve traction, while rival AMD has gained interest from major customers such as OpenAI, although it remains far behind Nvidia.
Turnaround efforts
Nvidia said in September it would invest $5bn in Intel, taking a roughly 4 percent stake in the company and becoming one of its largest shareholders as part of a deal that will see it working with Intel on future PC and data centre chips.
The investment followed a similar move by the US government to take a 10 percent stake in Intel.
Intel is also seeking to leverage its advanced chip manufacturing facilities to offer foundry services to third-party customers, a market that is currently dominated by the likes of Taiwan’s TSMC and South Korea’s Samsung Electronics.
