Intel is making another push into the artificial intelligence chip market, this time focusing on inference rather than training, a segment where Nvidia currently holds significant dominance. The company's previous attempt, the Gaudi series of GPUs, saw underwhelming sales, leading to the cancellation of its successor, Falcon Shores.
This new initiative, reportedly codenamed 'Crescent Island,' adopts a more strategic approach. According to Kevork Kechichian, Intel's general manager of its data center group, the focus is on data centers that handle AI inference. Inference involves processing tasks using already trained AI models, a less demanding application than the training process that requires substantial computational power.
To this end, Crescent Island GPUs are designed to eschew the need for large amounts of expensive High Bandwidth Memory (HBM). Instead, they will utilize more affordable LPDDR5X memory and are designed for simpler air-cooling systems, reducing overall costs and complexity compared to the liquid-cooled solutions often required for high-performance training chips. This approach could also potentially alleviate some pressure on the global memory market.
Nvidia is also exploring the inference market, notably through a collaboration with Groq to develop a new chip that combines a language accelerator with its Rubin platform. However, official launch dates for these inference-focused products from major chip manufacturers, including Intel's Crescent Island, have not yet been disclosed. Intel has only indicated a potential release sometime later this year.
Intel's initial foray into AI hardware with Gaudi GPUs, despite appearing competitive on paper and offered at attractive price points, struggled against the established presence of Nvidia's Hopper and Blackwell architectures in the AI training sector. The shift towards inference represents a deliberate pivot to find a more accessible segment of the rapidly growing AI hardware landscape.