Intel accidentally revealed that it’s preparing a Jaguar Shores AI accelerator to follow Falcon Shores

Intel announced the new Jaguar Shores AI accelerator, which is being prepared as a successor to Falcon Shores, mentioning it in a presentation during a technical workshop at the SC2024 conference. The presentation was dedicated to Gaudi chips produced by the company’s division, Habana Labs, reports the HPCwire resource. According to the source, the mention of the next generation chip in the presentation could have been accidental.

Falcon Shores is expected to enter production in 2025. Also next year, the Gaudi 3 AI accelerator, introduced back in February 2023, will go on mass sale. Otherwise, Intel prefers not to disclose details about its plans to release AI chips. In comparison, NVIDIA and AMD have already announced their GPU roadmaps with schedules extending all the way to 2026-2027.

Image Source: Intel

Last August, Intel told HPCwire that it was working on the Falcon Shores 2 chip, which is planned for release in 2026. “We have a simplified roadmap as we combine our GPUs and accelerators into a single offering,” CEO Pat Gelsinger explained at the time.

Since then, Intel’s financial position has deteriorated significantly, but the company continues to develop new AI accelerators. It’s unclear whether the Jaguar Shores will be a GPU or an ASIC, but Intel’s chip naming logic suggests it’s a next-generation GPU.

At the moment, Intel has lost the AI ​​training market to NVIDIA and AMD, focusing its efforts on inference using Gaudi AI accelerators. It is likely that Jaguar Shores will also be focused on inference, which Gelsinger identified as a larger and more promising market. However, to catch up with competitors NVIDIA and AMD that have moved ahead, Jaguar Shores must become a truly breakthrough product, HPCwire believes.

«Our investments in AI will complement and leverage our x86-based solutions, with a focus on enterprise, cost-effective data inference. Our roadmap for Falcon Shores remains unchanged,” an Intel spokeswoman told HPCwire several months ago.