It’s a long wait for customers hanging on for Intel’s next-generation GPU, Falcon Shores, which will be released in late 2025.
“Then we have a rich, a very aggressive cadence of Falcon Shores products following that,” said Intel CEO Pat Gelsinger when answering a question on the product roadmap in a financial earnings call on April 25.
Intel had previously not provided a proper release date for Falcon Shores. This situation most likely means the GPU will become broadly available in 2026.
Falcon Shores is now being redesigned by Intel after a change in market dynamics. Sources familiar with Intel’s plans said that the overnight emergence of generative AI dried up demand for the original Falcon Shares, which integrated GPU and CPU on a single chip.
Customers were interested in discrete GPUs and AI chips, which is why the company prioritized the Gaudi AI processor, which is Intel’s answer to Nvidia’s GPUs.
“We’ll be bringing that late next year when Falcon Shores, when we combine the great … performance of Gaudi 3 with a fully programmable architecture and all of that comes together with Falcon Shores,” Gelsinger said during the earnings call.
Intel is still in the early stages of the AI market, with Gaudi 3 giving some momentum. Intel made major Gaudi announcements, including availability and customer wins like Dell and Naver.
“We now expect over $500 million in accelerated revenue in the second half of 2024 and with increasing momentum into 2025 based on Gaudi 3’s vastly superior TCO as well as our own expanding supply,” Gelsinger said.
By the end of this year, Intel plans to add the Gaudi chip as a PCI-E accelerator that works alongside the server chips.
Intel’s next-generation server chip, Xeon 6 (codenamed Granite Rapids), will be released in the third quarter of this year.
The dense-server Sierra Forest server will also come out later this year. The server chip is based on low-power E-core parts.
“Next year is Clearwater Forest, the second generation of the E-core part, the leadership position on 18A in the server market, a very strong product for us,” Gelsinger said.
Intel is still an early entrant into the generative AI market, which is dominated by Nvidia. However, both companies are taking distinctly unique approaches to GPUs and the AI market.
Nvidia is moving toward highly integrated offerings, with its GPUs paired with its CPUs and proprietary networking technology. The company is steering customers to its proprietary CUDA software stack.
Nvidia also ships discrete GPUs, and its hardware works with other parallel programming frameworks.
Intel is taking a disaggregated approach. It is separating its CPUs, GPUs, and AI chips and relying on open technologies. The chips are lined up to work in tandem, and customers aren’t forced into buying the entire stack.
Networking is a big aspect of AI, and Intel is pushing its software and hardware stack on Ethernet with the Ultra Ethernet Consortium. The consortium includes high-speed communications protocols with reliable and secure links in line with AI requirements. Ethernet already supports many data-center AI operations.
“Ultra Ethernet consortium standardizing on scale up and scale out to Ethernet, increasing work for abstract levels of AI development with PyTorch, and the embrace of the open platform for enterprise AI that we rolled out,” Gelsinger said.
Besides Nvidia, ARM is also challenging Intel’s data-center dominance, with all major AI and hyperscale operators building chips on the architecture.
Google – which has relied heavily on x86 chips in the cloud — recently announced its homegrown Axion CPU, which it will offer customers in the cloud. Amazon already has its ARM-based Graviton CPU, and Microsoft announced last year that its Cobalt CPU is based on ARM.