Intel has recently shifted its focus from Xe-HPC GPU accelerators to the Gaudi platform, although it’s struggling to meet sales targets with its latest Gaudi 3 AI accelerator. Now, the company is setting its sights on the upcoming Falcon Shores GPU platform. Interim co-CEO Michelle Johnston Holthaus explained at the Barclays 22nd Annual Global Technology Conference that Falcon Shores, while an initial step, won’t be groundbreaking right away and will necessitate rapid evolution to become a viable product.
“Transitioning from Gaudi to our inaugural Falcon Shores GPU involves a steep learning curve,” Holthaus remarked. “Is it going to be perfect right off the bat? No, but it’s a crucial initial phase. We’ll learn a great deal about the platform’s operation, the software integration, and how the market reacts, all of which will enable us to iterate quickly and effectively.”
Intel conceded that its Gaudi 3 model would not meet its 2024 sales projections due to software issues. Further details emerged at the Barclays Conference, revealing that deploying the Gaudi platform, especially in large clusters for training purposes, poses significant challenges. Consequently, its application has been predominantly confined to inference tasks at the edge.
“Gaudi isn’t suited for widespread deployment; it’s not a GPU that can easily be integrated into systems worldwide,” Holthaus stated. “Its deployment ranges from major hyperscalers to smaller entities working at the edge.”
However, there are positives. The Gaudi platform has provided Intel valuable insights into both platform and software development. These learnings are expected to inform the development of future AI technologies, though it’s uncertain how these insights will translate to entirely new platforms like Falcon Shores.
Falcon Shores is rumored to feature a multi-chiplet design combining elements of Xe-HPC (or similar GPU technology) and x86 processors, unified by HBM memory. This integration of x86 CPUs and Xe-HPC GPUs within a single module using a unified memory architecture is anticipated to deliver a more than fivefold improvement in compute density, memory capacity, bandwidth, and efficiency per watt over platforms available in early 2022, according to a statement by Intel in 2022.
With advances in architecture and processing technology, Falcon Shores is expected to significantly outperform Intel’s 2022 offerings, which include the Xeon Scalable ‘Ice Lake’ processors and the first-generation Gaudi accelerator.
Ultimately, Falcon Shores will serve as a learning platform for Intel and its Independent Software Vendor (ISV) partners. Despite the limited success of Intel’s Data Center GPU Max ‘Ponte Vecchio’ in the AI sector, ISVs will need to become familiar with the Xe-HPC (or Xe-AI?) architecture utilized in Falcon Shores. As developing AI software is a lengthy process, Intel’s subsequent platform, dubbed Jaguar Shores, is likely to be the first with potential for widespread adoption.
