After abandoning the Rialto Bridge, Intel changed its plans for Falcon Shores. Instead of the originally announced XPU (a product that combines CPU and GPU cores), Falcon Shores will be released as a GPU. This is a big change from what Intel prepared in response to the AMD MI300 accelerators, which are a mix of CDNA3 and Zen4 architectures, also known as “Exascale APU”, and a similar solution. Grace Superchip by NVIDIA.
Intel has clarified that the Falcon Shores GPU will be part of the Data Center GPU Max series and will feature HBM3 memory. It will be a multi-chip modular solution (tiles in Intel terminology) like Ponte Vecchio that will have a single GPU programming interface. Today’s announcement also means that the Ponte Vecchio will remain Intel’s top HPC accelerator through 2025. The company didn’t announce a successor to its Gaudi AI chip, instead proposing Falcon Shores as the successor to Gaudi3 and Ponte Vecchio. It is assumed that the Gaudi architecture will be partially integrated into the Falcon Shores chip.
As for the reasons behind Falcon Shores’ move, Intel explains that the company’s strategy for HPC products is to initially offer two separate product lines — GPU and CPU — which should allow for more flexibility. Intel isn’t abandoning its plans for the XPU, but it won’t be part of Falcon Shores’ initial launch.