Intel confirms tweaks to Falcon Shores design

More than two months ago, Jeff McVeigh, the Vice President and General Manager of Intel’s Supercomputing Group, announced the cancellation of the subsequent improvement product, Rialto Bridge, for Ponte Vecchio. Additionally, Intel has decided to extend the release cycle of their data center GPU to two years.

The next data center GPU from Intel, Falcon Shores, originally planned as a hybrid architecture XPU integrating CPU and GPU, akin to AMD’s Instinct MI300 APU (Zen 4+CDNA 3) or Nvidia’s Grace Hopper, has undergone a change in strategy. Falcon Shores will now be developed as a pure GPU solution.

According to a report by TomsHardware, Intel has recently clarified certain aspects of Falcon Shores. It is part of the Max series product line of data center GPUs and will feature a multi-chip module design similar to Ponte Vecchio. It will also incorporate a “scalable I/O design” and offer various options for HBM3 with capacities of up to 288GB and a throughput of 9.8 TB/s.

Intel will introduce a universal GPU programming interface called oneAPI, emphasizing support for CXL as a key differentiating factor to ensure compatibility across architectures and platforms. Falcon Shores is scheduled for release in 2025, while Ponte Vecchio remains Intel’s leading HPC GPU solution until then.

Last year, Intel unveiled the second-generation deep learning chip, Habana Gaudi 2, designed for AI training. In the updated HPC GPU roadmap, Intel has integrated the Gaudi family into the Falcon Shores product line instead of providing a successor to Habana Gaudi 3. Due to the distinct purpose, the computational architecture of the Gaudi family differs significantly from standard GPUs. Further details on its operation will be disclosed by Intel in due course.