NVIDIA self-developed CPU debuts, Intel Ponte Vecchio GPU is coming

  At the 34th HotChips exhibition, NVIDIA announced the architectural details of its self-developed Grace server-level CPU, and topped the SPECrate2017_int_base test list with 740 points. Not to be outdone, Intel also introduced the potential of its own Sapphire Rapids HBM server CPU + Ponte Vecchio (2-Stack) GPU platform, especially the strength of the latter.

  Hong Jiang, Intel's chief GPU computing architect, pointed out in his speech that the Ponte Vecchio has three configurations: from a single OAM, to an x4 subsystem with Xe Links, the Ponte Vecchio GPU can not only run alone, but also be deployed on a Sapphire Rapids dual-socket server on the platform.

GPU parameters: Ponte Vecchio's flagship data center GPU provides 128 Xe cores, 128 Ray Chasing (RT) units, HBM2e memory, and 8 Xe-HPC GPUs connected together. In terms of performance indicators, the 2-StackPonte Vecchio GPU configuration (such as the configuration on a single OAM) can provide up to 52 TFLOPs of FP64 / FP32 computing power. Other aspects: Ponte Vecchio is manufactured using five different processes such as Intel7 and TSMC N7/N5. The number of transistors exceeds 100 billion, and it integrates up to 63 Tile unit modules (47 functionalities), with a total packaging area of 4844 square millimeters. Power consumption up to 600W.

For the performance comparison with NVIDIA, Intel's report card is: Sapphire Rapids HBM processor + Ponte Vecchio (2-Stack) GPU platform can achieve 2.5 times than the performance of NVIDIA A100.