Intel will not only step into the GPU market with Alchemist gaming graphics cards but also launch their Xe-HPC Ponte Vecchio GPU for servers and high-end workloads.
The company has powered on the Ponte Vecchio A0 silicon, and this is a huge announcement as it is a big bet Intel has laid, having made the first high-performance server GPU. On Intel Architecture Day 2021, Raja Koduri was there to unveil these amazing products, and Intel seems impressive at its first attempt.
Ponte Vecchio has achieved a whopping 45 TFLOPs of single-precision compute performance in its current A0 silicon version. Indeed, this is a new way of GPU as it is the first Xe-HPC-based processor to come with a multi-tile design, including Compute, Rambo, HBM, and EMIB tiles, a total of 47 tiles with 100 billion transistors.
As we know, a multi-tile design involves connecting the parts of the GPU tiles together and ensure they work correctly. The team at Folsom has measured a whopping 5 TBps of memory fabric bandwidth and 2 TBps of connectivity bandwidth. Intel has even compared Ponte Vecchio to NVIDIA’s; you see a huge jump in performance which Intel is claiming.
The main Xe-HPC Xe-Core, which is the building block of the GPU, will feature 8 Vector Engines and 8 Matrix Engines. Compared to the gaming-oriented Xe-HPG, Ponte Vecchio will have fewer Engines, but they operate at wider buses (512-bit and 4096-bit, respectively), while for HPG, those are 256-bit and 1024-bit, thus ensuring high bandwidth.
The Xe-HPC Slice will be combining 16 Xe-Cores, and interestingly, Ponte Vecchio will also be equipped with Ray Tracing Units, similar to the Xe-HPG, with each Xe-Core tied to a single RT unit. At the same time, Intel has demonstrated the use of these RT cores for Ray Traversal, Triangle Intersection, Bounding Box Intersection.
Ponte Vecchio will be made available in both 1 and 2-stack configurations, so specs go up to 8 cores, 128 Xe-Cores, and 128 Ray Tracing Units. While the 2-stack configuration will have 8 memory controllers for HBM2e, and Xe Link will be used to couple Ponte Vecchio’s in multiple subsystems.
Intel has even confirmed that the Aurora supercomputer will be utilizing the upcoming Ponte Vecchio GPUs. Intel also showed off a single Aurora Blade that will constitute one computational unit of the supercomputing cluster. The high-performance Ponte Vecchio GPU will be making its way at the beginning of next year.