In a new video update, Intel has asserted that it’s Xe-HPG (High-Performance Gaming) graphics cards have been successfully “powered on”, while the Xe-HP GPUs (for Data Centers and servers) are presently being sampled. This would fall in line with a recent leak where the Xe-HP GPU was described. It will supposedly feature HBM2 memory and be fabbed on TSMC’s 7nm class process node? Thirdly, Intel has also released the first revision of its chiplet-based HPC part (Ponte Vecchio) to manufacturing, and is waiting on their arrival to power it on too.
The Xe-HPG graphics cards are said to feature up to four GPU dies on a single substrate for 512-1024 EUs. That’s as many as 8K shaders, paired with a 256-bit bus and 8-16GB of GDDR6 memory. If Intel actually prices its new GPUs well and gets its drivers in order, it might just be able to properly challenge rivals NVIDIA and AMD.
The its dual-tile variant of its Xe-HP GPU is already being offered to select partners to run instances on Intel’s service, compile, and run Xe-HP compatible software, basically to get an idea of the performance and code adaptability for their workloads.
As for Ponte Vecchio, we’re looking at a chiplet based design with multiple GPUs connected together across a mesh interconnect via large chunks of (Rambo) cache and HBM 2 memory and running in tandem with Sapphire Rapids CPUs. These parts will be based on Intel 10nm SuperFin and SuperFin enhanced nodes.