As per a report from the Chinese Bilibili forums, Intel is presently testing engineering samples of its upcoming Xe-HPC (High-Performance Compute) GPUs. These parts are reportedly based on TSMC’s 7nm node and will feature HBM 2 memory:
This falls in line with Intel’s Architectural Day disclosure, according to which the compute tiles and the I/O die will be based on an external process. Seeing how miserably Samsung’s foundries have been performing, it’d be safe to say that Intel doesn’t really have any option but to go with TSMC’s 7nm node which is roughly on par with Intel’s 10nm process.
The use of HBM 2 memory also makes sense as you’re looking at a Data Center product aimed at accelerating neural networks and AI-based workloads which tend to be bandwidth-intensive. Furthermore, Intel has already confirmed the use of HBM 2 memory in its Ponte Vecchio HPC lineup, so this isn’t really a surprise.
Considering that we’re looking at a 7nm/10nm part, Intel should have at least some Xe-HPC products ready for launch alongside the Xe-HPG parts in 2021 if expects them to remain competitive.
