Intel’s Ponte Vecchio GPU Tips Up as an Add-in Card
Intel hasn't had much to say about its upcoming Xe graphics architecture lately, but we've seen quiet signs that the company is still working on its GPUs. Part numbers continue to tip up in import registries referring to various Ponte Vecchio parts.
— 188号 (@momomo_us) April 23, 2020
Ponte Veccio is a datacenter and HPC gpu intended for big-iron deployments, and it's supposed to be a major debut for the advanced interconnect Intel has been talking about since 2018. The upshot of the company's statements is this: While Intel hit roadblocks on CPUs, it didn't stop working on interconnect technology. EMIB and Foveros are both used in Ponte Vecchio to provide multi-die communication and to allow for optimized signal routing. A fully-enabled Ponte Vecchio GPU combines eight 7nm chiplets. Intel plans to pair them with HBM memory and then stack six of the GPUs into a system with two Sapphire Rapids CPUs.
This is the basic structure of the Aurora supercomputer currently under construction. It's also most of the “structure,” if you will, of the OneAPI intended to tie all of this computational capability together. Intel has created a software framework intended to support development on everything from IoT devices to HPC computing. This implementation of Ponte Vecchio is almost certainly a developer product, but it's interesting to see the hardware shipping on 10nm add-in cards rather than integrated into boards. These are likely 10nm implementations of the architecture, given that Intel's 7nm isn't expected to be commercially available until the end of 2021 and it's early for even risk production.
All of this work is being done to support Intel's push into the GPU market. That's supposed to begin this year with the release of a 10nm datacenter-focused part and continue into 2021 with other chips focusing heavily on compute and the low-power market. We've heard rumors of GPUs that draw up to 500W and are intended to compete with the top-end of what Nvidia and AMD can bring to bear.
It's going to be interesting to see how the HPC space evolves in the next few years. AMD hasn't been much competition for Nvidia in GPU HPC to-date, though that could change with further improvements to ROCm and the next generation of CDNA GPUs. Multiple companies are bringing specialized accelerators to market, but Intel is the only firm that covers the entire gamut of solutions, from embedded / mobile computer vision applications to CPUs, GPUs (both consumer and datacenter-focused), and HPC big iron.
The market coverage is factual. The quality of the hardware and drivers is unknown. All of this adds up to a decidedly interesting product launch. We should see the first 10nm Intel GPUs in consumer hardware this year if the company keeps to its previous roadmaps.