In context: Japanese website ASCII.jp posted an interview with Intel today, covering various topics about its upcoming graphics hardware. This includes some new details about the company's Arc Alchemist graphics cards, like why it chose TSMC's foundry over its own. They also shed some new light on the compatibility range for Intel's new upscaling tech as well as Intel's relationship with GPU partners.
Correction (09/21/21): Due to an error in translation, some of the information in this article was not accurate at the time of publishing. An Intel representative reached out to TechSpot to inform us that the original source, ASCII, had misquoted Raja Koduri. The ASCII article said Koduri had indicated that Intel chose TSMC because of its manufacturing capacity.
According to the corrected ASCII article, Raja said, "The factors considered for whether it's internal or external manufacturing is based on three factors – performance, cost and capacity. He did not comment that Intel did not have sufficient capacity for advanced nodes."
According to Wccftech's translation, Intel said it chose TSMC's six-nanometer processor node for the Arc Alchemist GPUs over its own node—probably Intel 7, due to manufacturing capacity. Intel also took operating frequency and cost into consideration.
Earlier this summer, Intel revealed its roadmap for upcoming node processes as part of its intention to compete with TMSC and Samsung's nodes, with Intel 7 as its latest process. Using its own process for future GPUs isn't out of the question, as Intel told ASCII.jp it will reconsider the same factors as before when the time comes for the next generation.
Intel confirmed that its upscaling technology, XeSS, will work with DG1 Xe-LP GPUs as well as 11th generation Tiger Lake integrated graphics. XeSS is supposed to use AI to upscale images from lower resolutions to save on performance using Intel's XMX cores, similar to how Nvidia's DLSS uses the tensor cores on that company's RTX GPUs. Some of XeSS's functionality, however, will work on GPUs without XMX cores as long as they have DP4a instruction.
Intel also stressed to ASCII.jp its intention to release regular driver updates for the Arc GPUs, including releasing new drivers for every major PC game launch. This is likely why Intel recently hired four gaming industry veterans to work with developers.
When it comes to GPU manufacturing partners making custom versions, Intel seems to have already shipped them its reference design. It told ASCII.jp that it's still deciding whether to sell its own reference-only design like Nvidia's founders edition GPUs, though it may have already teased such a reference design to the public.
Previous reports indicated the Arc GPUs would come with up to 32 of Intel's Xe cores. When talking about the scalability of the architecture in this interview, though, Intel left the door open for more than 32 cores in future cards.
https://www.techspot.com/news/91331-intel-covers-arc-gpu-fab-choice-xess-backwards.html
