Intel provides more details on its Arc GPUs, which will be made by TSMC

Earlier this week, Intel announced Arc, the branding for the new gaming GPUs that will face off with Nvidia’s GeForce and AMD’s Radeon cards when they launch early in 2022. Today, Intel provided a few additional details on Arc, its underlying Xe-HPG architecture, its graphics drivers, and the “XeSS” upscaling technology that will work with both Intel’s GPUs as well as GPUs from Nvidia and AMD.

The HPG in Xe-HPG stands for “high-performance gaming,” and it builds on the Xe-LP (“low-power”) graphics tech included in Intel’s 11th-generation Tiger Lake laptop processors and the low-end DG1 dedicated GPU. At a high level, Xe-HPG will support all of the features in DirectX 12 Ultimate, just like Nvidia’s RTX 2000- and 3000-series GPUs and AMD’s RDNA2-based Radeon RX 6000-series cards. This includes, most notably, support for hardware-accelerated ray tracing, variable rate shading, and mesh shaders.

Since at least 2006, Intel has talked about its GPU hardware in terms of “execution units,” or EUs. The company is doing away with that terminology for Xe-HPG, replacing it instead with the concept of the “Xe-core.” Each Xe-core is composed of 16 vector engines and 16 matrix (or XMX) engines, as well as L1 cache and some other hardware. Four Xe-cores combine with ray-tracing units and other fixed-function hardware to form a “render slice,” which is the bare minimum any Xe-HPG GPU will need to function (along with L2 cache and a memory interface).

Alchemist-based chips can include “up to” eight of these render slices, which implies that we’ll see at least a couple of different Arc GPUs with different levels of computing power. But Intel hasn’t disclosed any specific hardware configurations, and it also hasn’t gone into any detail about clock speeds, the memory interface, or the amount or type of RAM that the first Arc GPUs will include. (Quick terminology reminder: Xe-HPG is the name of the GPU microarchitecture, Alchemist is the codename for the chips based on Xe-HPG, and Arc is the consumer branding you’ll see on store shelves.)

The XMX hardware in Xe-HPG GPUs will be key to Intel’s AI-accelerated upscaling technology, also called Xe Super Sampling or XeSS. Like Nvidia’s DLSS, the idea is to upscale lower-resolution images with as little quality loss as possible. Doing so will make it possible to play a game at 4K resolutions without requiring the graphical horsepower necessary to actually render a native 4K image.

Intel says that you’ll get the best image quality and performance out of XeSS if you’re taking advantage of its XMX hardware, but there’s a separate version of XeSS that uses DP4a instructions common in modern GPUs from Nvidia and AMD as well. Intel didn’t call out specific GPU models, but DP4a instructions are included as far back as the Pascal architecture in Nvidia’s still-popular GTX 1060 and 1050 series cards. Nvidia doesn’t actually support either of these models with DLSS. Integrated GPUs will be able to benefit from XeSS as well. The SDK for the XMX version of XeSS will be available this month, with the DP4a version following later this year.

When they’re released, Arc GPUs will use the exact same driver package as Intel’s integrated GPUs, so improvements made to benefit one architecture can potentially benefit all of them (Intel boasted that it had overhauled its graphics driver in the past year to improve performance in CPU-bound games running on Xe-LP GPUs by between 15 and 80 percent).

Finally, the elephant in the room: Intel isn’t making Alchemist chips in its own chip factories, at least not now. The company announced that it will use TSMC’s 6 nm process for Alchemist chips, meaning that Intel’s GPUs will need to fight for capacity with AMD, Apple, and everyone else manufacturing cutting-edge chips at TSMC.

“The reason [for using TSMC for Xe-HPG] is simple,” wrote Intel SVP Stuart Penn in an accompanying press release. “Just as our designers use the right architecture for the right workload, we also choose the node that best fits that architecture. At this point in time, these foundry nodes are the right choice for our discrete graphics products.”