Intel’s Arc GPUs Will Include In-Driver Overclocking, up to 4,096 ALUs

This site may earn affiliate commissions from the links on this page. Terms of use.

There are a few bits of Xe data left over from Intel’s Tech Day last week that we didn’t hit the first time, so we’re circling back to include them. When Intel launches Xe, it’ll be the first time in two decades that we’ve had more than two companies competing for the GPU market. The last third-party manufacturer in the PC space was PowerVR, and while the Kyro and Kyro II enjoyed a moment in the sun, the Kyro III was ultimately canceled before coming to market. After 2002, the GPU market was a two-way party.

In a Medium post last week, Intel VP Roger Chandler shared some additional details on Intel’s Xe microarchitecture. Before we dive in, a bit of clarity: Xe HPG is the name of Intel’s GPU microarchitecture. Alchemist is a codename for this family of products. Arc is the brand name Intel has chosen for its GPUs.

I’m not sure this exactly lines up with how Nvidia handles branding, but there are overlaps. Nvidia’s latest GPU microarchitecture is codenamed Ampere. If there’s a specific codename for the Ampere family I’m not sure what it is, but externally Nvidia refers to these cards as the RTX 3000 family. Intel refers to Alchemist as a code name, however, not a brand. Ampere appears equivalent to Xe HPG and Arc is equivalent to GeForce, but it’s not clear if we’ll see the Intel Arc Alchemist + model number, or just Arc + model number.

Screenshot from Intel’s XeSS demo.

Roger’s post covers a few additional details. Alchemist GPUs can be stacked in configurations of up to 32 Xe cores per SoC. Each Xe core contains 16 256-bit vector engines and 16 1,024-bit matrix engines. Each vector engine contains 8 ALUs, which means there are 128 ALUs per Xe Core. A 32-wide Alchemist GPU would offer 4,096 GPU cores. Because Nvidia, AMD, and now Intel design GPU cores differently, GPUs cannot be compared strictly on the basis of core count. Core counts can help estimate performance within the same GPU family, but they don’t help when comparing across companies.

Intel’s DLSS competitor, XeSS, uses Xe Matrix eXtensions to render at maximum quality on Xe HPG hardware. Xe LP consumers, however, are not left out in the cold. If you have an Intel GPU with Tiger Lake graphics or an Intel Rocket Lake, you’ll have access to XeSS on the integrated GPU as well. We don’t know how good the image quality of the DP4a version will be compared with the XMX version, but integrated and ultra-low-end graphics solutions can use all the help they can get, and running these devices in 720p is much better for performance than tackling 1080p. With any luck, we’ll see integrated and low-end graphics take another leap upwards thanks to these technologies. The upcoming adoption of DDR5 won’t hurt anything, either.

As for overclocking support, Chandler writes: “We’re even integrating overclocking controls into the driver UI, to give enthusiasts the tools they need to push the hardware to the limit.”

Hopefully, Xe HPG and future Xe LP GPUs will support underclocking and undervolting as well as overclocking. If you find yourself trying to save battery power on a system with an iGPU, underclocking/undervolting the GPU and locking the frame rate to 30fps are probably the two best ways to save power, apart from generic tips like lowering the screen brightness. We would not count on the GPUs offering much headroom — that’s not the kind of thing we can judge without hands-on time — but whatever headroom there is, enthusiasts should be able to take advantage of it. Hat-tip to THG for noticing the blog post.

Now Read:

Comments are closed.