Intel Arc Alchemist

© GETTY IMAGES/KATSUMI MUROUCHI

Turning silicon into gold

Twenty-four years after exiting the dedicated graphics card market, Intel has returned. The landscape in those intervening years has changed so radically that what was once abandoned as an unnecessary side project now represents a major missing piece of Intel’s technology portfolio. Put simply, Intel needs a competitive GPU offering, and Arc Alchemist is where it starts.

This isn’t the first time Intel has tried to re-enter the GPU space. Larrabee was supposed to be a GPGPU (General Purpose computing on Graphics Processing Units), a chip that could do graphics but also other computational workloads, including scientific computing. Announced in 2008, a year later Intel said Larrabee would not be a consumer product but would instead go after the HPC (High-Performance Computing) market. Then Intel canceled Larrabee the following year, though some elements of the design lived on in the Xeon Phi offerings.

Now we’ve come full circle, as the third generation Xeon Phi, codenamed Knights Hill, was canceled in favor of a new high-performance architecture built from the ground up for the Aurora supercomputer. That’s the Xe-HPC architecture, and there’s also Xe-HPG (High-Performance Graphics) as a consumer variant that has now officially launched as Arc Alchemist. Let’s pull back the heat spreader and find out what makes Alchemist tick.

–JARRED WALTON

A GROUND-UP ARCHITECTURAL REDESIGN

Intel isn’t exactly new to the graphics arena—even in the past 24 years, it has released 12 generations (give or take, depending on how you want to count them) of GPU architectures. The old Intel i740 back in 1998? That was generation one—though Intel likes to claim that the ‘real’ generation one was in the first Core processors with integrated graphics in 2010. Either way, much has changed in the ensuing years. Where i740 had just a single pixel pipeline clocked at 220 MHz, the 12th generation Xe Graphics (used in Tiger Lake, Rocket Lake, Alder Lake, and now Raptor Lake) pack up to 768 universal shader pipelines and clock at up to 1450MHz.

Alchemist has some high-level similarities to the Gen 12 Graphics architecture, but it also changes a ton of stuff and delivers up to an order of magnitude increase in performance over the fastest Gen 12 solutions—and that’s just in normal graphics workloads. Alchemist also packs in a whole bunch of new features and hardware that aren’t available in any of the previous architectures.

Let’s put things in perspective, here. Tiger Lake has 96 Execution Units for its Xe-LP GPU, and the various other parts of the graphics engine— including display controllers, media encoding blocks, and others—consume around 40 percent of the entire die area. With a die size of just 146mm2, that means the Tiger Lake GPU only needs about 58mm 2.

In contrast, the smaller ACM-G11 Alchemist GPU measures 157mm2, nearly three times the size of the integrated GPU, while the larger ACM-G10 chip checks in at a hefty 406mm2.

Intel’s Arc Alchemist comes in two forms, the larger and significantly more powerful ACM-G10, and the smaller ACM-G11 that trims down core counts.

© INTEL, GETTY IMAGES/KATSUMI MUROUCHI

The performance delivered by the Alchemist GPUs is the best indicator of everything that’s changed, but let’s cover the specifics. The new architecture includes full DirectX 12 Ultimate support: ray tracing hardware, Variable Rate Shading (VRS), mesh shaders, and sampler feedback. And by all measures, Intel didn’t just tack on some hardware to check a box; the ray tracing units (RTUs) are capable and appear to be roughly on par with Nvidia’s Ampere RT cores.