In letter : Intel’s first wave of Arc graphics cards are not yet here, but the company has put forward the Xe-HPG kernel and support for Xe Super Sampling (XeSS). The first benchmarks have started to appear online, and if they are true, they suggest that Intel still has a lot of work to do if it wants to have “Alchemist” ready for a 2022 release.
Intel is currently working on its high-performance Arc GPUs for desktops and mobile form factors, but the first products of this effort will not arrive until the first quarter of 2022. Chipzilla has given hints on its plans with Arc and how it intends to challenge both Nvidia and AMD in the discrete GPU market, but other than a roadmap and a few nifty software tools, we’ve had little to do to get a feel for the way these newer GPUs might work.
A number of leaks have suggested that Intel’s upcoming “Alchemist” graphics solutions will be available in multiple SKUs, ranging from 128 threads and 4 GB of VRAM on a constrained 64-bit bus and up to a top model. range with 512 execution units with 16 GB of VRAM on a 256-bit bus, which would be roughly between an Nvidia RTX 3070 and an RTX 3080 in terms of performance.
Thanks to a few first markers identified by @Tum_Apisak, we can get a very rough idea of the performance of the mobile variant of the high-end Alchemist GPU. This is Geekbench 5 testing done with an Intel Tiger Lake processor, and we’re probably looking at a first engineering sample, but the picture isn’t pretty for Alchemist.
It looks like the engineering sample got an OpenCL score of 34,816 points when clocked at 1800MHz, which would be a rather disappointing result if taken at face value. For reference, this is roughly the same level as Nvidia’s latest generation GeForce GTX 1650 Max-Q, which is actually capable of scoring over 36,000 points on the same test.
Granted, this is just a singular test that measures the computational performance of the Alchemist GPU, which is again an early sample clocked at a rather low 1800 MHz frequency. Intel suggested during its Architecture Day presentation that we can expect the Xe-HPG graphics engine to have 1.5 times the performance per watt and the frequency boost of the Xe-LP found in the Iris card. DG1 released earlier this year.
This corresponds to a theoretical boost clock speed above 2 GHz, which is a further indication that we are looking at very early silicon here, and in a mobile form factor no less, which means it could be limited in power. In June, someone tested the Iris DG1’s playing performance and found it to be surprisingly decent in a number of popular games, despite a barrage of early benchmarks suggesting otherwise.
If anything, the same can happen with Alchemist, and it goes without saying that Intel only needs to offer entry-level and mid-range GPUs under current market conditions, which should persist for at least a year. If it can manufacture them continuously using its own chip factories, many players and cryptocurrency miners will buy them in the blink of an eye.