Intel's Flagship Arc Alchemist GPU Battles A GeForce RTX 3070 Ti In Leaked Benchmark
by
Zak Killian
—
Friday, January 21, 2022, 03:20 PM EDT
A lot of people were disappointed when CES 2022 came and went with nary a sign of discrete graphics adapters bearing Intel GPUs. Intel broadly promised Arc GPUs in Q1 of this year, but the thrust of the presentation was entirely focused on laptops, which are unlikely to include the hefty DG2-512 "Xe-HPG" processor meant to compete with AMD and NVIDIA's powerful desktop graphics cards.
We're excited to see the fruits of Intel's efforts, too, not just because we want more competition in the GPU space, but simply because having more GPUs in the market will help ease the cracking dam of pent-up demand. Arc Alchemist may have been delayed into the second quarter of this year—Intel removed all "Q1" messaging from its site recently—but we have a new data point that suggests that the GPUs may indeed be well worth waiting for.
Frequent leakster APISAK (known as @TUM_APISAK on Twitter) found a test result in the SiSoft SANDRA benchmark database that appears to clearly be an Arc Alchemist high-end desktop GPU. SANDRA lists the graphics card as having 4096 shaders in 512 cores, which matches earlier leaks about DG2-512.
If the SANDRA results database is to be believed, the card showed up with 12.8 GB of video RAM. That's an odd total, and it may be the case that SANDRA read the value incorrectly. It's also possible that this particular card has an unusual memory allotment, or that the Intel OpenCL driver is making only a part of the card's memory available.
Whatever the case, the really interesting part of the leak is the performance. Whoever was testing ran an OpenCL general-purpose GPU compute benchmark, and the results are pretty surprising. The presumed Arc Alchemist card scores 35,093 Mpix/s in half-precision compute, which puts it less than 1500 Mpix/s behind a GeForce RTX 3070 Ti.
If you move to single-precision compute, the performance drops off by over 32%, while the GeForce part only loses some 27%. That could reflect any number of things, from differences in the architecture, a weak point in the Arc driver, or some other unaccounted-for factor. It's a little concerning, only because single-precision compute is the most important type for gaming performance, though it's still a relatively small factor in that metric overall.
Perhaps most interesting to folks eager to hear about Intel's Xe-HPC datacenter compute accelerators is that the double-precision performance of Xe DG2-512 is nearly double that of the NVIDIA part. NVIDIA has been crippling FP64 float performance on its consumer GeForce GPUs nearly as long as they have been capable of GPGPU, but it looks like Intel has been more generous in that regard. The Arc part is even further ahead for 128-bit precision, although it's probably worth noting that few (if any) workloads perform that kind of math on GPUs.
Arc Alchemist GPUs will be showing up in laptops around the beginning of next month, but the future of the desktop parts is a little less clear. The last official messaging from Intel still indicates that we're looking at a Q1 2022 release date for Arc Alchemist desktop GPUs. Naturally, as soon as we know ourselves we'll be posting it here, so stay tuned.