Intel Arc Alchemist A370M GPU Benchmark Leak Strikes At The Heart Of NVIDIA's Mobile Line-Up
It's critically important when talking about processor performance, whether CPU or GPU, that we consider many types of applications. A chip that performs well in one application or workload may not perform very well in another. This can be down to hardware design, driver optimizations, or even application optimizations, and it's quite difficult to predict, which is why when we do benchmarks, we try to hit a wide array of tests.
The venerable SiSoft SANDRA includes numerous benchmark tests, but the one we're looking at today is the GPGPU benchmark. This test uses OpenCL to run a crunchy compute workload on whatever devices you select in your system. It supports any CPU or GPU that can run OpenCL, which is basically everything on the market.
The specific result of interest is this one, which comes to us by way of the sharp-eyed TUM_APISAK on Twitter. It's a benchmark result for an Intel Arc A370M GPU. This isn't the first time we've seen the A370M, but it is the first performance data of any kind that we've seen for this part. As APISAK notes, the result includes both the Arc A370M GPU as well as the system's integrated UHD Graphics 770, but he calls it a "misreport," and we don't think that's quite correct.
It's true that SANDRA says the test was run on a single device. Unfortunately, that's not informative because Sandra always says that in this test. This specific test is a general compute test using OpenCL, and when you run it, you can select which devices to use for the test. If you peer into the result, you'll note that the result lists two "threads." On this specific test, that actually means two OpenCL threads, or two devices. Don't believe us? Check out these other results on the SANDRA database that clearly include two devices and list "Devices/Threads: 1/2":
Because of that, we advise some caution when interpreting this data. Of the 160 EUs listed in SANDRA's report, 32 of them are from the UHD Graphics 770 part. That's 20% of the compute capacity of the combined parts, which means it's reasonable to expect the A370M on its own to score some 15-20% lower than the actual result here on OpenCL benchmarks like this one. This is corroborated by our own test of a UHD 770 in SANDRA, which scored 412 Mpix/sec, or about 17% of the combined score above.
It's also worth noting that despite this being called the "Arc A370M", it's actually being tested in a desktop machine. We know that beyond a shadow of a doubt because the UHD Graphics 770 part is not offered on any laptop chips. That probably doesn't affect its performance in this benchmark very much, but it's still odd.
As for its performance? If we subtract the UHD 770's contribution we end up with a score in the 1900 range, which is fantastic considering the bracket we expect this GPU to slot into. That puts it in roughly the same territory as a laptop GeForce GTX 1060 or GTX 1660. Because the UHD 770 is mixed into the results, we can't really rely on the clock rate or memory configuration information reported by SANDRA, but we expect this GPU to have 1024 shaders and either a 64-bit or 96-bit memory bus.
Of course, OpenCL performance doesn't actually tell us much about how the GPU will perform in typical client workloads. Indeed, this type of result, while informative regarding DG2's parallel compute performance, is almost completely unrelated to how Alchemist will perform in gaming, browsing, encoding, and other client workloads. It's a shame nobody seems to have run 3DMark on one of these things yet, but we'll let you know as soon as they do.