Synthetic benchmarks get a nasty rap. Sure, they are not at all times consultant of actual software efficiency — that a lot is inarguable. They’re nonetheless extraordinarily helpful, although. Not solely can they assist builders pinpoint efficiency issues for particular troubleshooting, however additionally they can assist customers affirm that their laptop is working accurately.

Yep, as you have in all probability guessed, we’re a bundle of artificial Intel Arc GPU benchmarks. Twitter consumer and GPU developer 0x22h has gotten his arms on what appears to be an Arc A380 desktop graphics card, presumably the exact same Gunnir card we wrote about earlier than, and he is run a number of benchmarks on it to look at the weird conduct.
A easy microbenchmark to check uncooked fill price.

Most of the benchmarks he ran are exams he appears to have devised himself, consisting of easy take a look at apps rigorously concentrating on one metric or one other. He comes out with uncooked fillrate of 62 Gigapixels/sec, which is round what we’d anticipate for this half. Another benchmark appears to point that the Alchemist components are utilizing a tile-based structure, just like PowerVR and up to date NVIDIA components.

More regarding are the benchmarks that show considerably worse efficiency than anticipated. Multi-sampling efficiency on the Arc A380 is “very poor”, with 2x MSAA dropping efficiency by practically 40% in one other take a look at. Meanwhile, OpenCL efficiency in AIDA64 is arising with a results of simply 2.5 TFLOPs, or somewhat greater than half of what we’d anticipate from this half.
AIDA64 consequence from David Huang (@hjc4869 on Twitter).

Well, that final consequence could also be all the way down to a bug in AIDA64. Eagle-eyed Locuza, well-known for his annotated die shots, identified that AIDA is barely counting 512 “cores” for the A380, half of its precise quantity. It’s seemingly that an update is required for the long-running system data and benchmark software to permit it to totally understand efficiency on Arc Alchemist.

The most brow-furrowing outcomes gathered by 0x22h are the cardboard’s efficiency in 3DMark. Unfortunately, 0x22h did not share any outcomes from the usual Fire Strike or Time Spy benchmarks, however did run the “Feature Tests” included with the app. The outcomes are blended.
Both the Sampler Feedback and Mesh Shaders exams accomplished accurately, and whereas the speedup is decrease than what we’re used to on GeForce and Radeon GPUs, we additionally should do not forget that that is an especially small GPU. The proven fact that the exams each accomplished accurately and supplied important efficiency positive aspects show that Intel has been focusing its improvement efforts on DirectX 12, and that these efforts have borne fruit.
3DMark API Overhead take a look at on Arc A380.
The sorts of outcomes we normally anticipate (GeForce RTX 3070 Ti).

The API overhead take a look at’s outcomes are extra problematic. The DirectX 12 quantity got here out completely high-quality; 11.5 million draw calls is an efficient consequence given the system in query. The DirectX 11 and Vulkan exams should not so nice, notably the DirectX 11 multi-threaded take a look at, which did not even break one million draw calls per second. This is probably all the way down to insufficiently-mature drivers, as DirectX 11 may be very reliant on the driving force for stable efficiency.

It’s vital to do not forget that these artificial benchmarks might in the end have little bearing on game efficiency. It’s additionally debatable that DirectX 11 efficiency is much less vital now than it as soon as was, however for our half, let’s simply say there is a motive AMD remains to be working to enhance its personal DirectX 11 and OpenGL drivers. Hopefully Intel can get its GPU software program all fastened up earlier than later.