Microbenchmarking Intel’s Arc A770 | Hacker News Reader

Microbenchmarking Intel’s Arc A770

chipsandcheese.com

97 pointspantalaimon15 comments

I'm kind of disappointed all the tests were without ReBAR. I'm curious as to what kind of difference that makes.

You are probably aware, but ReBAR has a huge effect on these card, so much that it is a requirement to get stable and acceptable FPS in games. I would assume that means the effect would be huge also in those smaller benchmarks.

edit: But was it really off in general, or only for the PCI-e 3.0 part of the article?

ReBAR only matters for transfers across the PCIe link, so for most of the tests in the article it's completely irrelevant. And even for the PCIe transfer tests, ReBAR should only affect the portions of the tests that are accessing more GPU memory than the default BAR size used for 32-bit compatibility.

Considering how much the Arc series loves to share system memory, I'm not so sure.

Intel's having to find and fix every place their drivers assumed that the CPU's RAM and the GPU's RAM were the same pool of memory. As a result, it's completely unsurprising that they're struggling with performance in real applications and games, but it doesn't provide a reason to expect microbenchmarks such as these to run afoul of an irrelevant problem.

It would make zero difference, except, maybe, for the PCIe bandwidth test. Because all other tests were stressed interfaces that reside within the GPU silicon itself or between the GPU and the DRAM.

Even the numbers of PCIe bandwidth test might not change much, even it's trying to test medium size memory to memory block transfers.

Apologies for the confusion, the tests were run with ReBAR. I've updated the article to reflect that

Shouldn't affect the conclusion for anything besides the PCIe copy to/from GPU tests.

These GPUs practically require ReBAR, performance without it drops significantly.

I would love to get my hands on one. They seem to be sold out everywhere.

If on Windows, you might want to check out the Gamers Nexus reviews on YouTube: the Intel drivers are a dumpster fire and build quality is... not great.

https://www.youtube.com/c/GamersNexus/videos

Is the benchmark suite available somewhere?

(Author here) See https://github.com/clamchowder/Microbenchmarks/tree/master/G...

It's very much a work in progress, as noted in the article. And some of the stuff that worked reasonably well on my cards, like the instruction rate test when trying to measure throughput across the entire card, went down the drain when run on Arc.

Have you tried reducing the register count in your FP32 FMA test by increasing the iteration count and reducing the number of values computed per loop?

Instead of computing 8 independent values, compute one with 8x more iterations:

    for (int i = 0; i < count * 8; i++) {
        v0 += acc * v0; 
    }

That plus inlining the iteration count so the compiler can unroll the loop might help get closer to SOL.

The problem is loop overhead matters on AMD, because AMD's compiler doesn't unroll the loop. Nvidia's does, so it doesn't matter for them.

unroll with #pragma unroll?

Reply on news.ycombinator.com