Originally Posted by FMJarnis
I'm sure you will describe your credentials as a DX12 developer and present proof of your assertions. I'll be waiting. I'm also happy to forward them to our development team in case they open any new (legit) questions on how Time Spy is implemented.
Of course not since it doesn't matter. I don't plan a war on you and don't plan to bring this down to a personal level since I have nothing against 3DM and I like synthetic benchmarks; but your claim that you can do it without IHV specific paths is imho wrong. You can do it, but it's not doing it the right way. GCN is so different to Pascal or Maxwell that it's hurting the performance. Preemptive-friendly approach is fine, but parallel async execution is significantly better solution for a real game environment.
You can limit yourself not to hit a glass chin of any of the architectures, you could design your renderer the way that would be easy on what NV lacks, but that is not exactly fair to AMD.
I've read Futuremark's disclaimer: it does actually make sense. As a benchmark, from a bechmark PoV, you do actually need to keep persistence in test content and any kind of IHV specific paths can hurt that.
It probably does not maximize GCN abilities, but it should give both vendors enough of the new API usage that it as closely as possible resembles game cases.
Real games with IHV specific paths might be better at utilizing specific architectures, but the difference should not be as drastic as I presumed.Edited by toncij - 7/26/16 at 2:49pm