Yeah, but RT is a very bad compute benchmark. VRAM access in inner loops, and even random access. Though, IC should help this a lot, and still it's slower.
Maybe NV has compute lead now even for me, but i need to see that first to believe it. If i get opportunity to compare, i will share results...
My personal argument here was: We would approach RT also without HW and its restrictions, and Crytek has proofed that. I did not expect similar results ofc., in terms of reflections, shadows, and GI all at once. But even this has been shown to be possible and practical from Epic, to quite some degree.
My work on compute GI uses raytracing, but algorithm can't be compared to the 'classical' approach of Crytek or DXR (also not with SDF or voxel tracing). In my case AMD indeed was much faster with all previous generations, which turned me into that 'AMD fanboy' i appear to be. Before my compute work, i preferred NV GPUs like most people.
In early discussion after DXR launch i was too optimistic towards classical RT using compute. I assumed i could find similar efficient solutions for high frequency details than those i've had found for low frequency GI. But i was wrong about that.
That said just to explain my stance, which is complicated and not the general perspective.
Thing is: My GI can solve the complete lighting model faster than any other solution i've seen so far. But i also want HF detail, thus i want HW RT.
Because the improvement from 'adding just HF detail' is not that big, i remain more critical about high costs of HW RT than others, simply because that cost / benefit ratio is smaller in my case.
On top of this, the limitation of continuous LOD not being possible, really adds a lot to those doubts, up to the point i see DXR as a 'failure / barely useful', etc.
But that's not because i'm a stubborn 'Glorious Compute Warrior'
I just need those API improvements to make HW RT 'worth it' for my case.
