10 Comments

Agree with Simon, very helpful analysis! I found your interpretation of the B580 limiting its efforts for Path Tracing in very challenging cases like Cyberpunk to make a lot of sense. Even a 4090 is taxed significantly by CP2077 with everything turned up to the highest settings, and you explained well how Path Tracing costs a lot of bandwidth and compute that is then missing elsewhere. Boils down to the graphics (drivers) team choosing wisely which battles to fight. 11 FPS just don't look good 😄.

Would be interesting to know how, for example, Nvidia's drivers deal with those choices across GPUs with very different capabilities. And, if someone from Nvidia reads this: Send Chester a 5090, so he can treat us to a great deep dive on Blackwell !

And, AMD, how about doing likewise with the 9070XT? RDNA4 sounds mighty good, maybe even better after a deep dive into it by Chester!

Expand full comment

I have a 9070 on-hand and I'm working through rdna 4 stuff. There's a lot to do, give me some time :P

Expand full comment

I would love to see an in-depth comparison between the recent architectures of the big players.

Expand full comment

👍🏻👍🏻👍🏻 Thanks, and looking forward to your article on RDNA4! I'm sure you're plenty busy right now 😄.

Expand full comment

Great article as always. Man, if only there were a big battlemage chip

Expand full comment

I also wonder what's happening with that! A B780 would be a welcome addition, especially if Intel decides to stay a bit ahead of the pack and provide 18 or 20 GB GDDR6 for a large Battlemage dGPU.

Expand full comment

>I also wonder what's happening with that!

I don't think it's too hard to see that there's enough happening at Intel anyway without having to spend big on another dGPU tape-out for a chip that, realistically, will have 0% market share anyway. ;)

Expand full comment

Great and helpful analysis, as always. Thank you!

(I might be wrong, but I think that any-hit shaders are used for shadows, not for transparency. With transparency, order matters)

Expand full comment

I think they're used for both, though in different ways. An any-hit shader gets invoked for any intersection with transparent geometry (https://learn.microsoft.com/en-us/windows/win32/direct3d12/any-hit-shader)

Expand full comment

>It reminds me of DirectX 9 GPUs striking a balance between vertex and pixel shader core counts.

Hopefully we might see yet another "shader unification" phase soon enough, then. :P

Expand full comment