AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)

Quake 2 RTX tests, the RTX 3080 is 190% to 200% faster than 6800XT:

'Must be drivers' - 'NV paid the devs' - 'RT isnt worth it'. In all seriousness, RTX gpus have the large RT advantage, AMD can make strides even in RT, but their behind. Its the same on the reconstruction front. They compete in rasterization though, atleast if BW isnt a problem.
The push for rasterization is there (on both fronts), but i doubt its the only way forward.
 
'Must be drivers' - 'NV paid the devs' - 'RT isnt worth it'. In all seriousness, RTX gpus have the large RT advantage, AMD can make strides even in RT, but their behind. Its the same on the reconstruction front. They compete in rasterization though, atleast if BW isnt a problem.
The push for rasterization is there (on both fronts), but i doubt its the only way forward.

NVIDIA did not even go "all in"...as far as I know Quake RTX does not support DLSS...the could have made the "numbers" seem even better when compared to AMD on that account.
 
Quake 2 RTX tests, the RTX 3080 is 190% to 200% faster than 6800XT:

CkE8Kw3pzex6T3ZyDTCyBU-970-80.png.webp


https://www.pcgamer.com/amd-rx-6800-xt-vulkan-ray-tracing/

Seems the best option we have to make some assumptions on RT perf NV vs. AMD...

Adding this for Turing:
upload_2020-12-16_12-16-27.png

Unfortunately i don't know the cost of denoising, but asking some people back then they guessed 5-10ms. Let's take 7. How much other non RT work is going on? Maybe with postprocessing we get 10. Rasterization is not used.
Subtracting this gives:
3080: 6ms for RT (took the 62fps result from DF which is much worse)
2080: 22ms
6800XT: 15ms

That's very vague, but i think it's better than looking at AAA games which do too much other stuff. I guess Quake BVH depth ends up much smaller than with current games details, but still.

My interest is again asking 'Is all BVH and traversal done with compute on AMD, so very flexible in theory?'.
I think 20TF 6800XT GPU beating a 10TF 2080 does not make that unlikely.
 
I remember first Quake RTX had some 'all surfaces become mirrors' mode, which turns off denoising and they used it as RT benchmark internally. Guess that's still in with some console command.
 
There's a mode in the menu that removes textures as a basis for the denoise, but I didn't see a fully mirrored world.
 
Seems the best option we have to make some assumptions on RT perf NV vs. AMD...

Adding this for Turing:
View attachment 5105

Unfortunately i don't know the cost of denoising, but asking some people back then they guessed 5-10ms. Let's take 7. How much other non RT work is going on? Maybe with postprocessing we get 10. Rasterization is not used.
Subtracting this gives:
3080: 6ms for RT (took the 62fps result from DF which is much worse)
2080: 22ms
6800XT: 15ms

That's very vague, but i think it's better than looking at AAA games which do too much other stuff. I guess Quake BVH depth ends up much smaller than with current games details, but still.

My interest is again asking 'Is all BVH and traversal done with compute on AMD, so very flexible in theory?'.
I think 20TF 6800XT GPU beating a 10TF 2080 does not make that unlikely.

Ryzen didn’t completely beat Intel in the first round. They focused on core count, not IPC.
For 6xxx AMD focused on rasterization, next they will focus on RT performance
 
Intel 14nm was supposedly superior to the 10nm AMD (TSMC?) process with which they already beat intel but anyway.
AFAIK, AMD never used TSMC 10nm in any shipping/shipped consumer products. They went straight from GloFo 14/12nm to TSMC 7nm.

Ryzen has beaten Intel because they are stuck on 14nm. Nobody is stuck anywhere in GPUs.
RDNA2 is also AMD's 8th(?) product family (more if we count actual unique dies) on TSMC 7nm! Plenty of experience there to go around.
 
Q2 RTX, Denoising takes 4.23ms on my RTX 2060, 19.8% of a frame time.

q2rtx_denoisek0k4j.png


Also ALU:FMA ratio is almost 1. It should be very friendly to Turing compared to Ampere, but bottlenecked by half speed RT units.
sdfsdf3rkf5.png

Are those numbers averages for the entire frame? I don’t think using those numbers will give an accurate ALU:FP ratio. There are tons of factors that affect instruction throughput during the frame. It’s better to look at instantaneous throughputs at specific points.
 
...am I missing something here?

In order to break 60fps with an RTX 3080 and full path-traced RT, we get something that looks like Quake II?

zzz
Yeah, this is DX10 and G80 vs R600 again. The hardware, even Ampere, just isn’t ready to say goodbye to rasterisation.
 
Back
Top