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

Is math throughput the bottleneck in DLSS cost?

 

I guess it will come down to how much compute do you need to render at something like 4k vs render at lower resolution + a super resolution solution.

 

I said it better than what they have before. Be patient, you can see the result and after you can say what you think about it.

 

Did you see the TAAU vs. DLSS 2.0 results from the DLSS 2.0 presentation? TAAU definitely has much lower quality from the same internal resolution as DLSS 2.0.

 

But AMD does not have it, this is bad, if Sony and MS have an ML based reconstruction it will be bad they are part of a conspiracy against AMD with Nvidia.

 

We should get one person with an RTX 3080 and one with an RX6800 to stand in the exact same spot in the game, with the same settings, and get the profiler shots from both cards. We can then normalize the numbers, and we can get an idea which card is better at which function. The shots are currently too different to make educated estimates.

 

If there is a save file for us to compare

 

Yes, we need to agree on resolution and settings, creating and sharing save file should be easy.

 

How's these settings or resolution?

Spoiler

 

Great, I like

Here is a link to my save in this spot:
https://we.tl/t-ji2oR2IqIG



Recording a video took about 0.5FPS to 1 FPS.
And my OC settings:

 

Nice, you have selected Temporal Upscaling as AA method though not Temporal AA. Time of day can change performance as well.

Running my 3080 at undervolted 1860MHz@836mV, 9500MHz MCLK, ~290W.

Temporal AA

Spoiler

Temporal Scaling

Spoiler

No performance difference, lol.

And balls to the walls setting: 2025MHz@981mV, 10875MHz MCLK, ~420W

Spoiler

Time of day:

Spoiler

cropped profiler sshots:
temporal aa ~290W


temporal upscaling ~290W


temporal upscaling ~420W

 

So the save file doesn't save time of day or sky settings. Makes sense, but a bit inconvenient for what we need.

This time grabbed screenshot properly with Temporal AA and same sky settings:

 

Do the sky settings or time of day influence the overall load seen in the profiler significantly? We can test this by maybe loading the save 5 times (for example) with the same settings, and looking at how different the profiler looks.

 

Looks like AMD is [at times] better at inhoherent rays (all the reflect ones), and worst at coherent ones (primary ray). Which is great and "logical", seems the cache does it's job.
BVH construction is probably naive for the moment, hypothetically a developer might be able to do a better version kinda quickly.

 

The difference in denoising (asvgf full) is pretty surprising. Maybe well optimized for Nvidia arch and not at all for RDNA2.

 

Or AMD is doing it all on CPU while NV is partially using GPU. Do we know how RDNA2 handle BVH builds/updates?

Denoising is pure math, is it? More flops win?

 

Excellent job with the table!

I have also now noticed your edit with 420W OC results, not the most efficient settings given gains of less than 1ms at the cost of 127W, but I love squeezing every drop of performance like that! Shame my card is so far power limited due to lack of BIOS flashing tool.
Now we need someone with 1st gen RTX card to add results ... Oh wait! I have 2060RTX in a laptop! Be right back

Had to do driver update for new version of Quake to work with Vulkan, so took a bit longer.

Please ignore wierd colors, they are fine in game, just MSI Afterburner's screen capture is not working properly.

https://ibb.co/dtzWqnV
https://ibb.co/cx56T1C

 

Sky setting doesn't change performance, just colour palette.

As for time, it doens't progress in the game, but you can toggle it via menu or by pressing "/" on the keyboard.

The results are as follows:
https://ibb.co/FHvnghk
https://ibb.co/GQSj7pV
https://ibb.co/ZYWsKSh
https://ibb.co/7v379pP

 

That looks close enough to me, despite the differences. Time doesn't vary the pattern that much.