GPU performance in ...

if you are ( I hope not ) play "games" then it is easy to find a zillion tests of any GPUs either for your "game" of choice or infer from a bazillion tests done for games with a similar engine / whatever... however if I wish to find a test of GPU for any raw converter ( or even FRV ) I think it is a much difficult task ... granted on some forums you can dig some threads with some semblance of tests or raw converters like C1 dump some cryptic numbers in logs that people do post too... naturally users of raw converters are a minority vs gamers, however - is there any way to approx. a performance of a GPU in any "game" tests towards at least some raw converters - of course it is understandable that GTX870 will be well beaten by GTX4090 - no need for tests.. but what if I want to have some idea how GTX4070 fares vs RX 6800 XT in raw converters - any chance to draw any conclusions from "game" tests ?

PS: yes, there are some tests published by sites like Puget Systems - but it is not exactly they test Camera Raw performance in stuff like Ai NR

I can point you to a raw processor that uses a GPU pipeline for the entire raw processing toolchain, from raw import to JPEG export:

github.com/hanatos/vkdt

It'll process my D7000 test NEF in about 18ms...

It's Linux-only right now, written by the original author of darktable. It organizes the processing shaders into modules that are connected using a node graph, which others have already used to define graphs for complex operations like hdr merging. It'll also read and render video streams.

it was not about that - but if one can infer how GPUs might perform in various raw converters from gaming tests for example ... it might as well be a stupid question, but I want to ask

FLOPS ought to be a reasonable metric to use, in the absence of specific tests. That's a basic measure of the number crunching ability of that GPU

Lies, damn lies, and benchmarks.

I'd think since vkdt does all it's processing in the GPU, it could be used as a "gold standard" against which to compare others. A lot of software selectively uses the GPU for just certain operations, will probably process the same image slower fur to both CPU operations and the marshaling in/out the GPU. Pick a single raw to use, to eliminate uninteresting variables...

The release of LrC V12.3 with AI Denoise has shown us that grunt power and VRAM is just part of the story. Ray tracing, tensor cores etc. are so crucial to the brave new world of AI. It is this that, in Nvidia speak, makes the RTX blow away the GTX.