AMD RDNA3 Specifications Discussion Thread

Thing is though the numbers don't add up for dedicated matrix math units like 512, 768 or 1024 ops/clock, they claim 2.7x faster AI and the gpu is 61 Tflops which is 2.7x 23Tflops (6900XT)? Looks like they're claiming shaders as AI accelerators

It’s not tensors but the way it was presented it does sound like dedicated hardware if some sort. Maybe it’s tensors just much narrower?
 
Looks like they were targeting at least 60FPS at 4k w/ RT and FSR2.
Because FSR3 will double the FPS? So they can hit +120FPS.
Makes sense but seems a bit lacking.
 
This should go without saying, but I'm finding having to repeat this message when it never should have to be said in the first place.

Do not turn this thread into silly comparisons across different companies. This is a new technical thread so keep unwanted behaviors out of it. They will be met with escalated moderation actions.
 
It’s not tensors but the way it was presented it does sound like dedicated hardware if some sort. Maybe it’s tensors just much narrower?
WMMA instructions. Given the small perf. multiplier, it is probably designed to run on and so reuse the same 32-wide SIMD ALUs (in particular their dot-product capabilities) as much as possible to do the job, eh.
 
WMMA instructions. Given the small perf. multiplier, it is probably designed to run on and so reuse the same 32-wide SIMD ALUs (in particular their dot-product capabilities) as much as possible to do the job, eh.
Correct.
They just labeled it like that so people on plebbit stop throwing a shitfit over it having "no AI".
 
Oh no, they're gonna gut ray tracing in every engine and game they partner with. Snowdrop engine, so Avatar, he mentions Splinter Cell remake.
That Halo RT example was a weird choice to showcase... barely noticed an difference. Slightly brighter color/lighting on the armor and a darker but detailed shadow.
 
Hopefully the Hyper-RX is easy to implement and brings genuine latency improvements.
 
Back
Top