AMD RDNA3 Specifications Discussion Thread

Still wondering what's the difference between 2nd gen RT and 1st gen.
Mentioning larger caches is off topic, imo.
'RT featuers for perf. & eff.' does not mention what's the features.
1.8x @ x GHz sounds like comparing per CU vs. the old gen at same frequency, which would be the speedup i would have expected in the best case. But benches suggest it's comparing the whole GPUs.

Well, ISA docs will tell...
The slides talk about up to a 1.5x boost per CU (https://images.anandtech.com/doci/17638/AMD RDNA 3 Tech Day_Press Deck 25.png). So the 1.8X must be for the full GPU.

But I found this qualifier: "Based on November 2022 AMD internal performance lab measurement of rays with indirect calls on RX 7900 XTX GPU vs. RX 6900 XT GPU." (https://www.anandtech.com/Gallery/Album/8202#70)

So it sounds like some kind of synthetic benchmark.
 
The slides talk about up to a 1.5x boost per CU (https://images.anandtech.com/doci/17638/AMD RDNA 3 Tech Day_Press Deck 25.png). So the 1.8X must be for the full GPU.
1.5 x per CU aligns with the wins we get from dual issue. So i would rant gen2 == gen1.
But we saw new RT instructions, so there must be something...

It sucks we wait for announcement, but then actual information is sparse, and it's mostly about marketing hogwash like image upscaling, frame interpolation, and all those 'progress indicators' which are just software. Speaking of both AMD and NVs recent GPU announcements.

Oh - wait - i've found one more while randomly clicking on this slide deck:
1667662182978.png
What could it mean? hmmmmm.... : )

However. I really like their slogan 'Worlds most advanced Gaming GPU'. What a nice sidekick. :D
 
Last edited:
What could it mean? hmmmmm.... : )

However. I really like their slogan 'Worlds most advanced Gaming GPU'. What a nice sidekick. :D
It means this:
The slides talk about up to a 1.5x boost per CU (https://images.anandtech.com/doci/17638/AMD RDNA 3 Tech Day_Press Deck 25.png). So the 1.8X must be for the full GPU.

But I found this qualifier: "Based on November 2022 AMD internal performance lab measurement of rays with indirect calls on RX 7900 XTX GPU vs. RX 6900 XT GPU." (https://www.anandtech.com/Gallery/Album/8202#70)

So it sounds like some kind of synthetic benchmark.
 
Anyone else seeing parallels between Fermi GF100 and N31? That is, there were lots of underlying changes at once while on a new process. GF100 had very poor clock and power scaling. To reach its original design/performance goals, Nvidia had to do a "quick" respin to get the anticipated performance, but had to ship something, so it shipped the 480 in March 2010. But in November 2010, Nvidia releases the 580, a respun GF100, which clocked higher and used less power and, I expect, represented Nvidia's original expectations for what became the 480.

AMD says RDNA 3 was "architected to exceed 3ghz" and no doubt that is true. But N31 doesn't appear to get near that mark based on the performance numbers and the footnotes from AMD's presentation. If I were a betting man, I'd lay money on the issue with power and clocks arising from the chiplet implementation. I wouldn't be surprised if the monolithic N33 exceeds 3ghz with relative ease.

Meanwhile, perhaps we can expect AMD to address whatever the bottleneck is on clocks and in a few months time release a 7950 XTX (or 7900 3Ghz Edition as someone mentioned).
 
It means this:
Oops, i've missed the first link from the post.
But the question is, if they have (full?) HW traversal as mentioned, can it be this only gives +50% perf, even if CU is free to work on other things in parallel?
That's hard to believe, so i guess they only have some instructions to help with traversal. But it's hard to imagine what those instructions might do and why they should be needed.

That's the most puzzling RDNA3 question still open...
 
At least we can expect a refresh next year with potential for a reasonable perfomance jump
 
Back
Top