AMD RDNA3 Specifications Discussion Thread

LordEC911

Regular
Do we need a RDNA4 speculation thread?

Rumor mill is talking about RDNA4 being ahead of schedule...
If they don't do anything too crazy with RDNA4's chiplet and/or interconnect... there are some (potentially) crazy silly-season thoughts bouncing around in my head.
 

Bondrewd

Veteran
If so, that would mean a mere APU introducing the next-gen CPU and GPU IPs together.
RDNA3 dGP to APU gap is like 3 months already so absolutely very much warranted as of now.
They've being aiming to win mobile for years and now they shall.
 

eastmen

Legend
Supporter
Why would AMD be announcing a even larger chip anyways? Raster performance isnt something they have to improve on so much, maybe clock speeds are enough there.
More cu's would offer better ray tracing performance. maybe a 100 cu chip will let them catch up with the 40x0 series
 

eastmen

Legend
Supporter
And the next gpu 200cu? At how many watts? Its not the way forward.

So you are saying they will never add more CU's ? Why did they increase CU counts from last gen to this gen then ?

There can certainly be a spot for a larger chip using the same additional chiplets that the current 7900xtx offers while offering more cus to get better performance.

Even NVidia has 2 dies the 4080 16/12 gig die and the 4090 die with what 250m tranistors more ?
 

Kaotik

Drunk Member
Legend
Supporter
So you are saying they will never add more CU's ? Why did they increase CU counts from last gen to this gen then ?

There can certainly be a spot for a larger chip using the same additional chiplets that the current 7900xtx offers while offering more cus to get better performance.

Even NVidia has 2 dies the 4080 16/12 gig die and the 4090 die with what 250m tranistors more ?
The cancelled 4080 12 used GA104, 4080 16 uses GA103 and 4090 GA102.
GA102 has 30.4 billion transistors more than GA103 and GA103 10.1 billion transistors more than GA104 (assuming TPU transistor counts are correct)
 

trinibwoy

Meh
Legend
Supporter
Sure. You would only need to be about 3-4x faster to be on par on RT.

It’s tricky because a lot of RT performance is dependent on general compute. RT optimization research is all about reducing the number of rays actually cast and filling in the gaps with intelligent math. RESTIR GI for example spends about the same amount of time casting rays as it does on the sampling, validation & projection. Other techniques only spend 10-20% of the total time casting rays and 80% on voodoo shaders.

So even if Nvidia 2x faster at that 20% the overall performance can be much closer.
 

DegustatoR

Veteran
It’s tricky because a lot of RT performance is dependent on general compute. RT optimization research is all about reducing the number of rays actually cast and filling in the gaps with intelligent math. RESTIR GI for example spends about the same amount of time casting rays as it does on the sampling, validation & projection. Other techniques only spend 10-20% of the total time casting rays and 80% on voodoo shaders.

So even if Nvidia 2x faster at that 20% the overall performance can be much closer.
I mean it's not like we have many games with RT where Nv is about 2+ times faster right now. Oh, wait, we do have such games and they are in fact the majority of such games with those where the gap is smaller being the exception more than a rule. And I wager that the number of such games will increase over time as RT will be used more and more prominently. So yeah while shading power is of course important for RT performance the actual RT part has already become the main limiter here and even being 2x faster in "rasterization" likely won't be enough to beat Nv h/w in RT. And you'd only need to double Navi 31 complexity to do so.

And I have honestly hoped that these fantasies on doing RT with FMA ALUs have left us for good around RDNA1 times.
 

trinibwoy

Meh
Legend
Supporter
I mean it's not like we have many games with RT where Nv is about 2+ times faster right now. Oh, wait, we do have such games and they are in fact the majority of such games with those where the gap is smaller being the exception more than a rule. And I wager that the number of such games will increase over time as RT will be used more and more prominently. So yeah while shading power is of course important for RT performance the actual RT part has already become the main limiter here and even being 2x faster in "rasterization" likely won't be enough to beat Nv h/w in RT. And you'd only need to double Navi 31 complexity to do so.

And I have honestly hoped that these fantasies on doing RT with FMA ALUs have left us for good around RDNA1 times.

I think you missed my point entirely. The optimization techniques I mentioned aren’t in any games yet. Also this is mostly relevant for diffuse GI. There’s less trickery possible with glossy reflections and there’s no substitute for casting more rays.
 

DegustatoR

Veteran
I think you missed my point entirely. The optimization techniques I mentioned aren’t in any games yet. Also this is mostly relevant for diffuse GI. There’s less trickery possible with glossy reflections and there’s no substitute for casting more rays.
I don't think that we'll be going into the direction of magic optimization techniques making RT lighter with time at all. Even if (that's a big "if" at this point in time) in some cases such optimizations would still happen whatever performance you will get from this will end up being spent on adding more RT into the frame most likely.

We are already in a situation with a varying ratio of games using RT "lightly" and games using RT "heavily" - but all of them will end up using RT eventually and if the last several years are anything to go off the number of "heavy" RT titles will increase with time, not diminish.
 
Top