AMD: RDNA 3 Speculation, Rumours and Discussion

[DegustatoR]

No such thing.
Below N31 goes N32 and after that a single die N33.

Which doesn't make much sense from production perspective so we'll have to wait and see here.

 

[Leoneazzurro5]

But these lower tier chips will be single chiplet presumably and it's just double the LLC per chiplet from what I've read? Not that that's still not awesome obviously.

My expectation at this stage is that the single chiplet N31 will be the direct replacement for the 6900XT with 50% more ALU, double the LLC, maybe faster clocks and improved IPC etc... for a more traditional performance uplift. But then we get a 6900XT X2 equivalent GPU at the very top end which harks back to the old Crossfire on a card days where they are just stupid prices for clearly Halo products. The difference here being that hopefully scaling is much less like 2 GPU's in Crossfire and more like 1 big GPU, with hopefully no compatibility issues. So outside of that one crazy expensive halo "X2" card, the rest of the stack will be a more traditional performance uplift of around 50% (a bit more if we're very lucky).

Single chiplet N31 has no memory controller or cache on its own, as N33 should be the monolithic, we can suppose it would be very similar to N21, with the architectural and process improvements.

 

[Bondrewd]

Which doesn't make much sense from production perspective

?
There's a big-big reuse point between 31 and 32.

 

[DegustatoR]

?
There's a big-big reuse point between 31 and 32.

So 32 will have MCD as well?

 

[Leoneazzurro5]

No such thing.
Below N31 goes N32 and after that a single die N33.

Now N32 is a very interesting topic of its own. So it would be basically a cut down N31 with a different/cut down I/O + LLC chip stacked on it, or is it made with diffreent GCDs (in this case, one less SE per GCD could be a possible configuration)?

 

[Bondrewd]

So 32 will have MCD as well?

Bingo.
Upper two parts are a funny slaughterfest, anything lower is single die and still funny enough.

So it would be basically a cut down N31

That would still be called Navi31.
We don't call different configs of the same part differently.

or is it made with diffreent GCDs

this

 

[Leoneazzurro5]

OK, so basically N32 should be a (around) 40 WGP (160 "old CUs") parts (two dies) stacked on the same I/O die of N31 (possibly with cut down LLC capacity). Is this correct?

 

[Bondrewd]

stacked on the same I/O die of N31

Oh not really.
It's not PVC or anything similar.

Is this correct?

Probably?
That config is plenty far away.

 

[DegustatoR]

OK, so basically N32 should be a (around) 40 WGP

If N31 is 30x2 then N32 won't be more than 30.

Bingo.
Upper two parts are a funny slaughterfest, anything lower is single die and still funny enough.

Well that makes more sense although N32 looks like an over-engineered part existing solely to make a point in this case. It will be interesting to see how it would compare to a single die solution of similar complexity.

 

[Bondrewd]

N32 won't be more than 30.

who said that?

Well that makes more sense although N32 looks like an over-engineered part existing solely to make a point in this case

Nah, that's 31.
32 exists to rob you of your moneh.

It will be interesting to see how it would compare to a single die solution of similar complexity.

Don't think a 160SM single die is really feasiable.

 

[Leoneazzurro5]

1 SE=10WGP, so N31=2x3SE= 60 WGP. If you remove 1 SE per die, then you get 2x2SE=40 WGP.

 

[DegustatoR]

1 SE=10WGP, so N31=2x3SE= 60 WGP. If you remove 1 SE per die, then you get 2x2SE=40 WGP.

Ah so it's still a 3 die solution with just some disabled parts. Even less interesting then from production POV.

 

[Bondrewd]

3 die solution

Not really.

just some disabled parts

Different GCD hencewhy different part number.

 

[DegustatoR]

Different GCD hencewhy different part number.

We're back to point A then with no apparent production benefits in such approach.
N31 will have 2 unique GCDs and an MCD.
N32 will have 1 unique GCD and the same MCD? Okay, some savings here probably?
N33 and below will be unique single chip GPUs.
What's the point in chiplets here besides the idea to make a card which will consume 500W and cost $2500?

 

[Bondrewd]

What's the point in chiplets here besides the idea to make a card which will consume 500W and cost $2500?

Literally the point.
It's not about costs or anything.
It's about winning more.
Shipping beyond reticle devices and all.

 

[Wesker]

@Bondrewd This is slightly off-topic, but do you know if Apple will remain a client of AMD's GPU solutions for Navi 2x/CDNA/Navi 3x?

The dual Radeon Vega II's and first-get Navi cards will need replacing in the Mac Pro. It's very doubtful that Apple's silicon engineering group will be able to come up with a GPU solution with similar, let alone better, compute performance.

 

[trinibwoy]

It's 60 WGPs but each WGP now has 256 ALUs, whereas one WGP in Navi 21 has 128 ALUs.

5120 ALUs on Navi 21 vs. 15360 ALUs on dual-chiplet Navi 31. Furthermore, Navi 31 has up to 512MB Infinity Cache, 4x the 128MB in Navi 21.

We should also expect N31's clocks to reach higher frequencies, since it's made on N5P instead of Navi 21's N7P.

It should be >3x Navi 21, even if power consumption jumps to RTX 3090 levels.

Got all that but we were talking about ray acceleration, not raw flops.

Your stance is correct if the ratio between RA and WGP is the same, I think Jawed was supposing that the ratio between RA and ALU stays the same, which could also be. Or it may be that ratio between RA and WGP is indeed the same, but the RA capabilities are increased... There are too few details atm for having a definitive answer. I would find very strange, however, if AMD increased the base shading power almost threefold (which is not the limit of RDNA2) while keeping Ray Tracing hardware (which is the weakest point of RDNA2) with moderate increase.

Yeah lots of assumptions of course this early in silly season. The thing is we don’t know if the ray accelerators themselves are a weak point. It could be that they’re starved on BVH data due to insufficient cache size/bandwidth. Or the bottleneck could be in the shader core. That would be strange given how wide the shaders are you would expect them to easily keep up with whatever the relatively narrow Ray accelerators throw at them.

Or maybe RDNA3 will do full traversal in the RA’s

 

[Bondrewd]

This is slightly off-topic, but do you know if Apple will remain a client of AMD's GPU solutions for Navi 2x/CDNA/Navi 3x?

Pretty sure we should've seen a Mac Pro N21 by now but still mia.
CDNAs can't draw tris nor do they have display cores thus useless for Mac Pro purposes.

It's very doubtful that Apple's silicon engineering group will be able to come up with a GPU solution with similar, let alone better, compute performance.

They're still doing it, duh.
Should be next year?
N5p and many-many cores.

 

[Nebuchadnezzar]

@BondrewdIt's very doubtful that Apple's silicon engineering group will be able to come up with a GPU solution with similar, let alone better, compute performance.

It's been their goal for literally half a decade now.

 

[Wesker]

It's been their goal for literally half a decade now.

Long before the release of the M1, we knew that the CPU performance of Apple's A-series processors overtook their x86 counterparts (when normalised for core count and power consumption). Anandtech has a series of articles documenting this.

The GPUs are a different story, however. While the GPUs in the A12, A13, and A14 (and M1) are good, they're just "kinda meh" when compared to their Nvidia and AMD counterparts.

I don't doubt that they have the potential of releasing a competitive GPU that can go toe-to-toe with the big chips from Nvidia and AMD. But it's not happening anytime soon, imo.

Pretty sure we should've seen a Mac Pro N21 by now but still mia.
CDNAs can't draw tris nor do they have display cores thus useless for Mac Pro purposes.

They're still doing it, duh.
Should be next year?
N5p and many-many cores.

Blame Intel for most of the delays. Rumour has it that Apple will be releasing a refresh to the Mac Pro soon, and that they were waiting on Intel for safe delivery of either Ice Lake or Sapphire Rapids. After a series of long delays, it looks like they'll have to go with Ice Lake. So they'll probably release N21 cards for the IL refreshed Mac Pro.

Interesting...