Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

Man I have been dying to make this one for a while now.

First rumours for RDNA3 are here so new thread time!

Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[TESKATLIPOKA]

It missed the cycle so the demand is none until the next year.
Better luck next time! haha

If this is true, then that's very bad. I wanted to see N32 based laptops and for how much they are sold for.

 

[adroc_thurston]

If this is true, then that's very bad

RDOA3 is very bad yes, just cataclysmically funny explosion.

I wanted to see N32 based laptops and for how much they are sold for.

Better luck next time!
haha

 

[insertcarehere]

Thanks for the slide.

Since AMD projected > 50% performance per watt increase (just months prior to launch) and got almost nothing shows that RDNA3 has a quite e bit of untapped potential that might possibly be harnessed one day by fixes in the silicon or drivers.

It's been 8 months since RDNA3 first launched and with N31, N33, and Phoenix have all been remarkably disappointing offerings on the graphics side.

Are we still doing this "RDNA 3.5" meme for N32? When are we going to acknowledge that just maybe, RDNA3 just isn't a good architecture and leave it at that?

I mean they've launched N33 just fine into a ton of attractive N23 deals so...

Have they? Laptops with N33 dGPUs have been hard to come by at best without looking through taobao and it has been months since these were released to the public (NBC's first review for a laptop with N33 was dated in February 2023).

 

[adroc_thurston]

Are we still doing this "RDNA 3.5" meme for N32?

3.5 is specifically Strix parts, it implements a few of the simpler RDNA4 features.

When are we going to acknowledge that just maybe, RDNA3 just isn't a good architecture and leave it at that?

This has absolutely nothing to do with the uArch itself and everything with how they did it on N5. It's a bonefide pre-Si miss, see their target perf/A uplift from a few pages ago.
N33 is ~around where it should be, at mobile bins anyway.

Laptops with N33 dGPUs have been hard to come by

It got meme amount of design wins given that half the stack is missing.
DT one very much launched and is available, even if it is a downclocked XL part.

 

[PJVol]

RDOA3 is very bad yes,

Bondrewd is back? Did you get banned @ b3d again? ))

 

[Tup3x]

Lumen can be made a helluva lot faster with programmable RTRT accelerators which is something we're not really having due to both h/w and DXR immaturity.

In 5.2 hw lumen is often faster while offering superior quality. It can be CPU heavy so it's not necessarily faster but it's pretty obvious that software solutions is not the way. We do need "T&L" for ray tracing and make it less CPU heavy, more dedicated hardware for ray tracing. Obviously API side improvements as well.

 

[adroc_thurston]

Did you get banned @ b3d again? ))

The product forums are still quarantined until the greens become less feral.

In 5.2 hw lumen is often faster while offering superior quality.

Yea I know but pretty shitty ROI so far.

Obviously API side improvements as well.

Gotta do this first and foremost, hoping for MS not screwing the pooch for some time longer.

but it's pretty obvious that software solutions is not the way

Yea but current DXR is a meme.

 

[RnR_au]

In 5.2 hw lumen is often faster while offering superior quality. It can be CPU heavy so it's not necessarily faster but it's pretty obvious that software solutions is not the way. We do need "T&L" for ray tracing and make it less CPU heavy, more dedicated hardware for ray tracing. Obviously API side improvements as well.

Not much faster (~10%) and not much better quality. And this on a RTX 4090. But we need to dedicate even more silicon to RT? Why not try to dedicate some silicon to optimise the current software only solution? Maybe the right hardware can run the software solution 50% faster at a 10th of the silicon needed for RT?

Anyways offtopic

 

[moinmoin]

I'm expecting a dedicated programmable RTRT API for the next console gen which then feeds into DX and Vulkan. With Nvidia not caring about making RT more flexible and mainstream beyond their own high end cards that seems to me the only feasible way changes could happen and actually get adopted.

 

[insertcarehere]

This has absolutely nothing to do with the uArch itself and everything with how they did it on N5. It's a bonefide pre-Si miss, see their target perf/A uplift from a few pages ago.
N33 is ~around where it should be, at mobile bins anyway.

Yeah and If we kick the goal posts around like that then logically speaking no architecture can ever be "bad". Might as well just say Bulldozer was a good architecture that was marred by disastrous execution at this point...

It got meme amount of design wins given that half the stack is missing.

Eliminate the Chinese brands that the rest of the world can't get and AMD's best buddy Asus and what's left?

 

[Heartbreaker]

Are we still doing this "RDNA 3.5" meme for N32?

I am.

My optimistic side sees the very long delay, as time to make it better/incorporate more features.

If they are all the exact same tech, there isn't much reason the N32 would be so delayed, unless they really just had a ridiculous excess of N21/22 chips, and it was a strategy to clear the channel. But it's been so long, I really can't see this being the reason anymore.

If you look at RX 6000 series results on both Mindfactory and Steam, it's RX 6700 XT that was their volume part in at both. It's kind of the sweet spot of the lineup, so in that sense N32 should be their most important chip, so you really would want that chip to just go MIA for no reason.

So, all I am left with is, they decided to fix/improve N32 because it's so important, and that is the reason for the delay, whether that results in an increase perf/mm2, and/or some new features to be called RDNA 3.5 remains to be seen. Or it's just the worse case, being an epic screwup, with a massive delay delivering no improvements at all.

 

[Ajay]

It would be interesting to see what silicon could come out if AMD went to Epic and asked what they would prefer to get hardware accelerated. But maybe this is already coming via their FPGA on their cpu's... if that is still coming.

FPGA on gpu's? Why not.

GPUs are basically advanced ASICS. Which are sort of pre-baked FPGA (many ASICS are prototyped on FPGAs). It would be expensive to use FPGA blocks/tiles to add features to a consumer GPUs (verification costs aren’t worth it). Consumer GPUs are refreshed every two years easier just to add fixed function blocks to the GPU, verify once and move on. Server/HPC is different because of longer life cycles and better ROI for both operators and manufacturers.

 

[PJVol]

The product forums are still quarantined until the greens become less feral.

Yeah, they reached really ridiculous levels at some point.

 

[Ajay]

Yeah, they reached really ridiculous levels at some point.

Which is too bad - was the best 3D graphics forum for along time. I learned allot there years ago.

 

[insertcarehere]

I am.

My optimistic side sees the very long delay, as time to make it better/incorporate more features.

If they are all the exact same tech, there isn't much reason the N32 would be so delayed, unless they really just had a ridiculous excess of N21/22 chips, and it was a strategy to clear the channel. But it's been so long, I really can't see this being the reason anymore.

If you look at RX 6000 series results on both Mindfactory and Steam, it's RX 6700 XT that was their volume part in at both. It's kind of the sweet spot of the lineup, so in that sense N32 should be their most important chip, so you really would want that chip to just go MIA for no reason.

So, all I am left with is, they decided to fix/improve N32 because it's so important, and that is the reason for the delay, whether that results in an increase perf/mm2, and/or some new features to be called RDNA 3.5 remains to be seen. Or it's just the worse case, being an epic screwup, with a massive delay delivering no improvements at all.

After the whole debacle with RDNA3 hype train, subsequent derailment and commentators backtracking on what they said *ahem*:

Good news, it won't.
AD102 barely edges ahead at 33% more SMs so...

You are a heck more optimistic person than I am...

 

[igor_kavinski]

So, all I am left with is, they decided to fix/improve N32 because it's so important, and that is the reason for the delay

I don't think the silicon will be different. It's just that it will be a much higher volume part than N31 while needing at least 70% of the cooling capacity so they are taking their time to improve upon and eliminate whatever mistakes led to the hotspot temp, vapor chamber and other related thermal issues. The last thing they want is people returning their cards on a massive scale due to manufacturing problems. I do wish that it is RDNA 3.5 but that is hoping too much from AMD.

 

[Heartbreaker]

After the whole debacle with RDNA3 hype train, subsequent derailment and commentators backtracking on what they said *ahem*:

I have no context/idea what that AD102 quote is supposed to show. Who is Bondrewd?

You are a heck more optimistic person than I am...

I seldom pay any attention to pre-release hype which is often driven by clickbait sites/tubers. Believing anything they say is usually a recipe for disappointment.

It just seems for what should be such an important part for AMD, to be delayed so long, there must be some constructive reason behind it. I acknowledged it could also just be an epic screwup, but to me it feels RDNA3 was the screwup and N32 is at least the partial fix.

 

[jpiniero]

It just seems for what should be such an important part for AMD, to be delayed so long, there must be some constructive reason behind it. I acknowledged it could also just be an epic screwup, but to me it feels RDNA3 was the screwup and N32 is at least the partial fix.

I think what happened is that they had no N32 laptop demand, and between that and the 7900 XT's lack of sales at MSRP, it was an easy call to not move forward with production.

Maybe the lack of N32 laptop demand was because OEMs had plenty of N22 laptops yet to be sold... but yeah it's getting late.

 

[Heartbreaker]

I think what happened is that they had no N32 laptop demand, and between that and the 7900 XT's lack of sales at MSRP, it was an easy call to not move forward with production.

Maybe the lack of N32 laptop demand was because OEMs had plenty of N22 laptops yet to be sold... but yeah it's getting late.

So you think we will never see N32?

 

[TESKATLIPOKA]

So you think we will never see N32?

We will in desktop for sure, I think even for laptops. Performance won't be bad and 16GB Vram is an advantage. It's not like we will see the next gen from both vendors sooner than 2025 for laptops.

 

[jpiniero]

So you think we will never see N32?

At this point it probably depends on how far out RDNA4 is and whether they can get some laptop OEM deals.

 

[Heartbreaker]

We will in desktop for sure, I think even for laptops. Performance won't be bad and 16GB Vram is an advantage. It's not like we will see the next gen from both vendors sooner than 2025 for laptops.

I also think they will release it, so we are back to the reason for the long delay...

12GB will be an advantage against 4060 Ti, where a cut down version will outright beat it's performance across the board, and doubly so in the few VRAM limited cases.

The 16GB advantage is more nebulous since presumable that part will compete against 4070/12GB and their really isn't as strong an argument that you need more than 12GB, and unless the do significantly better perf/CU than other RDNA3 cards, they will struggle to match 4070 performance, but if the delay improves the performance then it could make a strong 4070 competitor.

 

[GodisanAtheist]

I wonder if there really isn't anything beyond bad economics for the N32 delay (driven partly by missed performance targets), and as those economics aren't getting any better as this gen moves on.

There are fixed costs for going with a Chiplet design and at the price N32 would have to launch at and the kind of volume it would have to move to make sense just aren't working out.

What if AMD has scrapped N32 altogether and is working on a N33 style monolithic die on N6 to service this segment of the market (I know, that's nuts and no way that's going to happen etc etc etc).

 

[TESKATLIPOKA]

I also think they will release it, so we are back to the reason for the long delay...

12GB will be an advantage against 4060 Ti, where a cut down version will outright beat it's performance across the board, and doubly so in the few VRAM limited cases.

The 16GB advantage is more nebulous since presumable that part will compete against 4070/12GB and their really isn't as strong an argument that you need more than 12GB, and unless the do significantly better perf/CU than other RDNA3 cards, they will struggle to match 4070 performance, but if the delay improves the performance then it could make a strong 4070 competitor.

48CU N32 will be faster in raster than RTX 4060Ti, but in RT It will likely lose based on TPU's chart where this Nvidia is 33-38% faster at 1080p and 1440p.

16GB Vram is not so great an advantage over 12GB in desktop, true, but in laptops It should be received better. After all, you can't change the GPU in a laptop, and they are also expensive, so I would like to make It as future-proof as possible. On the other hand, weaker RT will hurt a lot.

I think raster performance comparable to RTX 4070 shouldn't be a problem for a 60CU N32, even with clocks at the level of other RDNA3 cards.
I would like to see a lot higher clocks, but the likelihood is pretty questionable.

 

[Heartbreaker]

I think raster performance comparable to RTX 4070 shouldn't be a problem for a 60CU N32, even with clocks at the level of other RDNA3 cards.

RX 7600 has less than 10% performance uplift over RX 6600 XT with the same CU count.

60CU N32 has to overcome a 12 CU (20%) advantage of the 6800XT, just to break even with it, which is about equal 4070.

Without some additional gains, it seems just breaking even with 6800XT/4070 would be the top end of expectations.

At best you have a slightly improved 6800XT. Then how do you price it?