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

 

[leoneazzurro]

It was a big subject until GN said, basically, you're using it wrong. At least you don't have to wait for Nvidia to make your card work, you can just carefully plug it in anytime.

DP1.4 does suck. But it isn't a deal breaker.

Nor is this SW issue for most of the people who buys an high end card, and DP 1.4 may be a deal breaker for people wanting to play 4K at higher refresh rates. In any case, this only demonstrates that certain people want AMD to be competitive only to buy Nvidia at a cheaper price.

 

[gdansk]

Nor is this SW issue for most of the people who buys an high end card, and DP 1.4 may be a deal breaker for people wanting to play 4K at higher refresh rates. In any case, this only demonstrates that certain people want AMD to be competitive only to buy Nvidia at a cheaper price.

It demonstrates that some people expected RDNA3 CU to be generally as fast as AMD advertised in raster. But unfortunately that was a bunch of cherrypicked games.

High refresh rate 4K will have to use HDMI. And any good monitor should include both. Heck, my 4K monitor - which isn't good - works fine at 144Hz using 2x DP1.4 cables.

 

[leoneazzurro]

It demonstrates that some people expected RDNA3 CU to be generally as fast as AMD advertised in raster. But unfortunately that was a bunch of cherrypicked games.

I was not speaking about expectations but actual product value in terms of performance/price. But then someone else deviated the discussion about other things.

 

[gdansk]

I was not speaking about expectations but actual product value in terms of performance/price. But then someone else deviated the discussion about other things.

In performance per dollar it is a better than Nvidia. But for me that's not the key metric. It's the relative improvement from my current RTX 3080 per dollar spent. And in that metric the 4090 turns out to be slightly superior.

 

[leoneazzurro]

In performance per dollar it is a better than Nvidia. But for me that's not the key metric. It's the relative improvement from my current RTX 3080 per dollar spent. And in that metric the 4090 turns out to be slightly superior.

Maybe, but rest of the world would use the perf/dollar ratio, which is another matter.

 

[gdansk]

Maybe, but rest of the world would use the perf/dollar ratio, which is another matter.

Not really, I think most people are upgrading from existing systems. And buying Nvidia in the end... so it doesn't appear that perf/dollar is the deciding metric for most buyers.

 

[DXDiag]

It's 5-10% faster than 4080

No it isn't .. TPU, Computerbase, pcgameshardware and Hardwareunboxed all say otherwise. Stop sugarcoating the situation. It's bascially a tie in raster, with a notable fall in ray tracing.

 

[Bigos]

Well, duh. That's were the "N31 broke" part comes in.

Are there still people around who are gonna tell me N31 is working as intended and the OG leak clocks and performance projections were just based on pure fabrications?

When we have data like this now?

View attachment 72763

Mesh shader is known to be "optimize this workload for one specific GPU arch". A new GPU arch is not being optimized the same as the old one - oh noes!

This graph makes as much sense as testing Zen 4 with AVX disabled, as in some software did not recognize Zen 4 and tested it on baseline x86-64.

It's just Mesh Shader being a terrible optimization target.

Mesh Shading for Vulkan

Deploying and developing royalty-free open standards for 3D graphics, Virtual and Augmented Reality, Parallel Computing, Neural Networks, and Vision Processing

www.khronos.org

 

[Saylick]

Not really, I think most people are upgrading from existing systems. And buying Nvidia in the end... so I doesn't appear that perf/dollar is the deciding metric for most buyers.

I've always wondered what portion of consumers only care about AMD's competitiveness only so that they could buy Nvidia at a lower price. Similarly, if AMD has such a small market share of the GPU market, does AMD's pricing of their GPUs even matter in the grand scheme of things, as in does AMD pricing even affect Nvidia pricing? If not, then is the hate towards AMD being not competitive even justified? It like people rooting against the underdog in a sport where the last ten seasons was won by the crowd favorite, but then also whining that the sport is dying.

 

[leoneazzurro]

No it isn't .. TPU, Computerbase, pcgameshardware and Hardwareunboxed all say otherwise. Stop sugarcoating the situation. It's bascially a tie in raster, with a notable fall in ray tracing.

Computerbase is the only one with a perfect tie, the other ones give a 2-4% advantage in 4K to the 7900XTX. And other reviews have bigger delta depending on the game choices and other factors (i.e. with Zen4 and resizable bar on there is a sensible increase). And there is the 200$ saving factor. I can stop "sugarcoating" but doomsayers should stop as well.

 

[Kaluan]

https://twitter.com/x/status/1602328567190372354

They also say the hardware faults are less pronounced with scaled down designs. So maybe N33 and/or Phoenix IGP can eek a better showing.

Mesh shader is known to be "optimize this workload for one specific GPU arch". A new GPU arch is not being optimized the same as the old one - oh noes!

This graph makes as much sense as testing Zen 4 with AVX disabled, as in some software did not recognize Zen 4 and tested it on baseline x86-64.

It's just Mesh Shader being a terrible optimization target.

Mesh Shading for Vulkan

Deploying and developing royalty-free open standards for 3D graphics, Virtual and Augmented Reality, Parallel Computing, Neural Networks, and Vision Processing

www.khronos.org

Oh yeah I bet RX 7900 not being able to properly cull geometry has no baring on the performance we've seen being lower than expected.

 

[Geddagod]

Not really, I think most people are upgrading from existing systems. And buying Nvidia in the end... so it doesn't appear that perf/dollar is the deciding metric for most buyers.

DIY vs Prebuilt buyers. I fully believe most people who build themselves go for best bang/buck, but people who just buy prebuilts or build a pc without any real input from YouTube or knowledgeable PC- enthusiasts (which I am going to say are pretty rare since most people don't just *know* how to build a pc) go Nvidia by default.
AMD needs to step up its marketing game imo, otherwise this scenario is going to stay the same till the end of time. Either that or go for the perf crown for 2-3 generations to build mindshare.

 

[gdansk]

DIY vs Prebuilt buyers. I fully believe most people who build themselves go for best bang/buck, but people who just buy prebuilts or build a pc without any real input from YouTube or knowledgeable PC- enthusiasts (which I am going to say are pretty rare since most people don't just *know* how to build a pc) go Nvidia by default.
AMD needs to step up its marketing game imo, otherwise this scenario is going to stay the same till the end of time. Either that or go for the perf crown for 2-3 generations to build mindshare.

Radeon team needs to step up its execution. It seems the neat features from the new RDNA3 don't work or aren't used.
Then they need to keep focusing on software too. Weird to see such bizarre power bugs in 2022.
Then they need to fix their marketing by not cherry picking abnormal results.
They're actually falling behind in all respects relative to 2 years ago at this time.

But it seems they focused on last generation's problem: availability. This should be an easier part to make in volume than the 6900 XT was.

But it's probably more expensive to make than the RTX 4080.

 

[Gideon]

If N32 doesn’t improve something, is there any reason to release it?

N31 has close to same price/performance to current 6800/6900 cards, so N32 has to be an improvement in this regard.

It probably will be improved upon, if rumors are to be believed.

Even if ends up with the same price/performance, economics would still dictate its release:

A 200 mm² 5nm die with 4x small 6nm chiplets (reusable over multiple products) has got to be cheaper to produce than a 520 mm² 7nm chip. You can get roughly 3x the Navi 32 dies from a single wafer than Navi 31 ones (presuming no defects, otherwise it's should be more) and while the MCD chiplets aren't free, it probably still ends up cheaper. Even if the costs end up in the same ballpark, it would still make sense, as AMD can then split their GPU production over 2 different nodes, freeing up 7nm/6nm wafers for Navi23 and later Navi 33 production. Otherwise there would be a hard cap on production, as Navi 21 is massive.

But overall I agree that It would be much better if top-of-the-line Navi 32 ends up much closer to 7900 XT (similarly to the 6750XT vs 6800 scenario) while being considerably cheaper. Say, a 5-10% gap at 1440p at most.

 

[Timorous]

It probably will be improved upon, if rumors are to be believed.

Even if ends up with the same price/performance, economics would still dictate its release:

A 200 mm² 5nm die with 4x small 6nm chiplets (reusable over multiple products) has got to be cheaper to produce than a 520 mm² 7nm chip. You can get roughly 3x the Navi 32 dies from a single wafer than Navi 31 ones (presuming no defects, otherwise it's should be more) and while the MCD chiplets aren't free, it probably still ends up cheaper. Even if the costs end up in the same ballpark, it would still make sense, as AMD can then split their GPU production over 2 different nodes, freeing up 7nm/6nm wafers for Navi23 and later Navi 33 production. Otherwise there would be a hard cap on production, as Navi 21 is massive.

But overall I agree that It would be much better if top-of-the-line Navi 32 ends up much closer to 7900 XT (similarly to the 6750XT vs 6800 scenario) while being considerably cheaper. Say, a 5-10% gap at 1440p at most.

If N32 is mostly fixed I could see N32 being within 5% of the 7900XT at current performance levels because the 7900XT is only around 15% faster than the 6950XT.

I do think though that there is a lot of performance to extract from the drivers and perhaps time spent working around hardware bugs was less time spent on performance tuning and extracting the ILP for the dual issue shaders.

 

[TESKATLIPOKA]

https://twitter.com/x/status/1602328567190372354

They also say the hardware faults are less pronounced with scaled down designs. So maybe N33 and/or Phoenix IGP can eek a better showing.

From that twitter only Phoenix and N32 will be capable.
N33 has the same problem as N31 and is made on a worse process. Not sure how It can perform as RX 6700XT, unless only 1/3 of shader engines and CU od N31 actually helps.

 

[insertcarehere]

But overall I agree that It would be much better if top-of-the-line Navi 32 ends up much closer to 7900 XT (similarly to the 6750XT vs 6800 scenario) while being considerably cheaper. Say, a 5-10% gap at 1440p at most.

As per Computerbase, 7900 XT (84CU) at 2.2ghz is approximately 9% faster than 6900XT (80CU) at 2.3ghz:

This points to roughly a 9% increase per CU from Navi 21 to Navi 31, iso-clocks.

From RX 6800 (60CU Navi 21) to the 7900XT is ~55% faster in both 1440p and 2160p.

Even if Navi 32 at 60CUs (as per Angstronomics who nailed specs for Navi 31 first) can clock 40++% over the 6800 (well over 3ghz), it then needs to be assumed that performance scales linearly with clock speeds (rather unlikely), while on equal/less bandwidth (faster clocked GDDR6 at 256bit but half the infinity cache).

It might be possible but I am not holding my breath on this.

 

[amenx]

--

 

[Gideon]

As per Computerbase, 7900 XT (84CU) at 2.2ghz is approximately 9% faster than 6900XT (80CU) at 2.3ghz:

This points to roughly a 9% increase per CU from Navi 21 to Navi 31, iso-clocks.

From RX 6800 (60CU Navi 21) to the 7900XT is ~55% faster in both 1440p and 2160p.

Even if Navi 32 at 60CUs (as per Angstronomics who nailed specs for Navi 31 first) can clock 40++% over the 6800 (well over 3ghz), it then needs to be assumed that performance scales linearly with clock speeds (rather unlikely), while on equal/less bandwidth (faster clocked GDDR6 at 256bit but half the infinity cache).

It might be possible but I am not holding my breath on this.

You bring up very good points, and I'm still not sold on it, but just wanted to add some facts you didn't mention:

• The 9% increase per CU already has a 37,5% deficit in infinity cache factored in (RX 6900 XT vs RX 7900 XT), 50% isn't that big of a jump from there IMO.
• While the RDNA3 cache is smaller, it actually has nearly 2x the bandwidth of the RDNA2 cache
• The RX 6750 XT has only 40CUs (vs 60), about 23% higher clocks than the RX 6800 XT and a 25% smaller Cache (with no offset in bandwidth).
  • FLOPS wise it's also 25% slower but at 1440p it's only about 11% slower in actual games. True enough, at 4K the gap goes up to 20%.
• We don't know how the hardware bugs affect performance.
  • It might not just be the Voltage/Power curve but those driver workarounds might cause very real performance regressions per clock as well (e.g. Phenoms infamous TLB bug workaround)
• And lastly (though it is a weak one, I admit) the drivers only make use of the VLIW2 in a limited number of games yet AFAIK. Hopefully the gap increases a bit in the coming months.

Overall I do agree that only a 5% gap at 4K looks very optimistic and would absolutely require clocks north of 3Ghz. I can totally see it at 1440p though (which could be the more realistic resolution if AMD prices the card right at <700€)

 

[Gideon]

Agreed. I think it will even be "passable" if it has about the same raster performance of the RTX 4080 (say 6800XT vs RTX 3080, within 5% of each-other) just nothing spectacular. In that case RT performance will equal 3090 only in lighter RT games (such as Doom Eternal, etc).

The 3Dmark leaks seem to indicate this is the plausible worst-case scenario. I obviously hope it will do somewhat better, but this is where I set my expectation bar currently, not to be disappointed again.

What is it with those "you can't be that pessimistic, therefore you are wrong" downvotes?

I wrote this just before the reviews in order to set my expectations low and not to be disappointed, which i clearly stated.

I absolutely nailed the raster-performance and while I was overly pessimistic about the RT perf, it certainly wasn't an inconceivable outcome (as I was leaing on AMDs own slides, where the RT uplifts were the same or a bit smaller than raster uplifts, and mostly in rather RT light games).

Luckily I was wrong and the RT performance uplift was bigger than the raster performance one. It ended up having about the RTX 3090 RT perf in very demanding games and RTX 3090 Ti perf in lighter or more AMD favored ones (with outlier like spiderman reaching 4080 performance).

All in all, those "prepare for the worst" expectations were totally justified. How can they be "wrong"?

 

[eek2121]

I hope we get a post mortem on this launch, part of the reason AMD is discussed more is the variability of their launches as well. NV is a known quantity, we generally know what they're going to do and there is a good amount of rabble rousing when they miss the mark (2080Ti only managing a 30% raster improvement over the prior gen).

I hope so too, but you know we won’t. All the big tech companies tend to be quiet about stuff like this.

Idk I think I'm the target demo for high end cards (way overpaid for a 3090 during the shortage, constantly buying dumb electronics, etc) and was already put-off enough with the 4090 to not get one. The physical size, price, and already high power consumption made it very unappealing. Someone I know who also has a 3090 is in the same boat.

I think Nvidia has found the limit for the average rube who buys high end. There will be an audience for even more ridiculous cards but I can't imagine it's very big.

I paid EVGA’s MSRP for the 3090FTW 3 Ultra during the shortage. I had hoped for an FE card, so technically I overpaid?

I will gladly pay $1,500 or so for a GPU, but they have gotten so large and power hungry, I won’t. I moved to a mini ITX system and I am not going back.

It demonstrates that some people expected RDNA3 CU to be generally as fast as AMD advertised in raster. But unfortunately that was a bunch of cherrypicked games.

High refresh rate 4K will have to use HDMI. And any good monitor should include both. Heck, my 4K monitor - which isn't good - works fine at 144Hz using 2x DP1.4 cables.

Some monitors don’t support high refresh rates via HDMI, like the Samsung G9.

Mesh shader is known to be "optimize this workload for one specific GPU arch". A new GPU arch is not being optimized the same as the old one - oh noes!

This graph makes as much sense as testing Zen 4 with AVX disabled, as in some software did not recognize Zen 4 and tested it on baseline x86-64.

It's just Mesh Shader being a terrible optimization target.

Mesh Shading for Vulkan

Deploying and developing royalty-free open standards for 3D graphics, Virtual and Augmented Reality, Parallel Computing, Neural Networks, and Vision Processing

www.khronos.org

Mesh shaders are not uncommon. AMD dropped the ball here.

I've always wondered what portion of consumers only care about AMD's competitiveness only so that they could buy Nvidia at a lower price. Similarly, if AMD has such a small market share of the GPU market, does AMD's pricing of their GPUs even matter in the grand scheme of things, as in does AMD pricing even affect Nvidia pricing? If not, then is the hate towards AMD being not competitive even justified? It like people rooting against the underdog in a sport where the last ten seasons was won by the crowd favorite, but then also whining that the sport is dying.

Valid competition holds market prices down while driving innovation on both sides. I personally don’t care as much about price as I do features, size, noise, and perf/watt

Radeon team needs to step up its execution. It seems the neat features from the new RDNA3 don't work or aren't used.
Then they need to keep focusing on software too. Weird to see such bizarre power bugs in 2022.
Then they need to fix their marketing by not cherry picking abnormal results.
They're actually falling behind in all respects relative to 2 years ago at this time.

But it seems they focused on last generation's problem: availability. This should be an easier part to make in volume than the 6900 XT was.

But it's probably more expensive to make than the RTX 4080.

Agreed except on that last part. AMD has a considerable cost advantage. NVIDIA is using N4 on a larger die. AMD is using a mix of N5 and N6.

Any AIB reviews yet?

 

[Panino Manino]

Again, barely matching the 4080 (just a few percent faster at most) is not the "between 4080 and 4090" in raster, far from it.

Changing topics, AMD will never "match" Nvidia in RT with this hybrid approach, it needs more dedicated hardware, which reminds me those IA accelerators. Couldn't it be used for FSR freeing shader resources? Is being used for what right now?

 

[Topweasel]

Agreed except on that last part. AMD has a considerable cost advantage. NVIDIA is using N4 on a larger die. AMD is using a mix of N5 and N6.

Any AIB reviews yet?

Not only a mix, but by splitting them up overall yields go up. I was worried about bandwidth on the IF connects and latency. But in the end even if they affected performance they seemed to get a pretty good balance out it.

 

[exquisitechar]

Agreed except on that last part. AMD has a considerable cost advantage. NVIDIA is using N4 on a larger die. AMD is using a mix of N5 and N6.

I think there is no cost advantage. AD103 is only 379 mm^2 of N5 related silicon and 4080 is not even using the full die. Navi 31’s GCD is 306 mm^2 of N5, along with 6 MCDs which are 37.5 mm^2 of N6. Add in its chiplet associated costs and higher VRAM amount and I think it’s clear there is no cost advantage for Navi 31.

The cost advantage would have materialized if Navi 31 was out of the 4080’s league in performance and closer to the 4090, against which it does have a massive cost advantage. Instead, it’s barely faster than the 4080 in rasterization and is well behind in RT performance. Still, AMD will be more than fine selling it at $1000 and it’s not a disaster like Vega, it’s just not better than Ada in that regard at all. If the performance was 25% higher like some are suggesting it should have been, then things would be different.

 

[Kaluan]

Sorry, I just don't see "RDNA3 & Lovelace = same node" as a hill worth dying on. Particularly if whoever insists that is the case was also someone who kept harping on about how Samsung 8nm is "oh so inferior to TSMC N7P"

AMD is on a slight node backfoot, just like nVidia was last gen. The gap may be ever so slightly smaller, but it's there. And I don't exactly get what the point of pretending otherwise is.

Anyway, it's pretty clear execution and (partly) software are the culprit here. Not the node(s).

And N31 RDNA3's problems clearly go beyond just missing clock targets.


BTW AIB 7900 card reviews are out. And my 2 takeaways are
1. Bigger OC headroom than we've seen since many generations ago.
2. Building on the above, I think this stems from the vestiges of initial clocking expectations. It's clearly a 3GHz+ on air architecture, but the v/f & efficiency curve just wasn't there for them in N31's current state to market a 10-20% faster N31 AND the 50% better power efficiency.