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

 

[Aapje]

Cache dies cost less than normal logic dies as they have a lot less layers to get through.

Yeah, so bigger L3 caches are a logical choice when doing MCM.

Did I read right that the first RDNA3 card being released will be the entry model that will use a legacy GPU die and not MCM?

I haven't read anything reliable about what card comes first. But it's true that the lowest tier is a mono-die. That's not legacy though. MCM is an engineering choice with upsides and downsides. It makes more sense the larger the die.

 

[Glo.]

Not only Paul talks about RDNA3, but also touts AMD Phoenix APU: larger CU count than 16 CUs, and supposedly new caching system, with Unified L3 cache between CPU and GPU, but NOT Infinity Cache.

 

[Mopetar]

Was there really anything special about infinity cache beyond the name? My understanding was that it was basically like a large L3 cache for a GPU.

Based on what's being proposed for Navi 31 (basically 3x the shaders at 25+% faster clocks than the 6900XT) that would be one hell of a card.

I'd say there's no way that's possible, but I remember the start of the RDNA2 thread where no one thought AMD was getting to 2 GHz and everything else that seemed to impossible to even imagine.

 

[Aapje]

AFAIK, infinity cache is just a marketing term for a large L3 cache.

They might use V-cache in RDNA 3, which means stacked caching.

 

[uzzi38]

Not only Paul talks about RDNA3, but also touts AMD Phoenix APU: larger CU count than 16 CUs, and supposedly new caching system, with Unified L3 cache between CPU and GPU, but NOT Infinity Cache.

The stuff on Phoenix is wrong. He also didn't talk about what I wanted him to talk about in regards to Zen 5. What a darn shame.

 

[Mopetar]

Zen 5 is mostly locked in as far as the design is concerned, but the actual product stack will change. There might be targets at this point, but they're moving. AMD has to respond to both what they can do and what the competition does and there's not enough certainty at the moment to do that with any kind of accuracy.

Out of curiosity what information were you looking for? There's not going to be any actual silicon outside of AMD to potentially confirm anything or base speculation on at this point. I don't think we'll get any kind of real solid information for at least another year.

 

[Saylick]

The stuff on Phoenix is wrong. He also didn't talk about what I wanted him to talk about in regards to Zen 5. What a darn shame.

Care to elaborate?

 

[uzzi38]

Care to elaborate?

I've hinted at it enough already.

 

[RnR_au]

They might use V-cache in RDNA 3, which means stacked caching.

Not 'might' according to greymon.

https://twitter.com/x/status/1515586908747083779

Which makes alot of sense. Why tie up expensive N5 square millimeters with dumb cache when that can be made on N6. Assuming 3D stacking is available for those processes.

 

[biostud]

Not 'might' according to greymon.

https://twitter.com/x/status/1515586908747083779

Which makes alot of sense. Why tie up expensive N5 square millimeters with dumb cache when that can be made on N6. Assuming 3D stacking is available for those processes.

Maybe it depends on which chiplet the cache needs to be placed?
But can you use N7 cache on a N5 chiplet?

 

[uzzi38]

Maybe it depends on which chiplet the cache needs to be placed?
But can you use N7 cache on a N5 chiplet?

Not right now. There's a diagram from TSMC somewhere showing N7 on N5 comes in the latter half of this year afaik.

 

[biostud]

Not right now. There's a diagram from TSMC somewhere showing N7 on N5 comes in the latter half of this year afaik.

So what we expect is N7 vcache on zen4, and maybe on RDNA cards? Or will it be N5 cache.

 

[uzzi38]

So what we expect is N7 vcache on zen4, and maybe on RDNA cards? Or will it be N5 cache.

I'd still think N7, personally. I'm not expecting Navi31 to come to market before very late this year/early next year anyway.

All rumours thus far have suggested Navi33 is coming first to the desktop.

 

[xpea]

Not right now. There's a diagram from TSMC somewhere showing N7 on N5 comes in the latter half of this year afaik.

That's right. it's the same but technically every customer will use N6 as it's cheaper and better suited for cache than N7 (N6 is just a refined N7 optimized for cost)

 

[Saylick]

It says N5 on N5 as Q3 of this year, so I assume it includes N6/N7 on N5 as well. Given that AMD has been working in close partnership with TSMC on the stacking tech, AMD should be able to implement it right out of the gate.

Ryzen V-cache was in production at the end of last year, which aligns with the N7 on N7 timeline, and it launches roughly a little over a quarter later. N5 on N5 being Q3 implies that RDNA 3 with stacked tech launches Q4?

 

[uzzi38]

It says N5 on N5 as Q3 of this year, so I assume it includes N6/N7 on N5 as well. Given that AMD has been working in close partnership with TSMC on the stacking tech, AMD should be able to implement it right out of the gate.

Ryzen V-cache was in production at the end of last year, which aligns with the N7 on N7 timeline, and it launches roughly a little over a quarter later. N5 on N5 being Q3 implies that RDNA 3 with stacked tech launches Q4?

Either Q4 2022 or Q1 2023 yeah

 

[Aapje]

AMD has promised a 2022 release, so I expect at least a paper launch in 2022.

 

[Saylick]

Well, we'll see how fast we get RDNA 3 come September... If this is true, looks like Computex in May might be just mostly a Zen 4 launch, then we get the RDNA 3 announcement a few months later.

https://twitter.com/x/status/1516334872419192835

https://twitter.com/x/status/1516333134916161538

 

[Mopetar]

Makes a certain amount of sense to launch the monolithic die first. Also makes for more GPUs which the market could still use.

Have there been any updates on Navi 33 since what was rumored last fall?

 

[Stuka87]

I'd still think N7, personally. I'm not expecting Navi31 to come to market before very late this year/early next year anyway.

All rumours thus far have suggested Navi33 is coming first to the desktop.

Isn't TSMC moving people over to N6? Which is why most of the recent AMD cards (such as the 6500) are on N6?

 

[Saylick]

Makes a certain amount of sense to launch the monolithic die first. Also makes for more GPUs which the market could still use.

Have there been any updates on Navi 33 since what was rumored last fall?

It does. Since Nvidia is going to launch Ada Lovelace in September, from AMD's point of view, they need to make some sort of announcement at the same time so that Nvidia don't get a huge lead with the first mover advantage. There's clearly a pent-up demand for GPUs due to the sky high pricing we've seen in the last year, so whoever can tap into that demand first is going to gatekeep the other company from those same customers (i.e. people who buy one GPU ain't going to buy another GPU). If Navi 33 can be quicker to market, they should release it ASAP. If Navi 31 and 32 take longer, the least they should do is give performance claims with the rest of the line-up so that people who are on the fence between Ada Lovelace and RDNA 3 sit on the fence a little longer. Time is NOT on AMD's side in my opinion with respect to Navi 31 and 32. They need to prevent as many people from upgrading to Ada Lovelace as possible if they want to make an impact with RDNA 3. It won't matter if Navi 31 is 20% faster than AD102 if everyone already upgraded to AD102.

The latest rumors say that Navi 33 is on TSMC N6 and is roughly on par with the 6900XT (faster than N21 at 1080p, tie at 1440p, slower at 4K). RT performance should be much better than Navi 21 (some say +50% improvement), 5120 shaders, close to 3 GHz game clock, 128-bit GDDR6 memory bus (18 Gbps? 8 GB RAM?), 128 MB Infinity Cache, 180-230W TDP, cost to make is similar to Navi 22 (6700 XT) so pricing should be $400-$500.

 

[Mopetar]

If it has 5120 shaders and upwards of a 3 GHz clock it should be faster than the 6900 XT in all cases except where the smaller bus is a bottleneck.

The 6900 XT has the same number of cores, but the clock tops out at 2.25 GHz. I expect that the clocks might only hit 3 GHz for a cutdown chip, but just being able to hit 3 GHz would mean that the GPU only needs about 4,000 shaders to have similar performance to Navi 21.

 

[jpiniero]

cost to make is similar to Navi 22 (6700 XT) so pricing should be $400-$500.

Too optimistic on pricing. Need more details but those specs sound like $700.

 

[Saylick]

Too optimistic on pricing. Need more details but those specs sound like $700.

No one is going to cough up $700 for an 8GB graphics card regardless of how fast it is.

 

[Aapje]

It might actually be a brilliant strategy if Nvidia launches Lovelace with the 4090/4080 at a price that is beyond what most people are willing to spend and AMD counters (or even preempts them) with a fairly affordable and much more available 7600 that is great at 1440p and pretty easily can stretch to 4k with FSR 2/RSR in any game (as RSR doesn't need game support, although it is only FSR 1 quality).

Cost is a huge concern of gamers right now and AMD could easily present themselves as caring about the non-rich gamers and trying to solve the GPU crisis, while Nvidia looks like they want to profit off the crisis. Even though Nvidia have always launched the top cards first, these market condition are perfect for a huge backlash against them doing it for the next series. I can see it cause huge damage to Nvidia's brand, especially if a bunch of things align, like:
- AMD gets credit for lower prices
- AMD gets credit for actually being able to deliver a decent number of next gen cards
- FSR 2 gets great reviews

If the 7600 performs close to a 6900 XT, then this becomes a very attractive proposition, where many gamers will not want to wait for a 4060. Quite a few might also want to give a shot to a seemingly nicer AMD due to the backlash that I expect.