Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[Qwertilot]

All likely makes sense when mediated by the overall die shortage.

Current policy definitely seems to be to release one die at a time, leaving a long time gap between each. They means they need to try to stretch things slightly to cover the market.

 

[beginner99]

Why would AMD offer the full bus from 6700XT to 6600XT? Seems a bit dumb which makes me assume there are some underlying technical reasons as to why. With reduced bus they could offer a 8 or 10GB 6600XT based on N22. in that segment 12GB is just a waste really.

 

[GodisanAtheist]

I wonder if this whole generation is going to last longer on account of the shortages. Why launch a brand new card when demand will be higher for longer for the old one since people cannot buy them?

Will be interesting to see if the 4xxx/7xxx generation gets pushed back a bit to make sure everyone that wants a current gen card buys one before they drop the next round...

 

[Leadbox]

Why would AMD offer the full bus from 6700XT to 6600XT? Seems a bit dumb which makes me assume there are some underlying technical reasons as to why. With reduced bus they could offer a 8 or 10GB 6600XT based on N22. in that segment 12GB is just a waste really.

Maybe the costs of differentiation by memory segmentation are no more or less than just leaving the bus intact and with full memory amount.

 

[moinmoin]

I wonder if this whole generation is going to last longer on account of the shortages. Why launch a brand new card when demand will be higher for longer for the old one since people cannot buy them?

That's not the POV AMD has. It sells everything it makes. Renoir is also seeing more demand than supply, and it's still in process of being replaced by Cezanne and Lucienne.

 

[GodisanAtheist]

That's not the POV AMD has. It sells everything it makes. Renoir is also seeing more demand than supply, and it's still in process of being replaced by Cezanne and Lucienne.

- Not disagreeing, but just to BS for a bit: I do wonder if it would give AMD an opportunity to move their core business (CPUs) down to 5nm where they have a really sweet margins despite the cutting edge node while leaving their RDNA2 GPUs to be produced on an older 7nm process for an opportunity to increase their margins, compete on price, increase their marketshare, and recoup some R&D costs.

Maybe launch top tier RDNA 3 on 5nm for the margins, but really slow drip everything else as volume opens up on the cutting edge node.

Would be nice to have product lines not competing with each other for the same fab space. But AMD is definitely not content to be the second player in any of its markets anymore, so I figure everything will filter down to the newest node as soon as it is reasonably possible, as soon as possible.

 

[moinmoin]

- Not disagreeing, but just to BS for a bit: I do wonder if it would give AMD an opportunity to move their core business (CPUs) down to 5nm where they have a really sweet margins despite the cutting edge node while leaving their RDNA2 GPUs to be produced on an older 7nm process for an opportunity to increase their margins, compete on price, increase their marketshare, and recoup some R&D costs.

Maybe launch top tier RDNA 3 on 5nm for the margins, but really slow drip everything else as volume opens up on the cutting edge node.

Would be nice to have product lines not competing with each other for the same fab space. But AMD is definitely not content to be the second player in any of its markets anymore, so I figure everything will filter down to the newest node as soon as it is reasonably possible, as soon as possible.

My impression is AMD would want to move all its own products to N5 as quickly as possible since demand for consoles is the biggest internal competitor for N7 fab space. Also with the introduction of IC RDNA2 seems ominously close to enabling MCM GPUs so if RDNA3 realizes this step it will give AMD much more flexibility. So my expectation is that AMD won't slow down in either of those two areas.

 

[Krteq]

RX 6700 XT

2589 MHz Avg. at stock
2809 MHz Avg with OC

 

[GodisanAtheist]

My impression is AMD would want to move all its own products to N5 as quickly as possible since demand for consoles is the biggest internal competitor for N7 fab space. Also with the introduction of IC RDNA2 seems ominously close to enabling MCM GPUs so if RDNA3 realizes this step it will give AMD much more flexibility. So my expectation is that AMD won't slow down in either of those two areas.

- I hadn't considered the IC angle, and if it is a stepping stone to chiplet based GPUs in RDNA 3 (I am skeptical that we'll see chiplets in RDNA3, but if we do it would be a huge coup for AMD) then it would definitely make sense to design a small GPU chiplet on the most up to date process and crank out a truckload of them as fast as possible.

I we're still looking at monolithic GPUs for RDNA3...

 

[andermans]

The other very strong driver for keeping the GPU and CPU on the same node is APUs. Unless the capacity constraints are very much at 5nm and not 7nm it would make a lot of sense to keep them in sync so you don't have to port a GPU to another process (adding cost and possibly also delays).

 

[Gideon]

Ok, FidelityFX Super Resolution is much further away than I anticipated

AMD FidelityFX Super Resolution (FSR) to launch this year for PCs

Even though it’s progressing well, we still have more work to do and not only internally but with our game developer partners. We want to launch it this year. We belive we can do that this year, but at the same time we a lot more work ahead of us. We need to make sure the image quality is there. We need to make sure it can scale from different resolutions. And at the same time that our game developers are happy with what we are producing.

— Scott Herkelman

This sounds more like "we hope we can deliver it this year, but it might very well slip". Based on that quote it's ridiculous to expect it before Q4 (and I think November/December is the most likely). Regarding GPU lifecycle that's horrendeous. Essentially FSR will be released when RDNA3 is already just a few quarters away.

 

[moinmoin]

Ok, FidelityFX Super Resolution is much further away than I anticipated

AMD FidelityFX Super Resolution (FSR) to launch this year for PCs



— Scott Herkelman

This sounds more like "we hope we can deliver it this year, but it might very well slip". Based on that quote it's ridiculous to expect it before Q4 (and I think November/December is the most likely). Regarding GPU lifecycle that's horrendeous. Essentially FSR will be released when RDNA3 is already just a few quarters away.

Honestly Herkelman's quote makes it sound like they are still squarely in the research phase looking for adequate solutions. They could find something tomorrow or not find anything workable for current GPUs in the next decade.

 

[Glo.]

Ok, FidelityFX Super Resolution is much further away than I anticipated

AMD FidelityFX Super Resolution (FSR) to launch this year for PCs



— Scott Herkelman

This sounds more like "we hope we can deliver it this year, but it might very well slip". Based on that quote it's ridiculous to expect it before Q4 (and I think November/December is the most likely). Regarding GPU lifecycle that's horrendeous. Essentially FSR will be released when RDNA3 is already just a few quarters away.

The DLSS competitor from AMD will come with the "yearly" driver, at the earliest. No chances of it coming to market before the December's large update, that we are getting on the yearly schedule.

 

[Glo.]

Honestly Herkelman's quote makes it sound like they are still squarely in the research phase looking for adequate solutions. They could find something tomorrow or not find anything workable for current GPUs in the next decade.

As per RedGamingTech sources, the development of FFSS, or whatever its called, is in very early stages of development: Alpha stages. The code is sent already to the game developers to work on it.

 

[Stuka87]

I rather them ship it when its a bit more flushed out then try and rush it out where they just end up with bad publicity, and a feature none of us can use.

 

[Mopetar]

I rather them ship it when its a bit more flushed out then try and rush it out where they just end up with bad publicity, and a feature none of us can use.

Well it's not like most of us have cards that can use it anyways, so even if it doesn't come out until the end of the year, it's not like most people are missing much.

 

[GodisanAtheist]

Having something good yesterday would have been ideal, but having the tech come out and work for RDNA2 cards on the tail end of their life is really where it would be most useful.

Will be interesting to see what AMD ends up with here, and if it's going to be more like DLSS 1.0 (no motion vectors) or 2.0 built into the engine with motion vectors).

It'll be like we're back in the early 2010's with 5-6 different Anti-Aliasing options available, each with their up and downsides.

 

[Leeea]

Well it's not like most of us have cards that can use it anyways, so even if it doesn't come out until the end of the year, it's not like most people are missing much.

I have a card that could use it! .... but I have yet to play a game that I would use it with.

I definitely care, but as long as they get it out before Cyberpunk goes on sale on steam for less then $10 everything should be good.

 

[Qwertilot]

Having something good yesterday would have been ideal, but having the tech come out and work for RDNA2 cards on the tail end of their life is really where it would be most useful.

Will be interesting to see what AMD ends up with here, and if it's going to be more like DLSS 1.0 (no motion vectors) or 2.0 built into the engine with motion vectors).

Or nothing - as people have noted. It is worth repeating just how hard what they're trying to do is. If they could simply 'clone' DLSS then of course they'd manage but without the tensor units that isn't really possible/effective.

So it really isn't obvious what they might be doing. It'd have to be something that people have been overlooking for a decade or two.

 

[Dribble]

RDNA 2 is gonna get whatever the consoles get - whatever MS/Sony manage to design. However MS/Sony have been doing this for many years and must have close to maxed what you can do for upscaling with shaders and gpu cores. The breakthrough is the AI tensor cores. I expect them to be added to AMD cards at some point (RNDA 3?) and only then is a real DLSS equivalent possible. Even then it will be tough for AMD as they don't really do software, and this software is complex. It's probably taken Nvidia about 5 years of R&D to get to DLSS 2. In addition MS/Sony won't have much incentive to help until they have consoles with AI cores too.

 

[moinmoin]

I doubt AMD is working on any software solution that needs hardware RDNA2 or rather the subset the RDNA based consoles don't have. The incentive is exactly that it doesn't only helps on PC but also on the now current gen consoles (that are bound define the technical base for AAA games for years to come once the last console gen stops being supported).

 

[Glo.]

RDNA 2 is gonna get whatever the consoles get - whatever MS/Sony manage to design. However MS/Sony have been doing this for many years and must have close to maxed what you can do for upscaling with shaders and gpu cores. The breakthrough is the AI tensor cores. I expect them to be added to AMD cards at some point (RNDA 3?) and only then is a real DLSS equivalent possible. Even then it will be tough for AMD as they don't really do software, and this software is complex. It's probably taken Nvidia about 5 years of R&D to get to DLSS 2. In addition MS/Sony won't have much incentive to help until they have consoles with AI cores too.

FFSS is not going to be only for RDNA2, but also for RDNA1 GPUs, and it remains to be seen - maybe for GCN4(Polaris and Vega).

 

[Mopetar]

I still think the technology is rather inane. I mean I suppose it would be nice if you could run it on 5600 XT and get frame rates like you had an $800 GPU, but . . .

*checks eBay*

Well damn, the 5600 XT is an $800 GPU right now.

Even this generations midrange cards like the 3060 Ti even runs games extremely well without using DLSS, so I'm not sure why anyone would buy the card with the idea that they'd use it, unless it was well into the later years of the card when it can no longer keep up with newer, more demanding titles like it used to be.

We're headed back to the days about arguments over image quality because both companies are eventually going to try to squeeze out more performance at the expense of how the end result looks. Frankly I think it should be treated as a nice extra feature and be completely excluded from reviews to avoid encouraging heading down that path any sooner than necessary.

 

[GodisanAtheist]

FFSS is not going to be only for RDNA2, but also for RDNA1 GPUs, and it remains to be seen - maybe for GCN4(Polaris and Vega).

- If it can run on shaders alone, it should really work with anything with programmable shaders (presumably DX12 compliant), not even just AMD.

Only way it would be RDNA2 only is if it required some proprietary element of the RDNA2 arch, such as the IC. But given PS5 and XBX don't have IC, I don't see that being the case.

Seeing how AMD want to release it for the 6900XT first, it looks like they're hoping the additional shader overhead leaves enough resources to actually run the game until they can refine a 2.0 version that isn't competing for as many resources.

That's my 2 BS's.

 

[Panino Manino]

IPC comparison between 40CU: https://www.computerbase.de/2021-03/amd-radeon-rdna2-rdna-gcn-ipc-cu-vergleich/

Somehow, RDNA2 is behind RDNA1 in most cases.
It's just immature drivers?