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

 

[Kuiva maa]

I am sure recommended specs for CP2077 are based on conventional graphics workloads. Also assuming that 1060 or Fury are required for 1080p/60fps max settings, 1440p/100fps+ will require a beefier GPU. Perhaps much beefier.

 

[DeathReborn]

They will be facing the same supply issues as Nvidia this year and the next. I don't why some don't understand this.

They will sell everything they print, at whatever price they sell them at. There isn't much reason to significantly undercut nVidia in price

While we don't know how TSMC will affect supply it is widely believed to be a more reliable node, they may still not make enough but I don't doubt for a second they can make more than Samsung. AMD also doesn't have a single VRAM manufacturer making a new type of VRAM so their supply chains do have some significant differences.

I have buyers version of blue balls at this point.

Ah, the old Blue Wallet.

 

[moinmoin]

That we have heard in the past when AMD GPU was in both consoles, and still Nvidia did, with throwing enough money at developers , changed the way the PC ports ran in Windows. I would like to think now is different, but history showed that Nvidia plays their cards very well, and this time has also precedence (second generation Ray Tracing in Nvidia), so I dont think AMD can get traction in RT this generation. I just hope the developers will bother implementing both ways of doing RT

RT is a major selling point for the new console gen. Other than that graphics are considered stagnating ("only" higher resolution/frequency). And unlike with PC games we can't expect all games to offer making graphical features like RT optional, so RT will more likely be commonly integrated where it offers visible improvements while not tanking resolution/frequency.

In the past console gen Nvidia was the clear PC market leader so PC ports rather naturally introduced Nvidia specific eye candies on the top. With the suggested prevalence of RT on consoles PC implementations wouldn't be independently on the top but alongside it. And Nvidia may not stay the clear leader in the PC market depending on the revelations in two days.

 

[undertaker101]

Look at the demand there was for Ampere's top chips? It's a nearly bottomless pit.

If AMD are broadly competitive (likely) then market share will just come down to who can produce more cards.

Everyone has won and all shall have prizes!

If that same level of increased demand hits Zen3, ps5, xbx etc then there really won't be much flexibility left for AMD to produce extra GPU's vs their original plans.

The console launches likely get top priority as there's vast amounts of future revenue from selling games at stake for MS & Sony.

NV do at least have SS almost to themselves, warts and all.

Problem is also anemic supply and inventory. One only need see the evga queues to confirm this, 150-200 total cards sold in 3 weeks since system was implemented. And that is for their top of line FTW3 model on which they are making a tidy amount, others have not even begun to move. So let's not forget that, an 8 oz cup is bottomless if you fill it a drop at a time.

 

[stebler]

So MALL is Infinity Cache and it's a 128 MiB LLC as expected.

 

[Mopetar]

I have been telling people for months now, that we will not see Ampere GPUs this year.

Outside of about 10,000 or so lucky individuals you're certainly still correct about us not seeing Ampere this year.

So let's not forget that, an 8 oz cup is bottomless if you fill it a drop at a time.

More like a Dixie cup to be honest. I'm not even sure if that's a good fit for this metaphor. Even a thimble feels like a Big Gulp based on the stock we've had so far.

 

[Olikan]

So MALL is Infinity Cache and it's a 128 MiB LLC as expected.

Proof Link for the infidels...

[PATCH 3/3] drm/amdgpu/display: add MALL support

lists.freedesktop.org

 

[blckgrffn]

Proof Link for the infidels...

[PATCH 3/3] drm/amdgpu/display: add MALL support

lists.freedesktop.org

Well thanks for that, because before it made no sense to me.

Now, it makes only almost no sense to me

 

[soresu]

As was mentioned in another post, RDNA was AMD's "Zen" moment and RDNA2 will be AMD's Zen2 moment: better in some things, slower in others, but a general return to competitiveness.

The "Zen2 moment" as it were was at least as much to do with the economical versatility of the chiplets strategy as it was the raft of uArch enhancements to the core.

No doubt if RDNA2 indeed pulled that particular stunt off without any warning it would be a massive coup for the market when combined with the enhanced perf/watt on the same process.

Alas I doubt that this will be the case, sad though it makes me to admit it.

 

[soresu]

Ah, the old Blue Wallet.

I thought that they called that "burning a hole in your pocket"?

Though indeed that also sounds dirty in the wrong context.....

 

[zinfamous]

I thought that they called that "burning a hole in your pocket"?

Though indeed that also sounds dirty in the wrong context.....

penicillin usually takes care of that.

....well it used to. Us people have gotten way filthier. :\

 

[soresu]

If AMD are broadly competitive (likely) then market share will just come down to who can produce more cards.

History has shown that he who holds the APU with a gazillion design wins has the market share bonanza to rule them all.

Knowing this it is clear why AMD have prioritised mobile/embedded markets for Renoir - we can only hope that Cezanne and Rembrandt follow suit.

Rembrandt especially for (hopefully) full RDNA2 feature parity.

 

[moinmoin]

The "Zen2 moment" as it were was at least as much to do with the economical versatility of the chiplets strategy as it was the raft of uArch enhancements to the core.

No doubt if RDNA2 indeed pulled that particular stunt off without any warning it would be a massive coup for the market when combined with the enhanced perf/watt on the same process.

Alas I doubt that this will be the case, sad though it makes me to admit it.

I just saw a MALL fly right over your head...

 

[Saylick]

As was mentioned in another post, RDNA was AMD's "Zen" moment and RDNA2 will be AMD's Zen2 moment: better in some things, slower in others, but a general return to competitiveness.

Yeah, and let's hope RDNA 3 will be their Zen 3 moment but without the price hikes of Zen 3. Chiplets being introduced for RDNA 3 hopefully let's AMD achieve higher margins without actually increasing the price by all too much.

 

[Olikan]

The "Zen2 moment" as it were was at least as much to do with the economical versatility of the chiplets strategy as it was the raft of uArch enhancements to the core.

Well, it's possible that big Navi is using SOiC from TSMC, and the Infinity cache is indeed a off chip

Anyway a high end gpu with only 256 bus is kinda nuts already... there is a non-trivial economic impact of the board complexity/cost... it gets more interesting at 5nm, when SRAM shrink 30% while the memory controlers only 10%

 

[BlitzWulf]

I just saw a MALL fly right over your head...

Is crossfire about to return with a vengeance?

 

[Glo.]

Is it possible that that cache is part of Infinity Fabric?

 

[moinmoin]

Is it possible that that cache is part of Infinity Fabric?

I guess the question is whether the GPU uncore in RDNA2 is organized in a similar way as in Zen, using SCF and SDF. If yes (and the high frequencies already being reported imo points to that), MALL would naturally be part of that.

 

[Mopetar]

Is crossfire about to return with a vengeance?

If that were true it would possibly be a bigger deal than big Navi performance being as good as it is. The 3090 is the only thing that offers SLI for NVidia.

 

[PhoBoChai]

That we have heard in the past when AMD GPU was in both consoles, and still Nvidia did, with throwing enough money at developers , changed the way the PC ports ran in Windows. I would like to think now is different, but history showed that Nvidia plays their cards very well, and this time has also precedence (second generation Ray Tracing in Nvidia), so I dont think AMD can get traction in RT this generation. I just hope the developers will bother implementing both ways of doing RT

This isn't gonna change. NV sponsors the PC port, throw out RDNA2 optimizations for raster & RT, add in GameWorks or now, RTX optimizations. Viola. NV has even more money these days to do these "marketing partnership" deals with studios. But the thing is, they never get all the major games. Studios have declined to work with NV in the past, for whatever reasons.

The only good thing for AMD is that Unreal Engine 5 is very optimized for Series SX and PS5, so more indie games should run good on RDNA2.

 

[dr1337]

Do we know if the consoles also have a last level cache or is this the PC versions secret sauce?

 

[AmericanLocomotive]

Anyway a high end gpu with only 256 bus is kinda nuts already... there is a non-trivial economic impact of the board complexity/cost... it gets more interesting at 5nm, when SRAM shrink 30% while the memory controlers only 10%

Do keep in mind that 128MB of cache is a huge chunk of silicon with an accompanying massive cost. Standard SRAM uses 6 transistors per bit, so 128MB of "LLC" in a GPU would be ~6.4 Billion transistors. 6.4 Billion transistors just for the LLC, never mind any other caches that may be there. The Radeon R9 290X's Hawaii core has less transistors total than just Navi2's last level cache. That is nuts!

 

[CastleBravo]

Do keep in mind that 128MB of cache is a huge chunk of silicon with an accompanying massive cost. Standard SRAM uses 6 transistors per bit, so 128MB of "LLC" in a GPU would be ~6.4 Billion transistors. 6.4 Billion transistors just for the LLC, never mind any other caches that may be there. The Radeon R9 290X's Hawaii core has less transistors total than just Navi2's last level cache. That is nuts!

One theory I saw is that Navi21 doesn't have 128MB of cache, but instead has a large pool of shared cache that would equivalent to 128MB of non-shared cache.

 

[Hans Gruber]

You guys realize that TSMC has been given little credit for the success of AMD products in the last few years. I think TSMC should be given much of the credit for their silicon.

 

[Glo.]

You guys realize that TSMC has been given little credit for the success of AMD products in the last few years. I think TSMC should be given much of the credit for their silicon.

Yep, it was TSMCs engineers which optimized and designed AMDs architectures for this process!

It had absolutely nothing to do with AMD! There you have it guys, final confirmation how bad AMD is!