Question Speculation: RDNA3 + CDNA2 Architectures Thread

Page 30 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

uzzi38

Platinum Member
Oct 16, 2019
2,628
5,934
146

uzzi38

Platinum Member
Oct 16, 2019
2,628
5,934
146
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.
4096 shaders is the current rumour. Take Navi23 and RDNA3-ify it.
 
  • Like
Reactions: Elfear and Mopetar

RnR_au

Golden Member
Jun 6, 2021
1,708
4,157
106
Way too many conflicting rumours... and some of them just sound absolutely ridiculous.
Engaging with the tech rumour mill is like flirting... never say never...

"AMD hasn't said that RDNA3 won't boost to 4Ghz"
"AMD hasn't said that RDNA3 won't have 512MB of 3D VCache"
etc etc...

But its a hobby :p
 
  • Like
Reactions: Frenetic Pony

biostud

Lifer
Feb 27, 2003
18,248
4,757
136
Personally? About 60% confidence. I don't have all that much faith in any of the RDNA3 rumours so far. Way too many conflicting rumours... and some of them just sound absolutely ridiculous.
If the rumors are to believed somewhat, Navi31 could be the "CF as a single GPU" card with 512mb cache and 32Gb vram and a 256bit bus and priced ~$2k, then Navi32 would be 256mb cache, 16Gb vram 256 bit bus and Navi33, basically a Navi21 in RDNA3 flavor on 6nm, with 128mb cache, 192 bit bus and 12Gb vram.
 

Mopetar

Diamond Member
Jan 31, 2011
7,837
5,992
136
4096 shaders is the current rumour. Take Navi23 and RDNA3-ify it.

4K shaders at 3 GHz will perform like a 6900XT anywhere it isn't bottlnecked by the memory.

I'm curious what AMD calls it and where it ends up in the product stack though. The MCM parts should clobber it on raw core count, but obviously AMD needs something below that as well.

Seems like it would be a 7600 - 7700 level part, but who really knows.
 

Timorous

Golden Member
Oct 27, 2008
1,611
2,764
136
If the rumors are to believed somewhat, Navi31 could be the "CF as a single GPU" card with 512mb cache and 32Gb vram and a 256bit bus and priced ~$2k, then Navi32 would be 256mb cache, 16Gb vram 256 bit bus and Navi33, basically a Navi21 in RDNA3 flavor on 6nm, with 128mb cache, 192 bit bus and 12Gb vram.

I think AMD might have a top '8k' N31 sku with 32GB and 512MB cache.

I doubt the entire N31 stack will be 32GB and 512MB.

As for N33. It sounds like a 7600XT to me with just 8GB ram.
 

GodisanAtheist

Diamond Member
Nov 16, 2006
6,808
7,163
136
4K shaders at 3 GHz will perform like a 6900XT anywhere it isn't bottlnecked by the memory.

I'm curious what AMD calls it and where it ends up in the product stack though. The MCM parts should clobber it on raw core count, but obviously AMD needs something below that as well.

Seems like it would be a 7600 - 7700 level part, but who really knows.

- To tie into your comment from the AD102 thread, it would be kind of wild if this came in as the 7600 level part, with the 1.5 die part as the 7800 series and the 2 full die part as the 7900 series.

6900XT performance in a x6xx series part might put NV in a sort of awkward position in terms of what die it will use for their own xx60 and xx80 series parts: AD102 or AD104...

Its possible AMD's naming convention gets *yet another* overhaul to accommodate chiplets and isn't really anything like their naming convention so far.
 
  • Like
Reactions: Tlh97 and Mopetar

Frenetic Pony

Senior member
May 1, 2012
218
179
116
Single compute die configs make a lot of sense. Not like they don't do that with CPUs already, including codenames. Thus the two multi die models are probably referring to IO/Controller dies, and not any compute dies and cache dies which can just be moved around depending on yield/sales/whatever.

Really the 7700 rumor would be more believable as a single smaller compute die variation card. Cut the "big" rumored die by a third and you get a 6900xt more or less. Pair it with 128mb of cache and 256bit bus with like, 18gbps ram? and hey 6900lc level performance (or better) that costs less to produce. Looking at it with a few guesses you'd get over 70% better yield per wafer, something similar to the current 6700 line. Could certainly charge $500ish msrp (assuming prices keep lowering) or lower if there's a salvage die (6800xt looking thing, $429?).
 
  • Like
Reactions: Tlh97

Mopetar

Diamond Member
Jan 31, 2011
7,837
5,992
136
6900XT performance in a x6xx series part might put NV in a sort of awkward position in terms of what die it will use for their own xx60 and xx80 series parts: AD102 or AD104...

I'm assuming that the 104 die will be able to square off against Navi 33. NVidia gets the advantage of making a bigger jump going from Samsung to TSMC and the usual trend of performance moving down based on product tiers will probably hold true for the next generation of cards.

The speculation here made it sound as though NVidia would launch their top end cards first as usual, but AMD would lead with this. Neither company would have direct competition with their new generation cards. NVidia could conceivably keep selling Ampere cards, but I doubt they could keep the 3080 or 3090 priced anywhere near what they are now against a 7600 XT that has similar performance, probably at half the power consumption.
 
  • Like
Reactions: Tlh97

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
I'm assuming that the 104 die will be able to square off against Navi 33. NVidia gets the advantage of making a bigger jump going from Samsung to TSMC and the usual trend of performance moving down based on product tiers will probably hold true for the next generation of cards.

The speculation here made it sound as though NVidia would launch their top end cards first as usual, but AMD would lead with this. Neither company would have direct competition with their new generation cards. NVidia could conceivably keep selling Ampere cards, but I doubt they could keep the 3080 or 3090 priced anywhere near what they are now against a 7600 XT that has similar performance, probably at half the power consumption.

Do we know that nVidia is switching? I swear I read some place that most chips would still be samsung.
 

jpiniero

Lifer
Oct 1, 2010
14,591
5,214
136
The speculation here made it sound as though NVidia would launch their top end cards first as usual, but AMD would lead with this. Neither company would have direct competition with their new generation cards. NVidia could conceivably keep selling Ampere cards, but I doubt they could keep the 3080 or 3090 priced anywhere near what they are now against a 7600 XT that has similar performance, probably at half the power consumption.

"Where they now" is like a grand for the 3080 10 GB and mostly 1800-2000 for the 3090. Which yes, won't last against Navi 33.
 

Timorous

Golden Member
Oct 27, 2008
1,611
2,764
136
Baseless speculation time.

N31 is 2 GCDs of around 7,500 shaders each for a 15k total. It also has some MCDs for cache, there are rumours there is an IO die as well due to the talk of an odd number of chiplets.

How would you fit this together.

Well 1st lets make an assumption, the GCDs will go on top because of heat. Pretty simple, sure MLID said the same thing as well and is pretty obvious when you look at the temps of the 5800X3D despite the lower clocks and voltages.

The next assumption we need is a die size estimate for the GCD. We have 50% more shaders than N21 but no real IO or cache so no idea on that front but we can put our fingers in the air and pull numbers out of our bottom so lets do that. N21 is about 110mm^2 of IO and probably about the same again for cache leaving 300mm^2 for shaders and other stuff. 50% more shaders bumps that to 450 which with N5 we can halve if AMDs numbers are accurate to 225mm^2. Lets add 10% to account for improvements to ray tracing and larger low level caches and other changes so call it a round 250mm^2 per GCD.

This means you have 500mm^2 of silicon on the top so you need around 500mm^2 of silicon under that if you are doing 3d stacking. We know 256 bit GDDR6 + 16x PCie4 on N7 is 110mm^2 so lets just assume that is the same for N31, some savings from N6 and some losses from PCIe 5 so lets call those a wash as a guess. That means 390mm^2 of space remains under the 2 GCDs for cache. We know AMD can fit 64MB in 36mm^2 on N7/N6 (do we actually know what node that cache is made on or do we just assume N7? I don't believe there have been any official announcements have there?) In any case that means you could easily fit 10 such cache dies and the IO die under the 500mm^2 of real estate the GCDs provide for. Another option here is that AMD can just use the same 64MB cache dies they do for Milan-X and Zen3D provided it works for RDNA3, heck they could even use if for Zen 4 as well.

That makes a 512MB cache part viable from a tech point of view. I can also see a case for a 7970XT 8k halo part to top the stack. Give it 32GB of ram, 512MB of cache and fully unlocked / top binned dies and you have your $2,000 monster part. I assume AMD have ways to disable cache dies so I expect in the event that a bond fails they can disable that individual die and re-use those parts as a 7900XT with 16GB of ram and 256MB of cache and also a full complement of shaders (or maybe not for a bit more differentiation). There would also be space for a cut down actually affordable 7800XT part as well to capture partially defective GCDs.

So while I have no idea if MLID has legit sources or if his information is credible from what we do know from more reliable leakers it seems 512MB is would physically fit and it also seems like there is a viable product opportunity that can fit in that space as well.

Another possibility with such a config is a single GCD design with a 128 bit IO die and less cache. Not viable in the N33 segment now due to lack of 3d stacking throughput and cost I expect but as a laptop halo product it might work or as a refresh to N33 or gap filler part it could happen in the future once capacity increases.
 

DeathReborn

Platinum Member
Oct 11, 2005
2,746
741
136
If its a leak then it means that something is leaking, and if something is leaking, it most likely is water, and if its water leaking then most likely its a ship, and if its a ship leaking water, then we should run, because its sinking.

Every ship I served on leaked water, it's fine. Taking water... now that's a problem.
 

Frenetic Pony

Senior member
May 1, 2012
218
179
116
New hypothesis: Navi33 is a dedicated version of the 20Cu chiplet they're making for APUs. Porting designs between nodes is still cheaper than new designs. It matches up with the rumored new CU organization in RDNA3. It'd also tap out to 8+ teraflops with enough wattage while being a relatively small chip even on N6. 128bit bus, 8gb of ram.

Could be an easy slot in for $250 and less category. Slightly less powerful than a 6600 but far moreso than a 65XX. The only question I'd have is LLC size. Do you just dedicate 32mb of SRAM on there, do you put like 16 for laptop SKUs and then add a stacked chiplet for high end/desktop? Well that's probably an unimportant detail.
 

JasonLD

Senior member
Aug 22, 2017
485
445
136
Would be a major disappointment if both N31 and AD102 end up performing less than 2X of current top-end GPUs. (Which is what I am expecting btw)