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

 

[Timorous]

The memory bandwidth question, i.e. can RDNA 3 be fully fed, is the million dollar question (2 million dollars now, adjusted for inflation).

Goal = 3x more effective bandwidth over N21

We have, thus far....
- 50% wider memory bus
- 11% higher memory clocks
- Higher bandwidth on the Infinity Cache
- Better caching algorithms so that the Infinity Cache is better utilized
- End-to-end data compression

Might be an effective 2x at the end of the day, hence why we should only expect a doubling over N21.

N21 is not memory bandwidth bound though so >2x is doable.

The 6950XT gains less than the memory bandwidth increase and if it was memory bound it would see gains closer to the bandwidth increase.

 

[Saylick]

And maybe a 3D cache version with the double amount of cache? Or is that rumor buried again?

I think that is still happening, at least according to main man @Kepler_L2. Full N31 + 192 MB Infinity Cache. My guess is that the same benefits that Ryzen got using V-cache is the same as what will happen with N31, i.e. a few cycles of latency increase with V-cache but 2x the size and thus higher hit rate.

 

[fleshconsumed]

Man, all these rumors make me anxious to see N32 performance/pricing figures... November can't come soon enough.

 

[GodisanAtheist]

I think we all accept that we're not going to be getting the current gen "low end" any time soon thanks to the stockpile of existing cards out there combined with the demand to upgrade off of ancient 9xx/10xx era cards.

I do hope we get to see not just N31 (and maybe even some derivatives of it) but also N32 and derivatives as well.

I will be both shocked and delighted if we get an N33 launch (even announcement) as well if AMD feels like their current gen stocks have sufficiently depleted by November.

 

[Mopetar]

View attachment 68003

Words because of forum rules

I wish my GPU could run on hype, because the forums generate enough to easily power two 3090s in SLI.

 

[Hans Gruber]

I still have my 7950 in a working computer (2013). I regret not buying an RX5700 in 2019 for around $300. I think I will be getting an RDNA3 card this time around. I like the beefy Vram that AMD cards have. I use 3x monitors and they would probably benefit from 16GB or more Vram.

 

[moinmoin]

I will be both shocked and delighted if we get an N33 launch (even announcement) as well

I think we get it. For mobile.

 

[Mopetar]

I still have my 7950 in a working computer (2013). I regret not buying an RX5700 in 2019 for around $300. I think I will be getting an RDNA3 card this time around. I like the beefy Vram that AMD cards have. I use 3x monitors and they would probably benefit from 16GB or more Vram.

If you had bought a 5700 for that in 2019 you'd still have the 7950 because you'd have sold the 5700 when they were going for ~$1,000.

 

[psolord]

I wish my GPU could run on hype, because the forums generate enough to easily power two 3090s in SLI.

...and memes. Man that Lovelace thread is firing on all 16 cylinders. 8 physical and 8 with dlss3.

 

[Timorous]

Man, all these rumors make me anxious to see N32 performance/pricing figures... November can't come soon enough.

N32 is pretty exciting.

AD104 has 7680 shaders and roughly matches the 6950XT with 5120 shaders at around 2.3Ghz.

N32 has 7680 shaders but with 3Ghz+ clockspeeds is offering 50% more shaders and 40% more clockspeed than the 6900XT which is a 2.1x increase in compute. At the low end I expect that to be equate to a 50% increase in performance over the 6900XT and if clocks are 3.5Ghxz or so on N32 then it may be more. That does land it between the 4080 16GB and 4090 in raster performance.

Also according to Dylan in his semi analysis article the BOM cost of full N32 is only 10% more than the 4080 12GB yet it should be faster than the 4080 16GB.

 

[Kaluan]

I think that is still happening, at least according to main man @Kepler_L2. Full N31 + 192 MB Infinity Cache. My guess is that the same benefits that Ryzen got using V-cache is the same as what will happen with N31, i.e. a few cycles of latency increase with V-cache but 2x the size and thus higher hit rate.

Zen3 V-Cache trippled the LLM, and while it's questionable if 288MB would bring any tangible gains over 192MB... if they can do the specialized 2x SRAM density libraries here again, I don't see why not just go straight to 96 + 192 route if applicable. It'll cost the same.

I think we all accept that we're not going to be getting the current gen "low end" any time soon thanks to the stockpile of existing cards out there combined with the demand to upgrade off of ancient 9xx/10xx era cards.

I do hope we get to see not just N31 (and maybe even some derivatives of it) but also N32 and derivatives as well.

I will be both shocked and delighted if we get an N33 launch (even announcement) as well if AMD feels like their current gen stocks have sufficiently depleted by November.

As long as 1. RDNA3 doesn't feature lock something (unlikely) and 2. RX 6000 prices continue to drop... I'm okay with that. But an N33 in Q1 2023 at the latest would still be preferred. Wonder how long 'till N34 rumors start poping up.

Also if 512bit dual compute die RDNA3 chip does exist... what are the odds of AMD doing the right thing and codenaming it N32X? 😂

N32 is pretty exciting.

AD104 has 7680 shaders and roughly matches the 6950XT with 5120 shaders at around 2.3Ghz.

N32 has 7680 shaders but with 3Ghz+ clockspeeds is offering 50% more shaders and 40% more clockspeed than the 6900XT which is a 2.1x increase in compute. At the low end I expect that to be equate to a 50% increase in performance over the 6900XT and if clocks are 3.5Ghxz or so on N32 then it may be more. That does land it between the 4080 16GB and 4090 in raster performance.

Also according to Dylan in his semi analysis article the BOM cost of full N32 is only 10% more than the 4080 12GB yet it should be faster than the 4080 16GB.

Is that for the entire board or just the die/packaging? N32 also has more memory, and I doubt 20Gbps GDDR6 is cheaper than bottom bin GDDR6X, particularly when there's 33% more of it and 33% more pin connections and traces on the PCB from the wider bus.

But no doubt about it. Full N32 will be at least 1 performance tier above "4070 Ti".

 

[Tigerick]

N32 is pretty exciting.

AD104 has 7680 shaders and roughly matches the 6950XT with 5120 shaders at around 2.3Ghz.

N32 has 7680 shaders but with 3Ghz+ clockspeeds is offering 50% more shaders and 40% more clockspeed than the 6900XT which is a 2.1x increase in compute. At the low end I expect that to be equate to a 50% increase in performance over the 6900XT and if clocks are 3.5Ghxz or so on N32 then it may be more. That does land it between the 4080 16GB and 4090 in raster performance.

Also according to Dylan in his semi analysis article the BOM cost of full N32 is only 10% more than the 4080 12GB yet it should be faster than the 4080 16GB.

Which Dylan are you refering to?

I have done my calculation: cost of N32 with 3 MCD seems slightly less than AD104....and with double pumped full SP, I am really excited what AMD brings to table

 

[coercitiv]

Which Dylan are you refering to?

This: Ada Lovelace GPUs Shows How Desperate Nvidia Is - AMD RDNA 3 Cost Comparison

 

[Kaluan]

Which Dylan are you refering to?

I have done my calculation: cost of N32 with 3 MCD seems slightly less than AD104....and with double pumped full SP, I am really excited what AMD brings to table

Patel.

Well, you wouldn't be wrong. Cut down N32 can certainly be quite a bit cheaper than the BOM of full AD104, while also trumping it decisively. As I suspected, it's only the increased memory cost that makes full N32 slightly more expensive.

2 things I got from that article,
AMD can maintain similar margins on a near full/full fat N31 to nVidia's insane 4090 ones by pricing their flagship $1200.
And N33 BOM ~ AD107 BOM 😨
Say goodbye to affordable next gen lower end cards from nVidia. AMD may save the day however (H1 2023+), but I doubt they'd want to.

Here's the article BTW:

Ada Lovelace GPUs Shows How Desperate Nvidia Is - AMD RDNA 3 Cost Comparison

Nvidia's first cost disadvantage versus AMD in a decade

semianalysis.substack.com

 

[moinmoin]

Say goodbye to affordable next gen lower end cards from nVidia. AMD may save the day however (H1 2023+), but I doubt they'd want to.

I suspect AMD will push hard cards with performance above the current gen. Anything below is best served by second hand now with crypto crashing. The big exception is mobile where due to the lack of competition by an exploding second hand market the full range can be offered.

 

[jpiniero]

This: Ada Lovelace GPUs Shows How Desperate Nvidia Is - AMD RDNA 3 Cost Comparison

I saw that... while I definitely agree that the chip part is decently cheaper, I feel like he's underselling the packaging costs a bit while also seemingly thinking that GDDR6X costs more than the new Samsung GDDR6+ that AMD is presumably using. I have no idea how you would be able to compare the memory cost difference other than number of chips.

 

[Mopetar]

Unless AMD and NVidia just postpone their low end GPUs they'll have to compete with the flood of used previous generation GPUs that were locked up in mining rigs until now. Even the sale of new high end 4080s and 7900s will put cheap 3080s and 6900s into the channel.

It doesn't matter what AMD and NVidia want to charge, it's what the market will let them.

 

[jpiniero]

Unless AMD and NVidia just postpone their low end GPUs they'll have to compete with the flood of used previous generation GPUs that were locked up in mining rigs until now.

That is what I expect them to do. You would think that N33 will be popular enough with mobile OEMs to avoid what happened with N24.

 

[Mopetar]

I'm assuming that outside of gaming notebooks that most will be happy with AMD's APUs and won't include dedicated graphics since the graphics power in the APUs is good enough for 1080p e-sport gaming and running other mainstream titles at lower settings.

We might even see a flood of used notebooks since there were some miners that were buying up gaming notebooks with mobile 3070s in order to mine.

 

[Glo.]

I'm assuming that outside of gaming notebooks that most will be happy with AMD's APUs and won't include dedicated graphics since the graphics power in the APUs is good enough for 1080p e-sport gaming and running other mainstream titles at lower settings.

We might even see a flood of used notebooks since there were some miners that were buying up gaming notebooks with mobile 3070s in order to mine.

Well, in 1080p, high(!) presets, Radeon 680M was already averaging 120 FPS in Overwatch.

And that game is heavily CPU dependent. If RMB's CPU would be closer in performance to Ryzen 5000 series, desktop, than to Ryzen 3000 series desktop, we could look at even above 120 FPS 1% lows in this game and presets.

Now, imagine what Phoenix will do for e-sports gaming, especially if CPU will deliver proper 5000 series desktop performance at minimum.

Who will need anything else, above it, for E-Sports?

And that is like, what? 80% of gaming market currently?

 

[jpiniero]

Well, in 1080p, high(!) presets, Radeon 680M was already averaging 120 FPS in Overwatch.

And that game is heavily CPU dependent. If RMB's CPU would be closer in performance to Ryzen 5000 series, desktop, than to Ryzen 3000 series desktop, we could look at even above 120 FPS 1% lows in this game and presets.

Now, imagine what Phoenix will do for e-sports gaming, especially if CPU will deliver proper 5000 series desktop performance at minimum.

Who will need anything else, above it, for E-Sports?

And that is like, what? 80% of gaming market currently?

Well... OEMs don't seem to care. And seem to love the 3050L. You can see why they wanted to try to replicate that with N33.

 

[Glo.]

Well... OEMs don't seem to care. And seem to love the 3050L. You can see why they wanted to try to replicate that with N33.

They will care about PHX .

 

[Kaluan]

I saw that... while I definitely agree that the chip part is decently cheaper, I feel like he's underselling the packaging costs a bit while also seemingly thinking that GDDR6X costs more than the new Samsung GDDR6+ that AMD is presumably using. I have no idea how you would be able to compare the memory cost difference other than number of chips.

So you're dissing his insider math data based on just your outsider presumptions? Ok then.

The 20Gbps GDDR6 that AMD is presumably using is not "GDDR6+", not sure where you got that from, or that they're Samsung. SK Hynix actually seems to have beat Samsung to 20Gbps mass production, it's more reasonable to believe that they will use Hynix, at least with the first batch.
GDDR6+ is just Samsung marketing, and it seems to be their future 24 and 27Gbps modules, which use a new fabrication method. SK Hynix also seems to have 24 and 27Gbps GDDR6 in the works. RDNA3 refreshes might use these.

Anyway, 16GB N32 has 33% more memory capacity than 12GB "4080", he clearly shows that's just ~20% higher BOM cost. Which makes sense, 21-23Gbps GDDR6X is not cheaper than 20Gbps GDDR6, not sure why anyone would think otherwise.

He clearly also weighs packaging costs, which obv should cost more for AMD (for N32 and N31)... but it's all outweighed by cheaper dies and less expensive cooling designs. Are we gonna pretend designing and manufacturing cooling systems than have to cool 450W+ is cheap? No wonder stuff like Lenovo is going 4 slot for their 4090 design lol

The benefit of more efficient designs also mean less cost dedicated to keeping it nice and cool and fault-free.

 

[jpiniero]

Samsung claims they have 20 and 24 gbps coming soon.

Anyway, 16GB N32 has 33% more memory capacity than 12GB "4080", he clearly shows that's just ~20% higher BOM cost. Which makes sense, 21-23Gbps GDDR6X is not cheaper than 20Gbps GDDR6, not sure why anyone would think otherwise.

He's claiming that GDDR6X is more expensive. Maybe it is. I don't know.

 

[TESKATLIPOKA]

One thing I find peculiar.

	N33	N32	N31
Shader Engines	2	3	6
Shader Arrays	4	6	12
Number of WGPs	16	30	48
WGPs per Shader engine	8	10	8
Shaders	4096	7680	12288

N32 having 32WGP(8192 shaders) would make more sense to me. That way, N32 would have 8 WGPs per shader engine.