Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

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...

forums.anandtech.com

 

[Timorous]

peanut dies, pretty easy pricing decisions.

Maybe I will be pleasantly surprised then.

 

[jpiniero]

peanut dies, pretty easy pricing decisions.

Expensive node though.

 

[maddie]

Expensive node though.

Back from vacation?

 

[jpiniero]

Back from vacation?

No... maybe haven't posted in this thread in a bit?

It's like people like to conveniently ignore the whole MLID thing and think wafer prices at TSMC are the same on every node.

 

[adroc_thurston]

Maybe I will be pleasantly surprised then.

Yeah gonna be alright.

Expensive node though.

N5 wafer pricing ain't that high anymore.

 

[jpiniero]

N5 wafer pricing ain't that high anymore.

Doesn't seem like it, when TSMC revenue was basically flat despite shipping a decent amount of less wafers (which I will guess is mainly N6)

 

[SolidQ]

Probably 300 and 500 bucks.

No cutdown versions for 200/400?

 

[Saylick]

$529 for 7900XT raster and 7900XTX-ish RT performance w/ 16 GB of RAM isn't bad at all imo. Like @adroc_thurston said, the price point shouldn't be difficult to decide given that the die is only 240mm2. Plenty of meat on the bone for AMD even if it were $499.

 

[SolidQ]

7900XTX-ish RT performance

Still mystery. Without BVH XTX level is ok, with BVH can be up 2x in heavy scenarios.
We only know from PS5pro is BVH8

 

[Mahboi]

Probably 300 and 500 bucks.
Maybe 329 and 529.

That's actually pretty great. 550 for a >7900 xt perf and ~4070 Ti RT perf or even 4070 Super is nothing to complain about (apart from the lack of ambition in chip size).

Interesting when RDNA4 will be announce Computex or Gamescon?

Kind of same question over here, because without question Computex is for Z5, and I don't believe AMD is going to be doing another double announcement all at once.
Gamescom would put us more than 2 months behind, RDNA 4 would be a Sept product. Which is kind of in line with what has been told around here I guess...

 

[MrTeal]

Why sell for $300 and $500 when you can launch at $400 and $600, get middling reviews and then drop the price to $300 and $500 two weeks later?

 

[adroc_thurston]

Why sell for $300 and $500 when you can launch at $400 and $600, get middling reviews and then drop the price to $300 and $500 two weeks later?

Because GB205 is somewhere next year and they need to time the MSS bump properly now that consoles are down bad.

 

[Mahboi]

Why sell for $300 and $500 when you can launch at $400 and $600, get middling reviews and then drop the price to $300 and $500 two weeks later?

Well they have been pricing pretty terribly because no matter how low they sell, people just buy NV anyway, so after awhile they just burn margins.
Could definitely go harder though.

Because GB205 is somewhere next year and they need to time the MSS bump properly now that consoles are down bad.

Down bad?

 

[jpiniero]

Down bad?

While the third year is usually the peak (this is the 4th year)... allegedly Sony has cut production in half and MS has temporarily stopped production entirely.

 

[Mahboi]

That's shocking. Not for MS, their console lineup is ridiculous. But Sony did a great job with the PS5, for the price that it has, you have a serious package. I remember Iceberg Tech trying to do a console equivalent PC video 3 whole years after, it was scraping the bottom of the barrel with the pinkie to try and get a very limited XSS equivalent.
If even that doesn't sell, I don't know what will.
I don't think the PS5 Pro will change that all that much either with their higher price, even with RDNA 3/4 in it.

 

[branch_suggestion]

I've assumed for some time that RDNA 3's success greatly hinged on RDNA 2's as well. RDNA 2 polished AMD's public image, and RDNA 3 enjoyed the afterglow of that.
How much are we talking accelerating RDNA 5? Dropping all the 3 top SKUs really accelerates next gen by 6 months, by a year?

Like 3 months, might be enough for a late 2025 release but don't hold your breath.

 

[Mahboi]

Like 3 months, might be enough for a late 2025 release but don't hold your breath.

Well, even if it's early 2026, it's at least better than RDNA 3 -> late 2022, RDNA 4 -> late 2024, etc.

 

[phoenix21]

Why sell for $300 and $500 when you can launch at $400 and $600, get middling reviews and then drop the price to $300 and $500 two weeks later?

RX 7700XT and RX 6750XT are up for $400 and $330 respectively on newegg, both 12GB cards. Do you really think an 8GB card worth single penny above $300 in 2024? And a GPU of N44 wouldn't be good for ray tracing either.
$250 is the most optimistic price for N44. $270 seems plausible for replacing RX 7600.

 

[branch_suggestion]

RX 7700XT and RX 6750XT are up for $400 and $330 respectively on newegg, both 12GB cards. Do you really think an 8GB card worth single penny above $300 in 2024? And a GPU of N44 wouldn't be good for ray tracing either.
$250 is the most optimistic price for N44. $270 seems plausible for replacing RX 7600.

There will be a 16GB N44 clamshell, pricing will be key for that. 8GB has a pretty strict price ceiling.

 

[Mahboi]

RX 7700XT and RX 6750XT are up for $400 and $330 respectively on newegg, both 12GB cards. Do you really think an 8GB card worth single penny above $300 in 2024?

When the 6750xt gets beaten squarely by N44, it'll sell.
I agree that 8Go is really weak in 2024 though. It's barely cutting it for 1080p, and the card is definitely capable of 1440p. Then again it's a tiny 140mm² die...hard to shove a big bus on it, if it used GDDR7 I'd definitely expect it to use 3Go chips but it's on GDDR6.

And a GPU of N44 wouldn't be good for ray tracing either.

You must be n-oh first message. Welcome.
If you had read it all, you'd have seen that yes, with a new BVH walker and extra general RT perf, the card should be doing quite decently. N48 may be on par with a 4070 Ti, which isn't bad at all for a 240mm² card. I expect that N44 will probably run pathetically but still be able to do 1080p RT. Closer to a 4060 Ti.

$250 is the most optimistic price for N44. $270 seems plausible for replacing RX 7600.

That's your wishful thinking, not what AMD will go for. And this card is going from the perf of a 7600 to near a 7700 xt, there's no point in them selling it as a same price card.

There will be a 16GB N44 clamshell, pricing will be key for that. 8GB has a pretty strict price ceiling.

AMD really loves those lel.
I sort of hate them, it's either too little VRAM or way too much.

 

[branch_suggestion]

Okay, so what die sizes can AMD realistically achieve with RDNA4 from what we know and can assume? (Note this is a rough guesstimate, not as scientific as what Locuza did here since I don't have the actual die shots).

First we must look at the past and find allowances based on what we believe is true.
Assumed configs and data:

https://twitter.com/x/status/1776244864856051785

Starting with N48, targeting a ~240mm^2 die.
4SE/8SA/32WGP/64MB MALL/256-bit GDDR6 so identical to 2/3rds of N31.
Lets also assume 4MB of L2, 16x PCIe 5.0 and 3 Display PHYs, plus we can discount the fanout PHYs and add a shrink to the former N6 parts, on top of a small N5>N4 shrink for the GCD stuff and then assuming the iso node area of RDNA4 blocks vs RDNA3.
Size scaling cannot be done lower than the SE level as the WGP is heavily reworked, and too much is unknown. We do know GDS is gone, a small area saving.
Area leftover will be divided by data fabrics and the main uncore blocks.

So starting with the SE's, I'll assume they are slightly smaller overall, 26mm^2 x4 for 104mm^2.
The L2 slightly denser, 9.5mm^2.
PCIe5 is an area hog, 5.5mm^2.
Display PHYs the same density, 2.35mm^2.
256b PHYs a touch smaller, 38.5mm^2
64MB MALL does get a decent shrink based on Z4's L3 having a 29% shrink vs Z3.
But I won't go that far, much of that shrink was the data fabric. Scaling for the large selection will be 20%, 51.3mm^2.

All of that equals 211.15mm^2, leaving 29mm^2 for the fabrics, frontend, Display Engine, Media Engine, PCIE control, memory controllers and remaining misc.
29mm^2 seems not quite enough, 39mm^2 for 250mm^2 overall is comfortable, somewhere between 240-250mm^2 is realistic. Once again so much about RDNA4 remains a mystery, but this shows that the area seems plausible at the least, only small differences in certain blocks could skew this a lot.

N44 is N23v3, same rules apply.
Only 8x PCIe5 which also means smaller control logic.
Targeting ~130mm^2.

To save the wall of text, everything is halved except display PHYs, it ultimately equals 106.75mm^2, leaving 23.25mm^2 for the other stuff.
This works out nicely, frontend and the amount of fabric needed is smaller and some stuff from the uncore can be cut off as needed, this can also apply to N48 to a lesser degree e.g. only having 3 display support instead of 5 on N31, bit of an assumption. Once again the 130mm^2 is a rough number, it could be a bit higher or lower but within 10mm^2 or so. At 140mm^2 it is 33.25mm^2 spare, so 130-140mm^2 is realistic as of right now. Could also have the smaller vGPR but that is WGP stuff so no speculation there.

Ultimately I think the leak from All_The_Watts has no obvious red flags, same with the clock/power figures that they gave afterwards and the expected performance given beforehand knowing how much room for improvement there was from RDNA3. Thus I'm fairly confident this is the best data we have right now, it is the only serious leak so far, after all.

 

[Kepler_L2]

There will be a 16GB N44 clamshell, pricing will be key for that. 8GB has a pretty strict price ceiling.

No there won't be.

 

[branch_suggestion]

No there won't be.

7600XT is a thing that exists, albeit pretty late in the RDNA3 cycle.
So I agree no need for it at launch, but eventually I think so.
Basically once the old 12GB cards are sold out it will be time to bring it out, and it helps shift some lagging inventory for people who refuse to buy 8GB.
Gotta be enough of a price seperation from cutdown 12GB N48 though.

 

[Timorous]

Okay, so what die sizes can AMD realistically achieve with RDNA4 from what we know and can assume? (Note this is a rough guesstimate, not as scientific as what Locuza did here since I don't have the actual die shots).
View attachment 96703
First we must look at the past and find allowances based on what we believe is true.
Assumed configs and data:

https://twitter.com/x/status/1776244864856051785

Starting with N48, targeting a ~240mm^2 die.
4SE/8SA/32WGP/64MB MALL/256-bit GDDR6 so identical to 2/3rds of N31.
Lets also assume 4MB of L2, 16x PCIe 5.0 and 3 Display PHYs, plus we can discount the fanout PHYs and add a shrink to the former N6 parts, on top of a small N5>N4 shrink for the GCD stuff and then assuming the iso node area of RDNA4 blocks vs RDNA3.
Size scaling cannot be done lower than the SE level as the WGP is heavily reworked, and too much is unknown. We do know GDS is gone, a small area saving.
Area leftover will be divided by data fabrics and the main uncore blocks.

So starting with the SE's, I'll assume they are slightly smaller overall, 26mm^2 x4 for 104mm^2.
The L2 slightly denser, 9.5mm^2.
PCIe5 is an area hog, 5.5mm^2.
Display PHYs the same density, 2.35mm^2.
256b PHYs a touch smaller, 38.5mm^2
64MB MALL does get a decent shrink based on Z4's L3 having a 29% shrink vs Z3.
But I won't go that far, much of that shrink was the data fabric. Scaling for the large selection will be 20%, 51.3mm^2.

All of that equals 211.15mm^2, leaving 29mm^2 for the fabrics, frontend, Display Engine, Media Engine, PCIE control, memory controllers and remaining misc.
29mm^2 seems not quite enough, 39mm^2 for 250mm^2 overall is comfortable, somewhere between 240-250mm^2 is realistic. Once again so much about RDNA4 remains a mystery, but this shows that the area seems plausible at the least, only small differences in certain blocks could skew this a lot.

N44 is N23v3, same rules apply.
Only 8x PCIe5 which also means smaller control logic.
Targeting ~130mm^2.

To save the wall of text, everything is halved except display PHYs, it ultimately equals 106.75mm^2, leaving 23.25mm^2 for the other stuff.
This works out nicely, frontend and the amount of fabric needed is smaller and some stuff from the uncore can be cut off as needed, this can also apply to N48 to a lesser degree e.g. only having 3 display support instead of 5 on N31, bit of an assumption. Once again the 130mm^2 is a rough number, it could be a bit higher or lower but within 10mm^2 or so. At 140mm^2 it is 33.25mm^2 spare, so 130-140mm^2 is realistic as of right now. Could also have the smaller vGPR but that is WGP stuff so no speculation there.

Ultimately I think the leak from All_The_Watts has no obvious red flags, same with the clock/power figures that they gave afterwards and the expected performance given beforehand knowing how much room for improvement there was from RDNA3. Thus I'm fairly confident this is the best data we have right now, it is the only serious leak so far, after all.

The big issue is that N31 seems to be fat. It uses 58B transistors to achieve its spec but it is no different to 3x N33 which uses 39.9B transistors for the same spec.

Question is why does N31 seemingly have 18B transistors for nothing? That is another N33 and then some. In theory you could have an 8SE 512bit, 128CU design in 53.2B transistors and that would include 32x PCie4 and 4 display outs.

When you look at N31 Vs N33 it makes almost zero sense.

 

[branch_suggestion]

The big issue is that N31 seems to be fat. It uses 58B transistors to achieve its spec but it is no different to 3x N33 which uses 39.9B transistors for the same spec.

Question is why does N31 seemingly have 18B transistors for nothing? That is another N33 and then some. In theory you could have an 8SE 512bit, 128CU design in 53.2B transistors and that would include 32x PCie4 and 4 display outs.

When you look at N31 Vs N33 it makes almost zero sense.

Transistor count is just a fun metric at this point, total die area and what perf it has are the relevant metrics.
N31 has a bunch of things that N33 does not, basically a lot of little things that all add up, and bigger chips have certain things that scale disproportionate vs smaller chips.