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

 

[adroc_thurston]

Chasing after NVidia on the high end hasn't worked out well for them

Which is why they're making Navi50. Aight.

With smaller dies AMD can produce more product and actually try to grow market (and mind) share for a change

The needle won't move till the beheading.

 

[Mopetar]

It's not as though I don't expect them to make a big die ever again, just that I don't think they need to in order to succeed. RDNA2 showed they were capable of competing at the high end, it's just that NVidia will do whatever it takes to be just that little bit better and that will keep a lot of people from buying AMD at the high end where people just want the best card and don't care as much about the price or even if it's a 500W toaster that makes pretty graphics.

The 7800XT was where AMD actually did well in the market. If Nvidia is releasing an 8 GB 5060 Ti or similar such products, AMD has a real chance to win customers with Navi 48. Having a halo product is good, but AMD loyalists aren't going to jump ship just because they don't. They stuck around through worse times.

 

[adroc_thurston]

just that I don't think they need to in order to succeed

Yes they do!
It's kill or be killed.
You need to make the shills scream and kids orgasm, and only the biggest stick could do that.

 

[Mopetar]

Yes they do!
It's kill or be killed.
You need to make the shills scream and kids orgasm, and only the biggest stick could do that.

Unless they can beat NVidia's best I don't think their high-end is as important. They also have to win by enough that NVidia can't just juice some card to the gills to scrape out a victory. They were as close to equal to NVidia as they have ever been in recent memory with RDNA2 and having that ultimately did nothing for them. There are arguments that RDNA3 having issues set them back and had that lived up to expectations we'd be seeing much different results, but that's all speculation.

 

[adroc_thurston]

Unless they can beat NVidia's best I don't think their high-end is as important.

Yes!
They know it!

They also have to win by enough that NVidia can't just juice some card to the gills to scrape out a victory

Oh they will make 600W ovens themselves.
It's not an issue some 150-200 WGPs can't solve.

having that ultimately did nothing for them

Well yeah RDNA2 was bad for mining.

 

[Heartbreaker]

Unless they can beat NVidia's best I don't think their high-end is as important. They also have to win by enough that NVidia can't just juice some card to the gills to scrape out a victory. They were as close to equal to NVidia as they have ever been in recent memory with RDNA2 and having that ultimately did nothing for them. There are arguments that RDNA3 having issues set them back and had that lived up to expectations we'd be seeing much different results, but that's all speculation.

It's not just performance.

Another reason NVidia has mindshare, is because they lead on introducing features. RT, DLSS, G-sync, Fake Frames...

You often end up waiting years for AMD to catch up, and when AMD offer the features it's often a lesser version.

 

[adroc_thurston]

Another reason NVidia has mindshare, is because they lead on introducing features. RT, DLSS, G-sync, Fake Frames...

Shill points.
Shills are easily bribed.

 

[GodisanAtheist]

@Mahboi is a little mad men but on point with his read on AMD culture.

I felt it back when AMD got curb stomped with the transition to DX10, the 2900xt getting it's stuff pushed in by the 8800GTX.

AMD went lean with the 3870, and then lean again with the 4870 while massive NV dies saddled with a bunch of compute stuff in its infancy ate power and die area.

AMD had a chance to go big with 5870, but they went lean again, with a 330mm2 top end die. NV's 480 was 550mm2 and only 15% faster.

Imagine if AMD had grown some teeth and scaled their arch to 550mm2 like NV? It would have been the ass kicking of the century. Maybe CUDA would have just died, and the whole trajectory of AMD would be different. There wasnt even a real feature difference at the time.

But AMD played it safe, they went with the "small die strategy" and effectively lost the window to ever really catch up to Nvidia ever again.

 

[linkgoron]

While funny - it didn't blow up in their face because of the chiplets. The dud wouldn't clock right even when monolithic (Phoenix) and on N4.

I've seen many posts here that say how supposedly great RDNA4 is, and how "AMD is back".

If the issue was RDNA3 and not chiplets, why did AMD cancel chiplet-based Navi 41 and Navi 42?

Where's the big-die monolithic Navi 4X? It doesn't exist because they bet too hard on chiplets and it didn't work out (for now).

 

[adroc_thurston]

why did AMD cancel chiplet-based Navi 41 and Navi 42?

You forgot Navi40.

Where's the big-die monolithic Navi 4X

Never existed.

It doesn't exist because they bet too hard on chiplets and it didn't work out (for now).

But it did work out.
They just focused on raw execution instead.

 

[maddie]

I've seen many posts here that say how supposedly great RDNA4 is, and how "AMD is back".

If the issue was RDNA3 and not chiplets, why did AMD cancel chiplet-based Navi 41 and Navi 42?

Where's the big-die monolithic Navi 4X? It doesn't exist because they bet too hard on chiplets and it didn't work out (for now).

1st gen chiplets (good), then 2nd gen (tricky). My understanding is the compute with chiplets is the issue.

 

[QuickyDuck]

I've seen many posts here that say how supposedly great RDNA4 is, and how "AMD is back".

If the issue was RDNA3 and not chiplets, why did AMD cancel chiplet-based Navi 41 and Navi 42?

Where's the big-die monolithic Navi 4X? It doesn't exist because they bet too hard on chiplets and it didn't work out (for now).

IMO chiplets did work out then the AI boom came. They just allocate their precious capacity to AI chips instead.

 

[gdansk]

I've seen many posts here that say how supposedly great RDNA4 is, and how "AMD is back".

If the issue was RDNA3 and not chiplets, why did AMD cancel chiplet-based Navi 41 and Navi 42?

Where's the big-die monolithic Navi 4X? It doesn't exist because they bet too hard on chiplets and it didn't work out (for now).

I haven't said anything about RDNA4.
But even the cutdown Navi 31 XT is 25% faster than full Navi 21 when both are at 300W. But 780M is only 10% faster than the 680M typically and it had a large process advantage too.

RDNA3 is relatively *worse* in its most integrated from.

 

[Mahboi]

I don't mind this and think it could be good for AMD in the long run. Navi 32 proved popular for AMD because they had a mass market product at a compelling price. Navi 48 looks like it could do more of that and help get some consumers to buy their first AMD card.

Chasing after NVidia on the high end hasn't worked out well for them and when they are constrained on wafers and can make more money producing more Zen dies for the server market, they don't have a lot of incentive to increase their GPU production.

With smaller dies AMD can produce more product and actually try to grow market (and mind) share for a change, which will help them if and when they can better execute in the future.

There's a lot to say there, but one thought occurred to me last night:
N48's final naming and price will tell a lot about the goals. Specifically, if they're dumb enough to redo RDNA 3 tier naming and call the damn thing 8800 xt, or god forbid 8800 XTX, I'll be convinced that RTG is led by brainworm infested donkeys. 8700 XT, nothing else. Compete with the future 5060 Ti, and compete with a clean naming scheme that says "not a big card, but a good card". Stop the damn shrinkflation with misleading naming to boot.

As for the price, I'm personally expecting anywhere between $550 and $450.
$550 is the highest price at which they can sell it for 18 months with a good market value. The 5060 Ti will probably be around $400 or $450 but definitely offer worse value, the 5070 should be stronger.
Below $450 is where they'd get into wasted margins because the card is effectively a drop-in replacement for the 7800 xt, same VRAM buffer/bus/1440p excellence/weak but decent 4K.

Anywhere between the two is the spectrum on "how much AMD wants margins" vs "how much AMD wants market share". Whatever they'll pick will tell about the goals with RDNA 4. I'm actually somewhat reassured with stuff like picking 18 Gbps GDDR6.

 

[jpiniero]

I am not expecting much from Blackwell's lower end parts... and it isn't coming out until next year regardless. So N48 is mainly going to compete against what's currently available.

 

[TESKATLIPOKA]

I am not expecting much from Blackwell's lower end parts... and it isn't coming out until next year regardless. So N48 is mainly going to compete against what's currently available.

I think low-end Blackwell could be also good depending on specs and price. I was thinking It could use only 96-bit, but I left It only for the cutdown version and only 24gbit modules will be used.

Model	Chip	Die size	SM(CU):Shader: TMU:ROP	Clockspeed - CBP 2077	TBP	L2 cache Infinity cache	Memory width	Memory module	Bandwidth	Vram	TPU perf. 1440p
RTX 4060	AD107 full	146 mm2 4N	24:3072:96:32	2618 MHz	115 W	24 MB	128-bit	17 gbps	272 GB/s	8 GB	100 %
RTX 4060Ti	AD106 cut	190 mm2 4N	34:4352:136:48	2756 Mhz	165 W	32 MB	128-bit	18 gbps	288 GB/s	8/16 GB	129 %
RTX 5060	GB107 cut	<190 mm2	32:4096:128:48 ?	2900 Mhz ?	150 W ?	32 MB ?	96-bit ?	28 gbps	336 GB/s	9 GB	~140 % ?
RTX 5060Ti	GB107 full	<190 mm2	40:5120:160:48 ?	3000 Mhz ?	185 W ?	32 MB ?	128-bit ?	28 gbps	448 GB/s	12 GB	~175 % ?
RX 7600XT	N33 full	204 mm2 N6	32:2048:128:64	2690 Mhz	190 W	32 MB	128-bit	18 gbps	288 GB/s	16 GB	106 %
RX 7800XT	N32 full	200+148 mm2 N5+N6	60:3840:240:96	2332 MHz	263 W	64 MB	256-bit	19.5 gbps	624 GB/s	16 GB	176 %
RX 8600XT?	N44 full	~130 mm2	32:2048:128:64 ?	3200 ?	?	32 MB ?	128-bit	18 gbps	288 GB/s	8/16 GB ?	?
RX 8800XT?	N48 full	~240 mm2	64:4096:256:96 ?	3200 ?	?	64 MB ?	256-bit	21.65 gbps	693 GB/s	16 GB	?

I didn't try to calculate any performance for RDNA4, because of BW. It's at best only 11% faster and even that only for N48.

 

[Aapje]

@TESKATLIPOKA

Your VRAM quantities don't match up with your bus sizes. A 96 bit bus means 6 or 12 GB. A 128 bit bus means 8/16. So your 5060 and 5060 Ti specs make no sense.

 

[Aapje]

N48's final naming and price will tell a lot about the goals. Specifically, if they're dumb enough to redo RDNA 3 tier naming and call the damn thing 8800 xt, or god forbid 8800 XTX, I'll be convinced that RTG is led by brainworm infested donkeys. 8700 XT, nothing else. Compete with the future 5060 Ti, and compete with a clean naming scheme that says "not a big card, but a good card". Stop the damn shrinkflation with misleading naming to boot.

As for the price, I'm personally expecting anywhere between $550 and $450.

I'm pretty sure that they will call it the 8800 XT, which wouldn't be a problem, since the 7800 XT price is exactly within your price range, at $499. So if they just keep the price and give 7900 XT+ performance, that would be a solid increase in price/performance, as the 7900 XT sells for $699 now.

 

[TESKATLIPOKA]

@TESKATLIPOKA

Your VRAM quantities don't match up with your bus sizes. A 96 bit bus means 6 or 12 GB. A 128 bit bus means 8/16. So your 5060 and 5060 Ti specs make no sense.

There is no mistake, because I mentioned in my post that I used only 24gbit modules.
24gbit = 3GB module
96-bit -> 96/32 =3 modules -> 3*3GB= 9GB
128-bit -> 128/32 = 4 modules -> 4*3GB= 12GB

or If there won't be enough 24gbit modules, then 16gbit modules in clamshell for 96-bit = 12GB and 128-bit = 16GB, but for that you need 2x more modules, so higher BOM.
I don't think only 8GB GPUs will be very sellable unless low price.

 

[Aapje]

@TESKATLIPOKA

The rumors say that those aren't yet available.

 

[Tuna-Fish]

They are not available for when the first cards launch. I would expect Blackwell release to mirror Ada in that the initial launch just has the high-end chips.

 

[moinmoin]

Unless they can beat NVidia's best I don't think their high-end is as important.

And that's exactly why AMD cancels its high-end parts that merely chase Nvidia's best instead clearly toppling them. So skip one gen and retry in the next one it is.

N48's final naming and price will tell a lot about the goals.

Naw. It will tell a lot about AMD's marketing dept though. I'm not expecting any positive surprises.

 

[Mahboi]

Naw. It will tell a lot about AMD'S marketing dept though. I'm not expecting any positive surprises.

Sorry my nerds but you really can't say "the marketing dept isn't part of AMD". It is. And their blunders are costing the corp.
I still remember Sherkelman's nipples through his RDNA 3 football shirt.
Don't Jebait me this way, bald man.

 

[Ajay]

But it did work out.
They just focused on raw execution instead

Well, missing the market window is a 'didn't work out' in my book. Raw execution is knowing when to move on instead of diddling. Hard calls. Good management team.

 

[moinmoin]

Sorry my nerds but you really can't say "the marketing dept isn't part of AMD". It is. And their blunders are costing the corp.
I still remember Sherkelman's nipples through his RDNA 3 football shirt.
Don't Jebait me this way, bald man.

Let's just say AMD would be a completely hopeless case if it were run by the folks in its marketing dept.