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

 

[moinmoin]

Yea they're building those anyway.

Not much halo left on those "halo" parts currently though.

build something NV can't reach while adhering to strict TTM schedule, and if you can't then burn it and wait for the next cycle.

What is the average TTM if half of it takes a literal eternity (being canned and all)? At some point this approach has to be unfeasible. Though if everything stops at A0 anyway it's probably cheap enough? Dunno. Weird approach that only works (or not, haha...) with pure GPUs I guess.

 

[adroc_thurston]

Not much halo left on those "halo" parts currently though.

Of course, they no longer exist.

Though if everything stops at A0 anyway it's probably cheap enough

Those never reached the TO stage.

 

[H433x0n]

Intel stuff is stillborn.

If AMD is killing their top 2 RDNA4 dies, Battlemage becomes a real threat. Even now, Arc A770 & A750 prices have remained steady despite an inferior product when compared to the 6700XT and 6650XT. The A770 16Gb regularly has a higher sale price than the 6700 XT.

If anybody should be celebrating today, it’d be Intel graphics division.

 

[jamescox]

It is going to trickle down, it is just a question of when and in what form.

1/4 of Mi300 is a good basis for future client products, both CPU and GPU.

Intel is already biting a bullet, adding the cost of an interposer to all client products. So, Intel is going to have to increase pricing of MTL parts. Which gives AMD some room to maneuver.

But also some pressure, if Intel later adds cache to the active interposer die.

It could very well be that both Zen 6 and RDNA 5 will be Mi300 derivatives.

It would be great if they could do that, but it seems like the package may be too expensive. The Mi300 base die also likely doesn't contain anything other than HBM interfaces, GMI, and whatever they are using for the base die to base die links. I don't know if that is known yet? It doesn't seem like a general IO die that could be used across many other products unless it has a lot of stuff there, but unused in Mi300. They could possibly do a base die setup with SoIC for a new IO die, but it is probably a lot cheaper and more flexible to leave it as a separate chip. There is no need for the level of bandwidth provided by SoIC chip stacking between cpu chiplets and the IO die. If they want to connect an accelerator, then that can be done via regular GMI links (50 GB/s in the lates iteration?) or maybe infinity fabric fan out links for really high bandwidth things. AMD wants to put video in everything, so I am wondering if next gen IO die will actually be essentially an APU with GMI links.

They can make products a lot cheaper with alternate tech. They don't really need HBM for many applications. Using MCD with GDDR seems to be a good replacement. The infinity fan out for MCD <-> GCD connections seems to have been explicitly developed to avoid using interposers or other stacking tech. If there are no issues with the infinity fan out connections, then it seems like that would be used instead. I believe the infinity fabric fan out is actually a scaled up version of something developed for mobile. I am hoping that they will make an APU utilizing it for power and bandwidth savings. An APU with on package memory and MCDs would be spectacular. They may also be able to use it to connect different acclerators, like something for video editing, more cpu cores, or even an extra gpu chiplet.

 

[adroc_thurston]

Battlemage becomes a real threat.

Not competitive in PPA even with RDNA3.

The Mi300 base die also likely doesn't contain anything other than HBM interfaces, GMI, and whatever they are using for the base die to base die links. I don't know if that is known yet?

It has MALL, CPs, HWS and ACEs, HBM PHYs and d2d USRs, plus a whole bunch of 32G PHYs for IFIS/PCIe and CXL2.

 

[darkswordsman17]

If true, this tells me that AMD is struggling to get gaming performance out of MCM, and wasting resources on large GPUs is stupid, so they're making shrewd, but I think smart/sensible decision. They literally would have to offer double the performance of NVidia at this point to make just inroads to market share. They will never ever be able to overcome the mindshare poison that Nvidia has been spending decades building. Even though basically everyone (including Nvidia fans) hate Nvidia at this point and it still doesn't matter.

People acting like you need a halo product are just nonsense, and that won't matter unless AMD doubles or triples NVidia's performance (and that's simply not feasible), and it'd have to happen for probably 5 years straight for even that to change things. Plus gaming stuff that gets the most sales and hype these days are stuff like the Steamdeck, Switch, and the like (the Steamdeck competitor devices that have come out recently). Heck I think it'd be more sensical (and likely far more lucrative) for AMD to partner with Samsung to make an Android based Switch ripoff.

Plus you guys can have fun paying $3500 for Nvidia's DLSS^3 (indicating it just makes up the path traces, the frames, and even the assets! - look ma, just check a box you don't even have to program a game or create assets for it, think of all the money you can make as a game developer then! honestly with the hilarity of AI generation of images that sounds more interesting than Call of Duty 57/Halo 15/etc) as enthusiast market products become prosumer/enterprise hardware.

Personally, I think AMD should ditch the current AIB dGPU market entirely. Especially since the cards are large enough to warrant it now. Just turn them into full blown dedicated external gaming boxes by integrating CPU cores either on die or MCM with the I/O die basically managing shared cache and memory system as well as a fast NAND controller setup (imagine what 10TB/s SSD bandwidth would enable as far as asset swapping and the like). Basically create a unified memory system hiearchy (large cache to sizable but not outrageous GDDR pool or HBM if it ever becomes viable again, to fast and large enough to load very large AAA game - so like 128-256GB of very fast NAND, then have a normal SSD slot for long term storage). They could possibly make gaming focused CPU designs, but not sure that's necessary with the large shared cache/memory setup. They could integrate dedicated hardware bits somewhere like they do for consoles.

For normal client focus on laptops and small boxes, which will drastically reduce costs for OEMs who can then integrate the small boxes into all sorts of stuff. Could even push that to other markets (streaming boxes with more capability, integrate as the control boxes for TVs, etc).

Which if they wanted to do a half-step to get there, they could. For an idea of what that could look like check out what Khadas is doing:

Khadas

Khadas

mind.khadas.com

They could also do other steps where they work their I/O to support and then require premium or gaming laptops to have like 2-4 USB 4.0 Gen 2, with each being able do 120Gb/s one way or 80/80 split which would make eGPUs much more viable (especially if you double or triple that; 120Gb/s). This would be a relatively quick solution instead of waiting for new external optical data standard for consumer, although that would be preferred. It lets the laptop cooling handle just the CPU, and 2-4 Gen 2 would offer 30-60GB/s of bandwidth. Plus if they add just a current PCIe Gen 4 level SSD slot to the eGPU immediately, letting it put game assets there - maybe with some pre-config for like decompression or something, utilizing the new DirectStorage, should further help reduce the bandwidth need between the eGPU and the CPU part as well.

 

[adroc_thurston]

If true, this tells me that AMD is struggling to get gaming performance out of MCM

Performance is the easy part.

so they're making shrewd, but I think smart/sensible decision

The very next thing (RDNA5) has the exact same insane halo part planned.

Personally, I think AMD should ditch the current AIB dGPU market entirely.

Never.

 

[H433x0n]

Not competitive in PPA even with RDNA3.

The A770 is a first generation card, it’s nearly irrelevant.

Also, consumers don’t care about PPA. If a cards PPA contributed to its success in the market then the 4060 and 4060 Ti would be flying off the shelves. A lot of Ampere cards had atrocious PPA and it still made a ton of money.

Clearly it’s already a threat since they’re able to maintain higher prices on their cards compared to the RDNA2 equivalent.

 

[adroc_thurston]

The A770 is a first generation card, it’s nearly irrelevant.

Very relevant, PPA is all in client GPUs.

Also, consumers don’t care about PPA

Patrick P. Gelsinger really-really does, since AXG is a sizeable red hole in Intel's already not particularly cash-rich biz.

A lot of Ampere cards had atrocious PPA

They're pretty good actually.

Clearly it’s already a threat since they’re able to maintain higher prices on their cards compared to the RDNA2 equivalent.

It's not, Intel is considered a joke by both AMD and NV as far as GPUs go.
I mean for gods sake, Gen's head honcho (Blythe) is at AMD.

 

[CouncilorIrissa]

Isn’t Arc Alchemist (DG2) technically a second-gen product, with DG1 being the first? It’s just that DG1 sucked so bad that it was a no-go in terms of making an actual product. So Intel kind of buried it to make sure no one remembers it, much like they did with Cannon Lake?

My point is that Intel can’t afford to be this behind in PPA. If they’re not reasonably close with Celestial, it’s not out of the question for them to pull the plug on their dGPU program (again)

 

[igor_kavinski]

I mean for gods sake, Gen's head honcho (Blythe) is at AMD.

Ah. So David Blythe is the real genius behind ARC and XeSS. Must have been irked by Koduri's over exposure and decided to jump ship to escape the boisterous monkey trying to steal the limelight.

 

[adroc_thurston]

So David Blythe is the real genius behind ARC and XeSS

Not sure if genius but he's definitely a decent guy.

Must have been irked by Koduri's over exposure and decided to jump ship to escape the boisterous monkey trying to steal the limelight.

Raja's also gone.
But Intel dGP prospects are grim; the less we talk about them, the merrier.

 

[H433x0n]

But Intel dGP prospects are grim; the less we talk about them, the merrier.

As of today their dGPU prospects appear less grim than Radeon. Battlemage will at least attempt to compete at the midrange.

 

[adroc_thurston]

As of today their dGPU prospects appear less grim than Radeon

Have you seen RDNA4?

Battlemage will at least attempt to compete at the midrange.

It will get pancaked by the tiny RDNA4 peanut much the same way ACM-G10 gets pancaked by N33.

 

[H433x0n]

Have you seen RDNA4?


It will get pancaked by the tiny RDNA4 peanut much the same way ACM-G10 gets pancaked by N33.

What do you consider “pancaked”? The performance is the same as 6650XT which performs the same as A770.

What is it with everybody that favors AMD wanting Arc to fail so bad?

 

[adroc_thurston]

What do you consider “pancaked”?

I adhere to this wonderful concept of "PPA triangle" where N33 is still pretty good despite being a relative miss on power.

The performance is the same as 6650XT which performs the same as A770.

Yea with a 204mm^2 die.
ACM-G10 is twice that.

What is it with everybody that favors AMD wanting Arc to fail so bad?

Not that I want it to fail, since whenever AMD claims their due it'll be rough times just like in CPU lands.
I just think they have little chances from what I've seen of their current products, their future roadmap and given the recent layoffs (that hit the thing formerly known as AXG pretty hard).

 

[PJVol]

Just wow ))

 

[TESKATLIPOKA]

What is going to happen to @TESKATLIPOKA what tables and projections will he make then?

Don't worry, I will continue doing It for Nvidia GPUs.

I kinda don't understand all this gloom and doom about AMD not releasing a highend chip in next gen.
So what If they don't release a high-end? It's not the end of the world and not many of us are willing to pay so much just to play a few hours per week.
The question is what will be the next midrange pricewise.
In my opinion, the $599 RTX 4070 is the current midrange card.
Next gen will go higher in price most likely. The question is by how much.

Although @adroc_thurston says that RDNA4 will be limited to ~$399, I don't believe It. Certainly not with 2 RDNA4 chips to come and also Strix Halo with up to 40CU IGP.
At least one of them should aim higher both in performance and price.

And here are my projections or what I would kinda expect for a next gen midrange. Hope you enjoy.

	Shader Engines	WGP (CU)	Shaders (FP32)	TMU	ROPS	Clockspeed	MCDs	Infinity cache	Memory width	Memory speed	Bandwidth	Vram	Price
N43 cutdown	2	16 (32)	2048 (4096)	128	64	3500 MHz	2	48 MB	128-bit	24 gbps	384 GB/s	8 GB	$299
N43	2	20 (40)	2560 (5120)	160	80	3500 MHz	3	72 MB	192-bit	20 gbps	480 GB/s	12 GB	$399
N42 cutdown	4	32 (64)	4096 (8192)	256	128	3500 MHz	4	96 MB	256-bit	24 gbps	768 GB/s	16 GB	$549
N42	4	40 (80)	5120 (10240)	320	160	3500 MHz	5	120 MB	320-bit	24 gbps	960 GB/s	20 GB	$699

Some info:
New MCD with 24MB IC and 64-bit gddr6.
20CUs per SE for both chip and 16CUs for the cutdown versions.

Performance estimate:
N43 cutdown -> faster than RX 6700XT at 1080-1440p, at 4K comparable performance unless Vram is insufficient.
N43 -> faster than RX 6800 at 1080-1440p, at 4K comparable performance unless Vram is insufficient.
N42 cutdown -> faster than RX 7900XT at every resolution.
N42 -> faster than RX 7900XTX at every resolution.

Nothing spectacular, but with those prices I think It would be a decent offering with much better perf/$. What I hate is the Vram regression, but I couldn't get rid of It unless they released 12gbit 24gbit chips.

 

[DisEnchantment]

Although @adroc_thurston says that RDNA4 will be limited to ~$399, I don't believe It.

Good joke, that is around 6700XT, six months ago that is not even the case.
I got my 6600XT at 420 bucks at launch, if 399 is their launch price then ~32CU cards is what they are targeting, give me a break.

Too much personification of a company which is profit driven, as if they won't grab the opportunity if it is there because their feelings were burnt by gamers choosing NV.
Fall behind over multiple years and they will discover one day that their architecture wont be suitable to run desktop games anymore without an emulation layer.

Why launch new RDNA3 based enthusiast card now then if at all right now late in the RDNA3 cycle.

If chiplets are in shambles they would do a monolithic die. Its not like GPUs need validation time like CPUs.

 

[jpiniero]

Why launch new RDNA3 based enthusiast card now then if at all right now late in the RDNA3 cycle.

Could easily be because the top RDNA4 is slower than the 7800 XT.

 

[TESKATLIPOKA]

Could easily be because the top RDNA4 is slower than the 7800 XT.

Slower than the current midrange? Highly doubt It, especially If they manage to fix the relatively low clockspeed.
With 3300-3500MHz clockspeed even a 48CU, 3072Shader,192TMUs,96ROPs chip + 64MB IC 192-bit 24gbps should be around 7800XT level of performance.

 

[Tuna-Fish]

I can understand the RDNA1 generation as AMD was totally broke for most of the time it was under development, but ...

Frankly, I don't get why they did it. Are the economics really just not there? Or rather is it a sign that chiplet-based GPUs aren't really working out yet?

If true, this late in the game I would assume that they attempted split compute GCD and it didn't work (either too bad yields or performance issues).

At this point, if your product stack depends on that and you can't do it, then you just cannot ship those products, and you will have a generation truncated in the high end.

 

[jpiniero]

If true, this late in the game I would assume that they attempted split compute GCD and it didn't work (either too bad yields or performance issues).

That's what it sounds like. And so you are going to get the one die version only...

 

[jpiniero]

I will add that I wouldn't be surprised they pulled people off to work on CDNA cuz AI Hype.

 

[A///]

I will add that I wouldn't be surprised they pulled people off to work on CDNA cuz AI Hype.

that and rdna3 was the most lukewarm serving of kickass in years from amd. unimpressive. it makes more sense to grab a 6900xt.