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]

It's taking Nvidia's naming scheme, and starting with a 9 makes this look like a one time thing

Naming scheme change is dumb and pointless but 9 is expected since 800/8000 is RDNA3.5 parts (strix/halo/kraken).

 

[Panino Manino]

They're never building another big config ever again.

They need to find a "GPU Jim Keller" to go tho the graphics division and slap them hard in the butt.
I understand why they would choose to bet on the "mainstream" but this will never give them money to do anything else, only a slow sad death.

Is it Lisa who is making this decision?

 

[gdansk]

Is it Lisa who is making this decision?

Gamers already made the decision. Sony and UDNA will keep it around but Radeon is dead.

 

[jpiniero]

They need to find a "GPU Jim Keller" to go tho the graphics division and slap them hard in the butt.
I understand why they would choose to bet on the "mainstream" but this will never give them money to do anything else, only a slow sad death.

Is it Lisa who is making this decision?

It's probally like Polaris... they see where the node pricing is going and unless they can get "nVidia - $100" ASPs and get people to buy it at those prices... it's not worth the resources with AI hype at full throttle.

 

[tajoh111]

Doable with N3P in 2026?

If AMD can do cheaply 3D cache for CPUs, why isn't it done for GPUs?

128 MB worked wonders in RDNA2 and after that it's gone downhill, AMD go BIG or go home - 256 MB unified 3D cache in UDNA1 is needed, their own stats shown that 128 MB is min for 4K and clearly gains will be very healthy for 256 MB.

The way things go in 2026 we'll finally have good 8k monitors and NVidia will yet again be ahead leaving AMD only scraps.

The bigger you go, the more easy it is to screw up. Vega, radeon VII and Fiji were painful lessons for AMD. Vega particularly is where AMD tried to go big but lost so much in the process. Particularly mindshare. AMD lost a ton of good will towards enthusiasts because it was so late, AMD guerilla marketing team kept stalling with fake leaks which made it appear the actual product would be worth the wait. We all remember how that turned out. RDNA 2 was the only time they got this right in recent years but that's mostly as a result of Nvidia going on a inferior samsung node.

Even RDNA3 while not being that big had it own engineering shortfalls. If Navi 4C was ever made, I think there is a high likelihood with AMD track record of being late, not living up to expectations and even if it beat Nvidia, it would be by barely anything while losing in ray tracing and ultimately not doing much good for AMD financials.

AMD graphic division with its current resources does not want to make such a bet.

 

[adroc_thurston]

They need to find a "GPU Jim Keller" to go tho the graphics division and slap them hard in the butt.

No.

I understand why they would choose to bet on the "mainstream" but this will never give them money to do anything else, only a slow sad death

They're choosing to disengage from the market.
Like they're gonna keep a token presence, but they're not gonna try to win it either, and client dGFX is explicitly the market where all hinges on a single massive win.

Is it Lisa who is making this decision

Yeah.

Sony and UDNA will keep it around but Radeon is dead.

Not such thing as UDNA and APUs very much live.

The bigger you go, the more easy it is to screw up. Vega, radeon VII and Fiji were painful lessons for AMD. Vega particularly is where AMD tried to go big but lost so much in the process. Particularly mindshare. AMD lost a ton of good will towards enthusiasts because it was so late, AMD guerilla marketing team kept stalling with fake leaks which made it appear the actual product would be worth the wait. We all remember how that turned out. RDNA 2 was the only time they got this right in recent years but that's mostly as a result of Nvidia going on a inferior samsung node.

Even RDNA3 while not being that big had it own engineering shortfalls. If Navi 4C was ever made, I think there is a high likelihood with AMD track record of being late, not living up to expectations and even if it beat Nvidia, it would be by barely anything while losing in ray tracing and ultimately not doing much good for AMD financials.

AMD graphic division with its current resources does not want to make such a bet.

That's a lot of babble about things you do not understand. Don't.

 

[RNR_Forte]

N3 please!

 

[jpiniero]

N3 please!

2 years from now.

 

[adroc_thurston]

N3 please!

Well N4c was that. But you're not getting it.

2 years from now.

Yes it's always 2 more years.

 

[linkgoron]

The bigger you go, the more easy it is to screw up. Vega, radeon VII and Fiji were painful lessons for AMD. Vega particularly is where AMD tried to go big but lost so much in the process. Particularly mindshare. AMD lost a ton of good will towards enthusiasts because it was so late, AMD guerilla marketing team kept stalling with fake leaks which made it appear the actual product would be worth the wait. We all remember how that turned out. RDNA 2 was the only time they got this right in recent years but that's mostly as a result of Nvidia going on a inferior samsung node.

I don't think Radeon 7 was really significant. I recall people thinking that it was a pleasant surprise, even though price-wise it was meh. It had a few nice things going for it (huge memory, compute etc), and for AMD it was mostly a pipe-cleaner and was relatively quickly replaced with RDNA1. Fiji and of course Vega were awful. Fiji's (overclocking monster) 4GB HBM killed it, and Vega (poor volta) was huge and still inferior to 1080 (although I thought at the time of the Fiji release that Fury Nano as a concept had potential if they had actually tried something similar in the future). Really, the amount of overhyping, overpromising and underdelivering in the Raja Koduri "era" was awful.

 

[yuri69]

They're never building another big config ever again.

Management decisions. ATi nearly imploded by the R600 fiasco. So management went for the "sweet spot" strategy aka never investing such large money in a very risky large die again. Yea, except they did Cayman like 4 years later.

 

[adroc_thurston]

Yea, except they did Cayman like 4 years later.

Cayman was a wee baby hit by the 32nm cancellation.
Took ATi 9 whole years to ship a new chungus (Fiji).

 

[SolidQ]

"sweet spot" strategy

if their new strategy cards 6/7/8 series up 1k$ i'm fine. We nothing lose, crazy people will always buy overpiced cards from NV

 

[tajoh111]

No.

They're choosing to disengage from the market.
Like they're gonna keep a token presence, but they're not gonna try to win it either, and client dGFX is explicitly the market where all hinges on a single massive win.

Yeah.

Not such thing as UDNA and APUs very much live.

That's a lot of babble about things you do not understand. Don't.

Then explain rather than give a passive aggressive response particularly towards me.

At this point, I feel your one of those AMD guerilla marketers(one who claims to be in the know and disappear when your predictions don't come true) in the past that I have outed in the past and have an axe to grind.

 

[RNR_Forte]

I mostly buy AMD cards but i have to say some crammed words.
Is based on the track record part of the insuccess of AMD in the gaming market. G80 was superated by RV770 but GT200 sold better because the Geforce track record was prime.
Cypress did not outsold FermiV1 due to bigger mindshare of Nvidia and FermiV2 outsold the last terascale VLIW4 based on Geforce Track record.
GCN1.0 sold well but the blaster Hawaii was made in a Hurry and performed hot and with too late Custom versions; Anyway the R9 290 non-x was a dream of price/performance. But Kepler had the performance-per-watt leading.
Maxwell came when AMD before LisaSU and was kinda out of VGAs because Zen. The GTX 970 was a blast as the R9 290 was when launched. Fiji was a bit late but hold the line nicely. Anyway GCN limitations did not allow Fury X to outperform the GM200 Halo.
Polaris bring solid performance but it had lost its power efficiency due to GloFo WSA. Polaris was made in a "Only for Zen" 14LPP. Vega1 was supposed to outperform GTX 1080 but like(And unlike because RX580 is still a relevant card) Polaris and vega was supposed to be made at 16FF+ at TSMC.
Turing was left AMD counterattack with a filler(And boring) Vega 7.
Then Lisa Su got the stage but launched just two midrange cards. Very well, but not what people expected. RDNA 2 was a marvel but undersold mysteriously. But I still think that amd could give more performance on DX12U aspects. The people needed to see more clear on how efficient RDNA2 was versus Ampere too.
And then came AD versus the chiplet(an option to make great cards with an affordable price and without assembling the card with second class components) RDNA 3. The power draw of RDNA3 was inferior compared to ADA(see the gpupwr of RDNA2 versus AMP and then see the same comparison for RDNA3 vs ADA) but the cards still sold well.
My hopes are made at how much RDNA4 took to be crafted. With Ryzen AI Max being a minor upgrade archicterally great but it needs more focus on the DX12U performance. The CPU sales are very well now but all about RDNA4 will be good.


PS: I'm all wishful of the AMD success in the graphics card market too.

 

[SolidQ]

RDNA 2 was a marvel but undersold mysteriously

Because people was brainwashed by DLSS/RT in 2020, especially RT, back then there was very minor number of games with RT.
I knew then, 3080 is crappy card with 10GB, because new consoles came with 16gb

The people needed to see more clear on how efficient RDNA2 was versus Ampere too

it's was, but people need only green sticker. RTX 3050 vs RX 6600 showed it

I'm all wishful of the AMD success in the graphics card market too.

they was taken market share with RDNA3 and without Halo card, but super series/RT/Dlss killed it. If AMD FSR4 and RT will be fine and prcie also, they can get some market share.

 

[adroc_thurston]

Then explain rather than give a passive aggressive response particularly towards me.

a) learn to be more concise
b) N4c is dead because the end value of halo client dgfx swimlane is questionable at best given the current market conditions. It's just an expensive program for something that might not drive even incremental volumes for Radeon in the end.

At this point, I feel your one of those AMD guerilla marketers(one who claims to be in the know and disappear when your predictions don't come true) in the past that I have outed in the past and have an axe to grind.

meds.

RTX 3050 vs RX 6600 showed it

yeah that hurt.

 

[Win2012R2]

It's just an expensive program for something that might not drive even incremental volumes for Radeon in the end.

MI300 seems to be generating decent money, and they can't or won't take some of it to push graphics division? Sooner or much sooner AI bubble will burst and then what AMD is left with - 10% market share for GPUs?

RDNA 2 did good - just bloody scale it, if 500 bucks CPU can get 96 MB of cache including 3D stuff then so can GPU. Instead of pushing this strategic direction AMD spent crazy money on acquisitions like Xilinx or that serverbox maker recently.

RDNA 2 was a marvel but undersold mysteriously

Eh? It sold out, mind you there was crazy shortages at the time, but it did sell out.

 

[adroc_thurston]

and they can't or won't take some of it to push graphics division

They'll use the money to push GPGPU etc further. Client dGFX guys have to justify their very existence after RDNA3 anyway.

RDNA 2 did good

It very much did not. A total flop.

if 500 bucks CPU can get 96 MB of cache including 3D stuff then so can GPU.

That CPU is server table scraps. Client gfx you gotta build with a purpose.

 

[Win2012R2]

That CPU is server table scraps

Ok, explain to me this - we get 3D consumer versions very quickly after (botched) launch of Zen 5, and at the same time AMD officially says that they don't plan to have Turin 3D because of "reasons". Why are they making 3D versions then so quickly?

Xbox is dead, there won't be meaningful semi-custom money from them and Sony alone might not carry it, they need consumer dGPUs to be successful - and work well with that ROCm malarkley, otherwise NVIDIA will continue to wipe the floor with them for the next decade, those tiny NPUs won't hack it.

Client gfx you gotta build with a purpose.

Well hire people to do it, part fund with AI bubble money since it's kind of in the ball park area anyway.

 

[adroc_thurston]

Ok, explain to me this - we get 3D consumer versions very quickly after (botched) launch of Zen 5, and at the same time AMD officially says that they don't plan to have Turin 3D because of "reasons". Why are they making 3D versions then so quickly

Because Turin-X was a thing but MS picked MI300C instead. But the part still remains.

they need consumer dGPUs to be successful

They in fact, do not.

Xbox is dead, there won't be meaningful semi-custom money from them and Sony alone might not carry it

Yeah they will, every xbawks not sold is a PlayStation sold.

and work well with that ROCm malarkley, otherwise NVIDIA will continue to wipe the floor with them for the next decade, those tiny NPUs won't hack it.

ROCm has beyond fringe relevance for client parts.

 

[SolidQ]

yeah that hurt.

Same situation with 4060ti vs 7700Xt despite prices very close

 

[branch_suggestion]

The inconvenient truth is this, client dGPU means little to both AMD and NV at this point, for both it is a declining share of their revenue.
Client dGPU is growing slowly, it is a pretty stagnant market and while ASPs are creeping up, unit numbers are going nowhere.
The TAM is nothing vs DC hardware and far smaller than client SoC.
Companies chase the largest TAMs, this is how it is. We are in a pretty pitiful period in consumer PC hardware honestly, and it is hard to say if the market will grow back into relevance long term.
Intel being utterly incompetent at basically everything has doomed us all, at least Patty G tried.

 

[Win2012R2]

while ASPs are creeping up, unit numbers are going nowhere.

It's linked - the more you charge, the less sales you get, unless it's a monopolistic inelastic market.

Companies chase the largest TAMs, this is how it is

NVIDIA appears to be able to do both - once they get 3 GB memory chips they will solve one of the only two major problems they have: limited memory, price they will keep for sure and people will be happy to pay it, frankly at this point (and I can't believe I am saying it) Intel is demonstrating bigger commitment to dGPU market than AMD.

 

[poke01]

The inconvenient truth is this, client dGPU means little to both AMD and NV at this point, for both it is a declining share of their revenue.
Client dGPU is growing slowly, it is a pretty stagnant market and while ASPs are creeping up, unit numbers are going nowhere.
The TAM is nothing vs DC hardware and far smaller than client SoC.
Companies chase the largest TAMs, this is how it is. We are in a pretty pitiful period in consumer PC hardware honestly, and it is hard to say if the market will grow back into relevance long term.
Intel being utterly incompetent at basically everything has doomed us all, at least Patty G tried.

client dGPUs at least in laptops are moving towards to SoCs. Just wait for next gen LP memory.

Intel and AMD won’t need NV GPUs anymore but that means Nvidia also can make large SoCs for the laptop market.


For desktop, Nvidia is king. That’s not ever changing unless leadership chaos at Nv.