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

 

[gaav87]

Well at least STRIX HALO looking good.
Damn apu = 6700xt (desktop)

 

[Josh128]

Well at least STRIX HALO looking good.
Damn apu = 6700xt (desktop)
View attachment 113726

Source? Thats actually pretty impressive. Basically RX6700XT desktop perf. My run.

 

[adroc_thurston]

Ryzen revenue is higher than EPYC

Sort of.

client margin is improving

Yes I know. Strix is good for you.

From my understanding rocm has come a very long way in the last year.

Yeah but his point is that it sucks for training. Which it does. Because it's not a priority, AMD's top2 customers rn (Meta and MS) are buying their stuff for inference.

 

[RNR_Forte]

N3something maybe?

Dreams are nice.

 

[linkgoron]

Bold:
Looks like you only want them to undercut Nvidia to lower Nvidia pricing. You already said you are going that route.

You quote a post of mine where I mention that I've only bought AMD GPUs since 2008, and say that all I want is an Nvidia card. That is something else. I would 100% prefer to buy an AMD GPU if it was priced right and competitive (e.g. the 7800XT was, IMO, priced well last gen and I probably would have bought one if I was looking for a GPU at the time). However, there won't be any competitive AMD GPUs this gen at the 5070 or above level. You can choose to wait two more years and see if AMD will actually release a decent GPU, but there are no promises that 1) they release a decent GPU and 2) that it will be priced well. Waiting for AMD's next gen is a tough ask.

Underlined:
Do you own a chiplet RDNA3 card? Doesn't seem a dud to me.

There's no doubt that RDNA3 vastly underperformed.

 

[Joe NYC]

Yeah but his point is that it sucks for training. Which it does. Because it's not a priority, AMD's top2 customers rn (Meta and MS) are buying their stuff for inference.

BTW, the latest OpenAI O3, according to my understanding, the advances are mainly on inference. Like totally tipping the scales from biggest advances coming from training (in the past) to the biggest advances coming from inference (in o3).

And then, the best inference queries can consume huge amount of compute resources. Over and over...

 

[adroc_thurston]

BTW, the latest OpenAI O3, according to my understanding, the advances are mainly on inference. Like totally tipping the scales from biggest advances coming from training (in the past) to the biggest advances coming from inference (in o3).

And then, the best inference queries can consume huge amount of compute resources. Over and over...

Yeah they're throwing compute (money) at the problem for ever more incremental gains.
We're peaking the bubble.

It's a funny desperate attempt to prevent investors from asking the hard questions (where is your path to profitability) and then breaking Altman's kneecaps soon thereafter.

 

[Joe NYC]

Yeah they're throwing compute (money) at the problem for ever more incremental gains.
We're peaking the bubble.

It's a funny desperate attempt to prevent investors from asking the hard questions (where is your path to profitability) and then breaking Altman's kneecaps soon thereafter.

Well, it is getting better at solving problems, but it is getting better at it by using astronomical resources on inference. Like $350,000 (inference costs) on a very complex query when the answer requested was to have very high degree of accuracy.

(at around this timestamp)

 

[Win2012R2]

I'm only relaying what Mark Cenry

He wants to pay max of 250-300 bucks for a GPU or less - so for him the future is fake generated frames, aggressive low-res upscaling (AKA "ultra performance" DLSS) going to be preferable die size wise than having proper muscle, which they certainly won't be able to afford in future hyper expensive nodes like N2

 

[Joe NYC]

Well at least STRIX HALO looking good.
Damn apu = 6700xt (desktop)
View attachment 113726

This looks quite good, considering that majority of current notebooks with dGPUs have 4050, 4060 and 4070. Only a small number has higher than this.

Strix Halo may not necessarily be greatly cost optimized to beat the dGPU solutions on price by a big margin, but overall system design, in intelligently sharing memory and sharing power envelope should make it a winner.

It would make a sweet casual gaming machine, perhaps return of Steam Machine.

 

[tajoh111]

Well at least STRIX HALO looking good.
Damn apu = 6700xt (desktop)
View attachment 113726

Impressive for an APU. This might be AMD best 2025 product if they price it reasonably.

 

[gaav87]

Impressive for an APU. This might be AMD best 2025 product if they price it reasonably.

Wait till you see the ppt 2xx+ xD

 

[DaaQ]

You quote a post of mine where I mention that I've only bought AMD GPUs since 2008, and say that all I want is an Nvidia card. That is something else. I would 100% prefer to buy an AMD GPU if it was priced right and competitive (e.g. the 7800XT was, IMO, priced well last gen and I probably would have bought one if I was looking for a GPU at the time). However, there won't be any competitive AMD GPUs this gen at the 5070 or above level. You can choose to wait two more years and see if AMD will actually release a decent GPU, but there are no promises that 1) they release a decent GPU and 2) that it will be priced well. Waiting for AMD's next gen is a tough ask.

There's no doubt that RDNA3 vastly underperformed.

So you are all in on a 5090 right?

You obviously don't frequent OCN where they have empirical data.

See you mentioned again priced well. It seems they are never priced well. Even if it is 20 dollars up. Like the 9800x3d 30 dollars up, oh my gaud.

 

[insertcarehere]

So you are all in on a 5090 right?

You obviously don't frequent OCN where they have empirical data.

See you mentioned again priced well. It seems they are never priced well. Even if it is 20 dollars up. Like the 9800x3d 30 dollars up, oh my gaud.

For somebody who bought high-end Ampere/RDNA2 and is looking to upgrade on a 4-5 year cycle, both fairly reasonable assumptions, is there a realistic upgrade path that doesn't involve Nvidia?

 

[blckgrffn]

For somebody who bought high-end Ampere/RDNA2 and is looking to upgrade on a 4-5 year cycle, both fairly reasonable assumptions, is there a realistic upgrade path that doesn't involve Nvidia?

If you’re looking at spending a $1k+ because that’s your budget, no, it doesn’t look like it and it hasn’t been that way since the 4090 was released. The 3090/3090ti situation was pretty similar although it was less expensive. The 6900XT wasn’t really a 3090 slayer in the same vein that the 7900XTX can’t shut the door on the 4080S.

This top card available level of enthusiast hasn’t looked to AMD since what, the 7970 days? For the largest segment of the buying public little is changing except ceding one more tier of top performance. Instead of it being the halo, it’s like halo+1.

If Next Gen AMD comes back with even a “serious” $1k card we will be back to business as usual.

I’ll keep an eye on the new Nvidia App for the time being. Despite what others have said, the AMD software (overlay, chill, power and clock controls, etc.) and the display hardware ecosystem (freesync premium is fantastic and available on all flavors of monitors at many price points) and I am have always been about the value so I’m not buying a $1k card or a $1.5k or $2k or $2.k card anyway. Spending $500-$600 might be my limit for now anyway.

 

[poke01]

That also big problem for AMD, they avoid fixes like plague. For real they need new software team, which can be creative, feels like AMD software team is 70-80 years guys

Software is hard. Really hard.

Well at least STRIX HALO looking good.
Damn apu = 6700xt (desktop)
View attachment 113726

this isn’t RDNA4, shouldn’t be in this thread. I’ll make a separate thread, for Strix Halo GPU

 

[marees]

Anyone pushing rumors that it performs better than expected are malicious. It is clearly not mere stupidity anymore.

I don't know why they do it.

And here's why they do it. To pretend to be mad when it performs as 32 WGP part should.

"AMD lied"

 

[adroc_thurston]

Software is hard. Really hard

Not a software thing, ATi always sucked at turning their world-class IP into desirable products and the curse never went away.

 

[DaaQ]

For somebody who bought high-end Ampere/RDNA2 and is looking to upgrade on a 4-5 year cycle, both fairly reasonable assumptions, is there a realistic upgrade path that doesn't involve Nvidia?

Yea, unless you want anemic VRAM or to pay the cost of 2 iPhones to get the top tier.

From what I've read, alot of you only settle for the midrange, 4060 variants, RDNA2 with lower RT, god forbid raster is good. Gotta have that DLSS.

Then we go down the road of power consumption on desktop. FFS, 600w PCIE connectors. No problem, but screw the 3 8 pin connectors because reasons.

There is a thread on OCN with actual owners, they list quite a few benchmarks, and yea XTX is not as low as y'all think it is. There is even a guy with a XT that is taking #1 spots on HWBOT, IIRC, it may be 3Dmark. It's been awhile.

 

[insertcarehere]

Yea, unless you want anemic VRAM or to pay the cost of 2 iPhones to get the top tier.

Answer the question, why would a 7900 XTX (if that doesn't get production discontinued) move the needle for these buyers in 2025 if they didn't move the needle for these same buyers in 2023?

And it's not as if the reason NV's flagship is going to cost 2 iPhones, has anything to do with a complete lack of competition to keep them more honest...

From what I've read, alot of you only settle for the midrange, 4060 variants, RDNA2 with lower RT, god forbid raster is good. Gotta have that DLSS.

Customers wanting additional performance in games when Raster performance has barely budged in those same mainstream pricepoints would explore quality upscaling solutions? *shocked Pikachu face*

Intel already has better RT/upscaling solutions in and they have only really entered the dGPU market in 2024, it is well past time for excuses.

Then we go down the road of power consumption on desktop. FFS, 600w PCIE connectors. No problem, but screw the 3 8 pin connectors because reasons.

There is a thread on OCN with actual owners, they list quite a few benchmarks, and yea XTX is not as low as y'all think it is. There is even a guy with a XT that is taking #1 spots on HWBOT, IIRC, it may be 3Dmark. It's been awhile.

Citing user-submitted data on one particular DIY website (a stagnant one at that) as a representative sample of all dGPU owners in the enthusiast space? Yikes and I remember the fits members in this forum made when people cited Steam surveys as a proxy marketshare..

 

[linkgoron]

So you are all in on a 5090 right?

With regards to performance, the 5070TI (if it exists), 5080 and 5090 will all outperform AMD's best (and probably 5070 but we'll see). 5090 is not the only option if you'll want something better than the top Navi 48.

See you mentioned again priced well. It seems they are never priced well.

I've already mentioned in the post that YOU'RE QUOTING that I thought that the 7800XT was priced well (IIRC the 7900GRE was also priced well).

Even if it is 20 dollars up. Like the 9800x3d 30 dollars up, oh my gaud.

What a joke - you're totally barking at the wrong tree. Here's a quote from a post that I MADE mocking that point exactly:

I don't get it - Nvidia raises prices by hundreds of dollars every generation and they get totally sold out everywhere, but AMD adding $30 to the price of what appears to be the best gaming CPU around is the end of the world?

If you’re looking at spending a $1k+ because that’s your budget, no, it doesn’t look like it and it hasn’t been that way since the 4090 was released. The 3090/3090ti situation was pretty similar although it was less expensive. The 6900XT wasn’t really a 3090 slayer in the same vein that the 7900XTX can’t shut the door on the 4080S.

6900XT wasn't a 3090 killer - but it was close enough that someone who preferred AMD could easily buy one without feeling that he's losing a lot (ignoring RT), especially as the MSRP was $1000 vs 3090's $1500. This is the performance at 4k from TPU at launch:



6800XT should have just been priced a bit lower. I'd buy one instead of 3080 even at MSRP but really, AMD just mostly price matched with Nvidia while having much worse RT. They could have undercut Nvidia by a bit more than $50 with the 6800XT, but they just didn't. The following is the price/perf chart at 4k:


Really, the RDNA2 lineup was fantastic. Sure, it was also close because Nvidia went with Samsung and not just thanks to AMD's engineering, but AMD missed a huge chance to increase their market share.

This top card available level of enthusiast hasn’t looked to AMD since what, the 7970 days? For the largest segment of the buying public little is changing except ceding one more tier of top performance. Instead of it being the halo, it’s like halo+1.

6800, 6800XT and 6900XT were good enough, even if the 6900XT lost by a bit to the 3090.

I’ll keep an eye on the new Nvidia App for the time being. Despite what others have said, the AMD software (overlay, chill, power and clock controls, etc.) and the display hardware ecosystem (freesync premium is fantastic and available on all flavors of monitors at many price points) and I am have always been about the value so I’m not buying a $1k card or a $1.5k or $2k or $2.k card anyway. Spending $500-$600 might be my limit for now anyway.

The fact is, that AMD just doesn't really provide much value over Nvidia at MSRP, to fight Nvidia's marketing machine (DLSS, RT, or previously stuff like PhysX or 3D vision). AMD just won't accept that they need to do more to get into gamers hands, and instead time and time again launches cards at bad MSRPs only for them to lower the price shortly after.

 

[SolidQ]

AMD missed a huge chance to increase their market share.

They increased it with RDNA3, but after super series and no alternative to DLSS + Path Tracing, dropped fast
RDNA4 have FSR4 as alternative for DLSS, only question Path Tracing, no doubt it's good in RT mode

Probably the 5070

it's have 12gb, so pretty meh card

 

[marees]

The fact is, that AMD just doesn't really provide much value over Nvidia at MSRP

Value is subjective to the users

AMD always gives better raster than Nv for same price point

FSR 3.1 better than PSSR. & soon we will have WMMA based FSR 4

DLSS FG is broken on 8gb cards.

Basically with NV you can buy only 2 cards: the xx90 or xx70 ti super. Everything else is poor value

If value was the argument you are not matching the value of 6800 for another couple of years atleast or more.
AMD 6800 sold for as low as $340 recently

 

[ajsdkflsdjfio]

RDNA4 have FSR4 as alternative for DLSS, only question Path Tracing, no doubt it's good in RT mode


it's have 12gb, so pretty meh card

RDNA4 may be an alternative to current version of DLSS, but what if Nvidia debuts yet another AI technology like their rumored "neural rendering" or just in general stay one step ahead by further improving DLSS/FG. As for RT, I just don't see AMD catching up in that department either.

Their still going to be playing catchup in many ways and yet they are rumored to be pricing their cards as if they are Nvidia's peer, plus even if that were true, AMD simply doesn't have the mindshare to be pricing their products at parity with Nvidia unless they want their marketshare to stagnate. Even if they quickly lower prices, it will leave a stain on RDNA4 like the horrible pricing left a stain on Zen 5's launch.

Value is subjective to the users

Even considering Nvidia's huge mindshare, AMD's marketshare has tanked throughout RDNA3. It's clear that offering +5-10% raster at the same price point while having generally worse features, software, and no CUDA is just not worth it for most people.

Basically with NV you can buy only 2 cards: the xx90 or xx70 ti super. Everything else is poor value

If value was the argument you are not matching the value of 6800 for another couple of years atleast or more.
AMD 6800 sold for as low as $340 recently

This might have been true at their launch MSRPs but it is no longer true now. 4080 super and 7900xtx have been within 50-100 dollars for a long time with essentially the same raster performance. Although the 4080 super is slightly more expensive it offers much better RT performance, superior drivers and upscaling technology, and has CUDA. For someone who only plays rasterized games and doesn't like to use upscaling technology maybe the 7900xtx would be a clear winner, in any other scenario though the 4080 super is competitive if not better.

The rx 6800 and other RDNA2 cards continue to be excellent value options but they aren't really mainstream products. To my knowledge AMD no longer produces these cards and the cards currently being sold are just the large inventory of RDNA2 cards that existed before its discontinuation. I think most of the buyers are more likely to flock to newer generations of cards even if an older generation provides excellent value. AMD isn't winning anything when AIBs sell their massive backlog of RDNA2 cards, at much lower margins than newer products.

 

[poke01]

I agree RDNA2 was fantastic, why can’t AMD do a full lineup like that for RDNA4 and 5?