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]

it doesn't mean anything, need real gaming scenario
in port royal 7900Xt beat 4070ti, but in real gaming scenario different story

Spoiler

Oh it means
(0 RT older games) = 7900gre
Light RT games possibly UE5 = 7900xt
Heavy RT games 7900xtx<

 

[gdansk]

It seems to land between the 7900 GRE and XT exactly as predicted months ago from the WGP count and this guy says

Well guys time to pack your bags

🤡

 

[beginner99]

It seems to land between the 7900 GRE and XT exactly as predicted months ago from the WGP count and this guy says

🤡

Agree, the issue is the rumored $599 price tag which does barley nothing once again for generational performance/dollar uplift. you could get a GRE for that price some weeks ago. (actually still now available for that price in my local market and currency adjusted and it is actually then $599 incl. all taxes, while with the $599 MSRP + taxes plus "EU tax" it will end up with worse performance/$ than a 7900 GRE)

 

[jpiniero]

Agree, the issue is the rumored $599 price tag which does barley nothing once again for generational performance/dollar uplift. you could get a GRE for that price some weeks ago. (actually still now available for that price in my local market and currency adjusted and it is actually then $599 incl. all taxes, while with the $599 MSRP + taxes plus "EU tax" it will end up with worse performance/$ than a 7900 GRE)

That's why AMD letting the Fire Sales happen is so problematic.

 

[gaav87]

Just buy 6nm stock. Slap on 24gt/s samsung mem instead of 18GT/s +increase clocks thanks to 6n
Dont change, a damn thing from 6800xt except the above call it 8800xt 499$ = mss xD No need for R&D and marketing = lots of $$$ xD

 

[blckgrffn]

I still want to see better perf/watt. I know it's not that sexy but if I could get that perf in the same general envelope as the 6800 that'd be swell.

I did the 290X thing back in the day (and ran it for years!) but it was $330 and the 7900XTX is... a lot more than that.

Personally hoping it ends up closer to 7900XT vs 7900GRE but actual gameplay will reveal that. All this last minute pontificating is sorta exhausting to keep up with.

 

[SolidQ]

It seems to land between the 7900 GRE and XT

let's wait gaming perfomance, then we can happy or dissapoint. I'm never trust in those 3d mark scores
2 weeks left

 

[gaav87]

It seems to land between the 7900 GRE and XT exactly as predicted months ago from the WGP count and this guy says

🤡

Yep time to move to green first time since 8800GT no other choice.

 

[SolidQ]

since 8800GT no other choice.

you forgot HD 5xxx, HD 6xxx, HD 7xxx. Especially 7xxx was real winner vs kepler

 

[ToTTenTranz]

Because it's cope and he should talk less.
Sometimes the best way to play is not playing at all and they won't be playing client dGFX anymore (at large).

because words cost nothing......

Sorry but I don't buy this.

Huynh is the executive of a $200B tech company, not a teenager on the internet. He's Senior VP and General Manager of the Graphics and Computing Business Group, meaning he knows very well what AMD's current business strategy on GPUs is. He probably has more agency over business strategy than David Wang does, as the latter is focused on Engineering.

He had little to gain in conceding the interview, let alone lying his ass off to the interviewer. If he said in September that AMD's strategy was to gain marketshare on consumer dGPUs, then that's what I will assume AMD's strategy was in September.
That might have changed, plans change. If the top-end N48 comes out for 700€ then I'm sure plans did change. But AFAIK all we've got so far are rumors and for now I'll trust the statements from an AMD official over rumors.

You already just wanted a "stop gap" card anyway. Just buy a 5070ti or 5080 and call it a day.

Because I'm not a drone who follows trends/orders, and because everyone buying nvidia regardless is the main reason we're in this mess.
I will buy nvidia if I'm out of decent options elsewhere, but I'd rather not.

 

[gaav87]

you forgot HD 5xxx, HD 6xxx, HD 7xxx. Especially 7xxx was real winner vs kepler

Oh i still have hd4850 hd5870 and hd7870

 

[reaperrr3]

Nope, nothing stops you from building a chungus if you have class-leading PPA.
That was quite literally the issue with small die strategy: they won majorly on PPA and capitalized exactly zero times on it.
And then NV went back to PPAmaxing with Kepler and there you go.

Of course I don't know if they might've lied, but I remember an article (might've even been here on anandtech) according to which DAAMIT's big R700 would also have been only 800 ALUs, for RV770 they only stripped down other things to reduce die size (possibly including ROPs though, which *did* matter for perf/clk at that time).

But yeah, it's been maddening how AMD never did the right thing with the right IP.

- RV670 with a 5th SIMD: Would've been way more competitive vs. G94.
- A big R700 with 2 SE, 2x6 SIMDs (960 ALUs) and 32 ROPs would've still been sub-400mm² and demolished GT200.
- RV870 with 24 SIMDs @ 900 MHz would've still been below 400mm², relatively cool and efficient, and enough to beat the GTX 480 in anything but tesselation.
The 58x0 were also priced too aggressively, AMD was far too afraid of Fermi being good and self-deterred themselves into making no money with that gen, even though they could have.
- Tahiti with 4 SE, 40 CUs and 64 ROPs (close to Hawaii specs) would've been ~450mm², but would've demolished GK104 and given GK110 a hard time.
- Pitcairn with 24 CUs would've given anything below the GTX 680 a hard time.
- Hawaii with just 4 CUs more and some GCN3 tech (DCC, doubled L2) could've kept the mem interface power consumption at saner levels and would've had an easier time competing at least against GM204.
- Tonga was the worst PPA part of all GCN gens, so much wrong with that design, a waste of space and opportunity in so many regards
- A big Polaris with 52-56 CUs, 64 ROPs and 384bit MI could've beaten the 1070 at least, and would've been only like 360mm², much smaller than Vega10 and not much bigger than GP104, so margins wouldn't have been much worse than Nvidia's.
- Navi10 with 48 CUs would've done a lot better against Turing, allowing higher prices.
- Navi22 with 48 CUs would've been only like 7% bigger, for at least ~12-15% higher perf, enough to do better against the 3070(Ti), allowing higher prices.

The sheer length of this list speaks volumes.

It's like some managers at AMD were so obsessed with GPU PPA they forgot that higher performance allows for higher prices, and that sometimes just a few more SIMDs/CUs make for better $PA.

Meanwhile, AMD kept wasting lots of area and margin on APU IGPs that were completely bandwidth-starved and would've barely lost performance by removing 25-40% of the SIMDs/CUs from Llano all the way to Picasso, before sanity returned. Mindboggling stuff.

Irrelevant bling.

As much as I want AMD to become fully competitive in dGPU again (and no, not to buy NV cards cheaper, but rather because I'd buy a good AMD card competitive in PPW in a heartbeat):
As 3060 Ti owner, who was able to play MW5 Clans at an image quality rivaling the good ol' 4xSGSSAA DX9 days only thanks to DLSS, I unfortunately have to strongly disagree.

I could care less about RT, but when implemented well, DLSS is a game changer AMD currently has no proper answer to, at least in terms of image quality.
And I'm frankly skeptical about how much FSR4 will improve things.
AMD surely won't spend big money to build server farms just to train their FSR4 algos, so I have a hard time imagining they can catch up and are at risk to fall behind further, actually.

Nvidia has reached a critical mass of financial advantage that allows them to literally buy themselves feature advantages that even a perfect hardware gen of AMD would only help so much against, which is worrying but in part AMD's own fault for botching too many opportunities in the past.

 

[gdansk]

Yep time to move to green first time since 8800GT no other choice.

You were waiting for something that was not going to happen. It's a 32 WGP part. If you wanted higher performance from AMD it has been available for two years.

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

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

 

[SolidQ]

Hidden pr for FSR4

 

[adroc_thurston]

that the chiplet strategy has been a disaster for AMD's client GPUs

No it's good.

This is similar to what happened with Polaris, with almost exactly the same things being said at the time

Polaris was a worse uarch. Not comparable at all.

Nvidia has more software eng than hardware ones

Good news, Meta (the actual real driver of the CUDA roadmap) has probably 10x that, too.

Especially 7xxx was real winner vs kepler

it wasn't, not on launch anyway.

Huynh is the executive of a $200B tech company, not a teenager on the internet. He's Senior VP and General Manager of the Graphics and Computing Business Group, meaning he knows very well what AMD's current business strategy on GPUs is. He probably has more agency over business strategy than David Wang does, as the latter is focused on Engineering.

He has no agency because no matter how much he wants the chainsaw, no matter how much they need it for brand revival, no one at AMD would be able to justify the program costs to quite literally gamble.

I could care less about RT, but when implemented well, DLSS is a game changer AMD currently has no proper answer to, at least in terms of image quality.

Native image is very much that. next.

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

It's a bandwidth thing.

 

[blckgrffn]

It's a bandwidth thing.

A GDDR7 refresh addressable thing or organizational bandwidth thing?

 

[adroc_thurston]

A GDDR7 refresh addressable thing or organizational bandwidth thing?

GDDR7 is more cost for a very mainstream part so they won't be doing it.
But the IMC supports it iirc.

 

[blckgrffn]

GDDR7 is more cost for a very mainstream part so they won't be doing it.
But the IMC supports it iirc.

Right on. A year from now it might be a different calculus. Not saying I’ll hold out for it but “maybe” it could happen. Thanks.

 

[linkgoron]

Sorry but I don't buy this.

Huynh is the executive of a $200B tech company, not a teenager on the internet. He's Senior VP and General Manager of the Graphics and Computing Business Group, meaning he knows very well what AMD's current business strategy on GPUs is. He probably has more agency over business strategy than David Wang does, as the latter is focused on Engineering.

He had little to gain in conceding the interview, let alone lying his ass off to the interviewer. If he said in September that AMD's strategy was to gain marketshare on consumer dGPUs, then that's what I will assume AMD's strategy was in September.
That might have changed, plans change. If the top-end N48 comes out for 700€ then I'm sure plans did change. But AFAIK all we've got so far are rumors and for now I'll trust the statements from an AMD official over rumors.

I've posted this before at least twice, but I'll post it again - here's AMD's TAM strategy for Polaris the last time they were extremely behind with just a mainstream lineup (basically 480 was kind of competitive with the 1060, Nvidia was alone with 1070/1080 and of course 1080ti later on). I remember that the "cheap" 4GB (for $200) was mostly marketing and in reality it was just for the initial launch or for a very short while and then they actually stopped making them (so actual entry price was $240), but I couldn't find a source to back my memory up. When AMD is extremely behind they always claim how they're going for TAM or whatever. What are their other options? Saying that they have a weak brand? that Nvidia is pushing for RT and DLSS and that they're behind?

They could say that they've failed with their chiplet strategy for client, and how they wrecked four years of client GPUs (RDNA3/RDNA4 at least), but they don't want to do that. They've failed before with going all-in on HBM on client, and now they've made a similar mistake with chiplets.

Because I'm not a drone who follows trends/orders, and because everyone buying nvidia regardless is the main reason we're in this mess.
I will buy nvidia if I'm out of decent options elsewhere, but I'd rather not.

I get it, I've had solely AMD cards since the Radeon HD 4870 (last Nvidia card was the 6600GT), but given AMD's recent behavior even if they'll have a winner, they won't significantly undercut Nvidia. They had a competitive lineup with RDNA2 and didn't really undercut Nvidia at launch MSRP-wise with the 6800xt and 6700xt. RDNA3 also (IMO) didn't really undercut Nvidia enough. They've shown this time and time again. We'll see what happens with RDNA4, but given RDNA3's pricing with the 7900xtx and 7900xt - I'm not hopeful. Only when it was very clear that RDNA3 was a dud, only then did AMD provide decently priced cards - with the 7800XT and 7900GRE.

 

[adroc_thurston]

They could say that they've failed with their chiplet strategy for client, and how they wrecked four years of client GPUs (RDNA3/RDNA4 at least), but they don't want to do that.

it did not.
what makes you even think that?

 

[GodisanAtheist]

it did not.
what makes you even think that?

Devil's Advocate: 530mm2 of silicon on a relatively complicated packaging process just only matched NV's 380mm2 die size in raster with far worse ray tracing performance and "feature set" (regardless of its value to you).

Of course, RDNA3's issues seem to stem from its architecture rather than its chiplet packaging, but it can be hard to separate the two without direct acknowledgement from AMD on what went wrong.

 

[adroc_thurston]

Devil's Advocate: 530mm2 of silicon on a relatively complicated packaging process just only matched NV's 380mm2 die size in raster with far worse ray tracing performance and "feature set" (regardless of its value to you).

That's not a chiplet issue at all.

but it can be hard to separate the two without direct acknowledgement from AMD on what went wrong.

Very easy, no amount of tiling impacts fmax like that.

 

[linkgoron]

it did not.
what makes you even think that?

You're right, AMD's chiplet strategy with RDNA is a massive success. RDNA3 totally destroyed Nvidia's relatively simpler monolithic solutions and chiplet based RDNA4 didn't get completely cancelled leaving AMD with a small-die monolithic design that can just about compete with Nvidia's third or fourth tier cards.

 

[adroc_thurston]

RDNA3 totally destroyed Nvidia's relatively simpler monolithic solutions

It did okay despite a massive fmax miss?

and chiplet based RDNA4 didn't get completely cancelled

You can do bigger and better things instead.

leaving AMD with a small-die monolithic design that can just about compete with Nvidia's third or fourth tier cards.

Yeah, that's the point. They gave up.
Because there's no market condition where client dgfx ever makes them money back.

 

[gaav87]

Guys remember 7900xt and 7900xtx leaked Timespy and Firestrike scores ? They were 10-13% lower vs release drivers.

First AMD Radeon RX 7900 XTX/7900 XT 3DMark TimeSpy/FireStrike scores are in - VideoCardz.com

AMD Radeon RX 7900 in 3DMark We now have the very first confirmed 3DMark results of AMD’s upcoming Radeon 7900 series. The RX 7900 series cards perform just as expected and akin to what we know from RDNA2 series. This means better 3DMark FireStrike performance, which is a DirectX11 benchmark...

videocardz.com

Real 9070XT score is 26k with release drivers.

9070xt release drivers:
We detected, a driver timeout has occurred...
"Close" / "Report issue"