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

 

[branch_suggestion]

Probably wise from AMD, RDNA4 is console driven, this being for the mid gen hardware release. And the console business is good, shame Nintendo isn't trying to pursue a built for purpose SoC and using sloppy seconds again from a company that couldn't think less of them, even after the whole easily avoidable hardware/firmware loophole fiasco.
Extreme GPU configs rely on having good enough displays (4K240 self emissive for <$2k) and demanding killer apps to really draw mainstream desire.
Considering other factors, like demand elsewhere and the node timing from TSMC it is wise to make a killer midrange part and focus on executing the DC godzillas.

A config I have in mind is based on MI300 packaging tech, stacked GCX on base die.
GCX: 4SE/32WGP N4P ~150mm^2
Base die: 192-bit GDDR7 768GB/s 18GB/48MB MALL/IO N6 ~175mm^2

Yes hybrid bonding adds costs and whatnot, but I prefer this to reusing the MCD concept from RDNA3. It would allow for a really fast and low latency LLC and really low power overhead for a chiplet design. IO like display, media, PCIE and other accelerators could technically be another chiplet to the side of the base die, but this is probably not worth it. Could also be monolithic, but that is way too safe, this design needs to be a proof of concept, even if it is missing multiple coherent compute dies.
Total die area is similar to N32, costs would also be similar but performance would be significantly higher, main thing is if RDNA4 hits design goals it would trounce the scuffed RDNA3 in PPAC. IPC should also have a nice little bump. 1.5x N32 at iso-power? Well that would mean it basically matches N31 with a smaller config and less power draw, so maybe a bit ambitious. RDNA3 however is pretty broken in all implementations based on it having the worst frequency curve in games since Fermi, sans woodscrews. Turns out doing pre-Si estimates based on headless compute (which works as intended) might not cut it.

 

[Saylick]

IMO, this is another nonsense rumor. Only going to about Polaris level ( about RTX 5060 competition) next generation makes no sense at all.

I could see them abandoning the very top level of competition, but not the midrange.

Their best selling RX 6000 part is the 6700 XT.

Having only low end parts means they could may also even get buried by Intel Battlemage GPUs.

Nothing about his makes sense, thus it's nonsense.

I think you're reading into it too literally when they reference Polaris level performance.

To me, it seems like the top SKU for RDNA4 will only roughly match current AD102 levels of performance. Basically, it's what N31 should have been from the start. Fix the HW bugs, optimize the drivers, and just get the product out there while you still can salvage a chiplet based approach.

Considering that the economics and cost-benefit isn't in AMD's favor to make gigantic dies (no pro vis market to support it), I'm not surprised they decided they didn't want to chase the flagship against Nvidia when the odds simply aren't in their favor.

Put in a more positive point of view, if AMD could fix N31, get RTX 4090 levels of raster performance, and keep prices relatively constant, they can still target the vast majority of the market with a decent product. Who here wouldn't want a product that gives them 4090 levels of raster performance for $1000?

 

[Heartbreaker]

I think you're reading into it too literally when they reference Polaris level performance.

Not really. Why say that, if they didn't intend that comparison. Polaris was a GTX 1060 competitor. In current terms that would be RX 7600, or RX 8600 next generation.

There are MUCH better examples if it was intended to be some kind of midrange cutoff. Like RDNA 1, RX 5000 series, that topped out with Navi 10, RX 5700 XT.

 

[exquisitechar]

OUCH. GG if true.

https://twitter.com/x/status/1687388515670056960

Judging by the predating @uzzi38 reaction this looks to be true:

https://twitter.com/x/status/1687379360276185088

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?

AMD can only blame themselves if this is true. They can’t release more than one solid GPU generation before imploding like they did with RDNA3. Terrible execution with equally terrible marketing for years on end. They hadn’t made a serious competitor for many years before RDNA2 and then they blew it again with RDNA3. If all they keep delivering is worse HW and worse SW than Nvidia with nearly equal pricing, they may as well pull out of the market entirely and not just the high end, because near zero percent market share is all that awaits them in the long run.

But hey, maybe this negative hype train means AMD doesn’t suck for once and RDNA4 brings a solid midrange competitor with a fixed uarch and more than the bare minimum for RT acceleration.

 

[Saylick]

Not really. Why say that, if they didn't intend that comparison. Polaris was a GTX 1060 competitor. In current terms that would be RX 7600, or RX 8600 next generation.

There are MUCH better examples if it was intended to be some kind of midrange cutoff. Like RDNA 1, RX 5000 series, that topped out with Navi 10, RX 5700 XT.

I mean, the Tweet does say "Think of it like Polaris or RDNA1"

I interpret that as meaning the top SKU targets the -70 class GPU of Nvidia's next generation lineup. It's not going to be the fastest, but it does cover a pretty big segment of the overall market.

 

[coercitiv]

Relax folks, if AMD still aims for up to ~5070 class performance then we're covered up to $999 anyway

 

[gdansk]

Relax folks, if AMD still aims for up to ~5070 class performance then we're covered up to $999 anyway

Yeah that's what I was thinking. Who knows what the next generation 5070 will cost if they can get away with a $1200 4080.

AMD not trying to make something to compete with a rumored 512 bit beast is wise. And maybe it marks a return to trying to offer a better value (probably not).

 

[JasonLD]

Could it be that RDNA4 big die was never planned to begin with and high end was going to be 2x MCM solution but that didn’t quite work out? So AMD got stuck with only a midrange for RDNA4.

 

[Leeea]

Relax folks, if AMD still aims for up to ~5070 class performance then we're covered up to $999 anyway

How big the market over $999 anyway?

AMD did make a lot of rx580s.

 

[adroc_thurston]

Could it be that RDNA4 big die was never planned to begin with

Wasn't really a die per se.

 

[adroc_thurston]

AMD not trying to make something to compete with a rumored 512 bit beast is wise

They had something way, way way more horrific than that.
Deferred till later, though.

 

[gdansk]

How big the market over $999 anyway?

Big enough when that's the only option. Some day soon.

 

[moinmoin]

They had big plans and cancelled them because they felt they didn't have the time to chase them.

There are potentially two reasons why:

1. Validation for what they were trying to do would bring the halo parts too close to RDNA5 in terms of timescale.

2. They're giving up (or if you want to phrase it nicely, deprioritising) desktop graphics.

If you want to ask me which one I think it is, then to put it kindly, I'm leaning towards the second personally.

I could VERY easily be wrong, but this is the way I'm leaning for now.

There are essentially four areas AMD has to do R&D for "graphics" compute: servers, consoles, iGPU and dGPU.

• Consoles are financed through big longterm corporate partners in Sony, Microsoft and potentially Valve and others.
• Servers is by far the biggest margin market available, the AI craze will let sales of non-CPU compute soar there.
• iGPU is a necessity in mobile and desktop chips nowadays, a requirement to get access to the steady market of enterprise and OEM sales.
• That leaves dGPU which for AMD so far is a high risk low reward market. How high the risk is RDNA3 is showing: Disaggregation of GPU dies, extensive reworking of the platform, with ambitious targets regarding performance and efficiency. Those targets were missed to a degree that RDNA3 barely makes a difference over RDNA2, meaning the investment in a whole dGPU gen is essentially being wasted there.

In which of these four areas would you invest your own money for further improvements?

 

[Joe NYC]

Relax folks, if AMD still aims for up to ~5070 class performance then we're covered up to $999 anyway

View attachment 84039

5070 $999
5080 $1,499
5090 $1,999
5090 Ti $2,499

That's the pricing gamers deserve, pricing they have earned.

 

[MrPickins]

IMO, this is another nonsense rumor. Only going to about Polaris level ( about RTX 5060 competition) next generation makes no sense at all.

I could see them abandoning the very top level of competition, but not the midrange.

I know tiers have gotten skewed a bit the past few generations, but Polaris and x060/x065 are midrange GPUs.

Polaris was a solid product by AMD; I'd be fine with a similar release in the 5000 series.

 

[Heartbreaker]

I know tiers have gotten skewed a bit the past few generations, but Polaris and x060/x065 are midrange GPUs.

Polaris was a solid product by AMD; I'd be fine with a similar release in the 5000 series.

Would you be fine if the only thing they released for this generation was the RX 7600?

 

[adroc_thurston]

That leaves dGPU which for AMD so far is a high risk low reward market. How high the risk is RDNA3 is showing: Disaggregation of GPU dies, extensive reworking of the platform, with ambitious targets regarding performance and efficiency. Those targets were missed to a degree that RDNA3 barely makes a difference over RDNA2, meaning the investment in a whole dGPU gen is essentially being wasted there

This has more pronounced impact on APUs.

Again, they're still focusing on building big halo parts first and foremost; just that those get killed if they look late and the $next_thing schedule might vaguely slip due to them.

 

[Saylick]

5070 $999
5080 $1,499
5090 $1,999
5090 Ti $2,499

That's the pricing gamers deserve, pricing they have earned.

We'll likely never see a 5090 Ti, but I agree that the price tiers will just move on up. Little to no perf/$ gain from Nvidia at all. If you want X% more performance, you're paying X% more money. It's what happens when there's no competition.

At some point, we'll just see stagnation from Nvidia because 1) no one can challenge them, so they are basically competing against themselves trying to convince previous owners to upgrade, 2) density increases and reduced $/xtor from new nodes are harder to come by, thereby limiting performance gains, and 3) they'd rather invest in more lucrative market segments and would not mind decreasing desktop GPU volume if they know they won't lose market share.

 

[tajoh111]

I think this is sensible for AMD. Make one or two SKU and make it good.

When AMD didn't focus on the highend they had 30-40% marketshare. Lower margins but the volume made up for it. 300-400 million a quarter is a lot better than they are doing now for discrete.

The idea AMD didn't get marketshare when they offered better value from the onset is a lie. 4870, 5870 and 7970/290x(after they dropped pricing and did a renaming) is when they had peak marketshare. The 7970/7870 marketshare collapse to the resurgence of them gaining marketshare after the price drops and renaming show that people will buy AMD if they are substantially better value. Polaris was not bad either. A good mid range is even more important now when you consider how much more prominent laptop gaming is.

This 15% marketshare is the result of them trying to price themselves like Nvidia and using hate marketing to convince customers to pay Nvidia prices for AMD card. Only a small portion of the population is willing to do it.

With AMD divesting away from discrete and likely more into AI and datacenter, this was likely a way to get more engineers without spending.

 

[adroc_thurston]

If you want X% more performance, you're paying X% more money.

That's GPUs in general now.
Xtors aren't getting cheaper and every % of perf on GPUs costs a lot of 'em.

 

[MrPickins]

Would you be fine if the only thing they released for this generation was the RX 7600?

The 7600 isn't exactly a Polaris analogue. Polaris was highly competitive at a reasonable price point and had more VRAM than its direct competitor. The 7600 is a disappointment, and, like the 4060, should be branded a 50 series card, IMO.

 

[adroc_thurston]

When AMD didn't focus on the highend they had 30-40% marketshare.

They're not interested in MSS; they're interested in margins.

This 15% marketshare is the result of them trying to price themselves like Nvidia

Yes, good for margins.

that people will buy AMD if they are substantially better value

AMD will never provide 'better value' anymore.
Forget about it.

Polaris was highly competitive at a reasonable price point and had more VRAM than its direct competitor.

Yea death to margins; something that is forbidden at modern day AMD.

 

[Heartbreaker]

The 7600 isn't exactly a Polaris analogue. Polaris was highly competitive at a reasonable price point and had more VRAM than its direct competitor. The 7600 is a disappointment, and, like the 4060, should be branded a 50 series card, IMO.

It's the x60 series competitor, just like Polaris was the 1060 competitor.

 

[Saylick]

That's GPUs in general now.
Xtors aren't getting cheaper and every % of perf on GPUs costs a lot of 'em.

Correct. I would not be surprised if most of the marketed performance gains from Nvidia in the next 5 years will be more from software features rather than raw hardware gains. More dedicated silicon and acceleration engines for their proprietary software, less emphasis on general compute units.

 

[maddie]

There are essentially four areas AMD has to do R&D for "graphics" compute: servers, consoles, iGPU and dGPU.

• Consoles are financed through big longterm corporate partners in Sony, Microsoft and potentially Valve and others.
• Servers is by far the biggest margin market available, the AI craze will let sales of non-CPU compute soar there.
• iGPU is a necessity in mobile and desktop chips nowadays, a requirement to get access to the steady market of enterprise and OEM sales.
• That leaves dGPU which for AMD so far is a high risk low reward market. How high the risk is RDNA3 is showing: Disaggregation of GPU dies, extensive reworking of the platform, with ambitious targets regarding performance and efficiency. Those targets were missed to a degree that RDNA3 barely makes a difference over RDNA2, meaning the investment in a whole dGPU gen is essentially being wasted there.

In which of these four areas would you invest your own money for further improvements?

These aren't exclusive. 1,3 &4 are almost the same thing with 2 also sharing tech to a lesser degree.

If you have to do 1,2&3, then 4 is not that an expensive proposition compared to only 4. Marginal costs.

It might mean slowing the multi core die strategy and the maximum size single core product. Means competing in smaller GPU core markets only. This drops your expenses a lot for the desktop card market and still allows you to service the largest sales segments.