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

 

[insertcarehere]

No it's a case of ATi doing a market suicide in 2008.
They should've never, EVER gone with the small die strategy. They handed NV a free win.
Gave them a 9-year gap between R600 and Fiji to do whatever they wanted, and by the time Fiji came, it was already too late.
Franky even after that they didn't try.

Who cares, PPAmaxxing is all that matters.

Lol prioritizing maximum PPA as a strategy was the exact reason for RV770's inception in the first place - i.e the same small die strategy that as per your claim, backfired on AMD hard.

Sticking to that is only going to lead to more 'RV770s' (basically RDNA 1-4), which had good 'value' but:
- Behind in feature sets (Ray-tracing/up-scaling for consumer and god knows what for professional) that AMD management thought didn't matter - turns out they do and are a differentiating factor for customers purchase decisions.
- As a result of the above, being 2nd place in feature sets means dealing with software developers (game developers/pros/institutions) that are used to working with NV's hardware stack for said features first -> more touchpoints for NV to work with software devs, fast forward that for a few years and say hello to the CUDA moat.
- Constantly behind in performance regardless, which does wonders for brand equity, sure hope there's no third player willing to (at least temporarily) operate off lower margins to grab marketshare!

An AMD management that is not willing to unstick its head up you know what and change up its strategy on the GPU side is only destined to be the 'value' option at best (i.e give away the high margin segments), if not become vulnerable to new disruptors in the market.

 

[adroc_thurston]

Lol prioritizing maximum PPA as a strategy was the exact reason for RV770's inception in the first place - i.e the same small die strategy that as per your claim, backfired on AMD hard.

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.

- Behind in feature sets (Ray-tracing/up-scaling for consumer and god knows what for professional) that AMD management thought didn't matter - turns out they do and are a differentiating factor for customers purchase decisions.

Irrelevant bling.

- As a result of the above, being 2nd place in feature sets means dealing with software developers (game developers/pros/institutions) that are used to working with NV's hardware stack for said features first -> more touchpoints for NV to work with software devs, fast forward that for a few years and say hello to the CUDA moat.

IHVs do not touch gamedevs directly, it's all on MS/Khronos/etc to define core API featuresets (for better or for worse).

- Constantly behind in performance regardless, which does wonders for brand equity, sure hope there's no third player willing to (at least temporarily) operate off lower margins to grab marketshare!

There can be only two.

An AMD management that is not willing to unstick its head up you know what and change up its strategy on the GPU side is only destined to be the 'value' option at best (i.e give away the high margin segments), if not become vulnerable to new disruptors in the market.

their strategy is to ignore client dgfx exists.
2-3 dies keep tokenish presence which is enough.

 

[insertcarehere]

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.

The inherent philosophy of priortizing PPA means targeting sweet spots, which necessarily means trading off absolute performance and features. NV's biggest dies basically never have the best PPA out of all their lineups, for example.

NV's direction with its GPUs stopped being about prioritizing PPA from c. 10 years ago (about when Turing is getting through its design stages).

Irrelevant bling.

For you maybe, market share trends over the last half decade say otherwise for the marketplace.

Sony think this was relevant enough to put work implementing and marketing a semi-custom solution for their PS5 Pro.

IHVs do not touch gamedevs directly, it's all on MS/Khronos/etc to define core API featuresets (for better or for worse).

Yes being first to market with certain features has no influence whatsoever on what becomes future 'core' features down the line, or indeed how they are implemented, got it.

There can be only two.

Please explain why this market is one in which there can only be a natural duopoly composed specifically of Nvidia/AMD.

their strategy is to ignore client dgfx exists.
2-3 dies keep tokenish presence which is enough.

Doing real well competing on the professional side with 1/16th the market cap and 1/32nd the net earnings (majority of that being CPUs), great strategy right there.

 

[adroc_thurston]

The inherent philosophy of priortizing PPA means targeting sweet spots

No, it means maximing the PPA.

NV's direction with its GPUs stopped being about prioritizing PPA from c. 10 years ago (about when Turing is getting through its design stages).

Turing was the only outlier, everything after went back to PPAmaxing.

For you maybe, market share trends over the last half decade say otherwise for the marketplace.

Market share is trending towards NV victory since 2006.

Yes being first to market with certain features has no influence whatsoever on what becomes future 'core' features down the line, or indeed how they are implemented, got it.

Yea?
MS always does its own thing, which is how we got crimes against humanity known as Geometry Shaders in dx10.

Please explain why this market is one in which there can only be a natural duopoly composed specifically of Nvidia/AMD.

Everyone else sucks at making usable GPU IP.

Doing real well competing on the professional side with 1/16th the market cap and 1/32nd the net earnings (majority of that being CPUs), great strategy right there.

Yeah, they're making real cash in there and forcing NV to aggressively defend their MSS.

 

[SteinFG]

Starting a new naming scheme with 9 at the begining is really funny. Not even AMD believes in itself to carry it for more than 1 gen.
Next one is probably gonna be called Radeon Ultra 6070 XT (to be on par with Nvidia's 6000 series).

 

[marees]

Next one is probably gonna be called Radeon Ultra 6070 XT (to be on par with Nvidia's 6000 series).

Or probably Radeon 185 AI Maxx Ultra pro etc.

 

[HurleyBird]

No it's a case of ATi doing a market suicide in 2008.
They should've never, EVER gone with the small die strategy. They handed NV a free win.

And at that point Nvidia didn't have the unshakeable mindshare they do now. Before G80 it was back-and-forth and people still remembered the 9700 Pro fondly.

ATi had to come back with a vengeance after R600. They had an amazing architecture relative to Nvidia's, yet didn't do anything with it.

By far the most maddening part of that whole fiasco was how smug and self congratulatory AMD acted over the sweet spot strategy which was actually their doom.

 

[gaav87]

Predicted a 20% improvement over at 7800xt last year.

I was the only one here that predicted RDNA3 performance remotely correct last time around.

Those timespy results are indicating I could be right again. Lets see.

Nah i also predicted rdna3 +-5% and was spot on rdna4 6month's ago

 

[poke01]

And at that point Nvidia didn't have the unshakeable mindshare they do now. Before G80 it was back-and-forth and people still remembered the 9700 Pro fondly.

ATi had to come back with a vengeance after R600. They had an amazing architecture relative to Nvidia's, yet didn't do anything with it.

By far the most maddening part of that whole fiasco was how smug and self congratulatory AMD acted over the sweet spot strategy which was actually their doom.

NV having full nearly full mindshare is on AMD. Its self inflicted.

Sony think this was relevant enough to put work implementing and marketing a semi-custom solution for their PS5 Pro.

Adroc doesn't like things AMD is behind in. Sony has to market full RT and AI for PS6 so AMD has no choice but to design and deliver with RDNA5/UDNA.

 

[adroc_thurston]

Adroc doesn't like things AMD is behind in.

They're just not very relevant.
Intel has the stuff but it's not gonna help them an iota.

Sony has to market full RT and AI for PS6 so AMD has no choice but to design and deliver with RDNA5/UDNA.

They can always do custom.
But RDNA5 will have matrix slop either way due to MS.

 

[poke01]

https://videocardz.com/newz/leaked-amd-gpu-render-confirmed-by-hardware-insider-as-radeon-rx-9070-xt

Its a nice design

 

[ToTTenTranz]

Why not just get a 7900xtx?

I'm not buying a 2 year-old consumer GPU. It's just not a good choice.

It's supposedly about equal to the N48 in RT, but still faster by a margin in raster, and would cost you much less than 1500€.

N48 should perform above the 7900XTX in heavy RT scenarios, and FSR4 should run a bit faster in RDNA4 than in RDNA3 as matrix throughput should still be higher in N48.

No it works pretty well already but there are still many things to be done wrt training.

Then it doesn't work pretty well...

Everyone else sucks at making usable GPU IP.

Depends on power target. AMD and Nvidia seem to be behind in GPU IP for ULP, where Qualcomm and Apple (PowerVR?) reign.

 

[Joe NYC]

Of course it's a monodie. People who can make tiled GPUs work are busy with other things.

Just to keep things in perspective, NVidia is reports $3.3 billion in gaming revenue, which is nearly as much as AMD EPYC and Instinct combined - $3.5 billion. Each one - EPYC and Instinct is about half of NVidia gaming revenue.

So, this market is not irrelevant. AMD would benefit tremendously from taking a bigger slice of this market, at reasonable margins.

 

[gaav87]

6800xt/6950xt/7700xt/7800xt upgrade path -> 9070xt +20% 599$ or 5070ti +50% 899$ xD
The more you buy the less they give.

 

[adroc_thurston]

Then it doesn't work pretty well...

Not the target market for it.

AMD and Nvidia seem to be behind in GPU IP for ULP, where Qualcomm and Apple (PowerVR?) reign.

Sort of? RDNA3.something in e2400 is very competent despite an inferior node.
QC and AAPL GPU IP is more big boy than it used to be but it still has issues.

AMD would benefit tremendously from taking a bigger slice of this market, at reasonable margins

High opportunity cost for incredibly pitiful ROI.
Just don't (and AMD isn't).

 

[ToTTenTranz]

High opportunity cost for incredibly pitiful ROI.
Just don't (and AMD isn't).

Then why did AMD's Jack Huynh say the exact opposite just 3 months ago?

AMD deprioritizing flagship gaming GPUs: Jack Huynh talks new strategy against Nvidia in gaming market

The battle seems to be over before it starts.

www.tomshardware.com

, I pressed Huynh for information regarding the company's plans for the high-end GPU market with the RDNA 4-powered Radeon RX 8000-series. His comments sketch out a plan focused specifically on gaining market share in the GPU market above all else

TH: Price point-wise, you have leadership, but you won't go after the flagship market?

JH: One day, we may. But my priority right now is to build scale for AMD. Because without scale right now, I can't get the developers. If I tell developers, ‘I’m just going for 10 percent of the market share,’ they just say, ‘Jack, I wish you well, but we have to go with Nvidia.’ So, I have to show them a plan that says, 'Hey, we can get to 40% market share with this strategy.' Then they say, 'I’m with you now, Jack. Now I’ll optimize on AMD.' Once we get that, then we can go after the top.

TH: This is specifically a client strategy [consumer market]?

JH: This is a client strategy.

 

[adroc_thurston]

Then why did AMD's Jack Huynh say the exact opposite just 3 months ago?

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).

Like they weren't making N4c's for kicks. You need the chainsaw. They know they need the chainsaw. But they can't make one anymore.

 

[itsmydamnation]

Then why did AMD's Jack Huynh say the exact opposite just 3 months ago?

AMD deprioritizing flagship gaming GPUs: Jack Huynh talks new strategy against Nvidia in gaming market

The battle seems to be over before it starts.

www.tomshardware.com

because words cost nothing......

 

[linkgoron]

Then why did AMD's Jack Huynh say the exact opposite just 3 months ago?

AMD deprioritizing flagship gaming GPUs: Jack Huynh talks new strategy against Nvidia in gaming market

The battle seems to be over before it starts.

www.tomshardware.com

Because he won't flat-out say that AMD is getting completely outplayed by Nvidia and that AMD just couldn't deliver and that the chiplet strategy has been a disaster for AMD's client GPUs. This is similar to what happened with Polaris, with almost exactly the same things being said at the time. There's no way getting killed in the midrange by Intel from below and Nvidia from above is somehow a viable strategy for AMD.

I'm not buying a 2 year-old consumer GPU. It's just not a good choice.

You already just wanted a "stop gap" card anyway. Just buy a 5070ti or 5080 and call it a day. Why live with inferior performance? I hope that AMD will deliver with UDNA/RDNA5/whatever but given RDNA3 and RDNA4 I wouldn't plan for a great product or a competent launch, and you might just end up with Nvidia anyway.

N48 should perform above the 7900XTX in heavy RT scenarios, and FSR4 should run a bit faster in RDNA4 than in RDNA3 as matrix throughput should still be higher in N48.

I suggest we wait and see. RDNA4's raster performance is already under what you were expecting, let's see what happens with RT.

 

[vanplayer]

N48 should perform above the 7900XTX in heavy RT scenarios, and FSR4 should run a bit faster in RDNA4 than in RDNA3 as matrix throughput should still be higher in N48.

It depends on model. The gap between reference & non-reference card is bigger than previous gens.

 

[poke01]

that's a meme. Neither is an s/w company.

Nvidia has more software eng than hardware ones

 

[CouncilorIrissa]

https://twitter.com/x/status/1871185600310899030

SKU	avg score @ Port Royal
4070 Ti	13978
7900 XT	14186
4070 Ti Super	15716

 

[SolidQ]

avg score @ Port Royal

big problem it's just simple RT bench, need test overdrive mode in C2077 and other heavy scenarios where 4070 easily beat 7900XT

 

[gaav87]

Well guys time to pack your bags. 1.39x over stock 6800xt in RT and only 1.08x over in raster vs my stock 6800xt at 2350mhz and 240w ppt uv + 9800x3d

 

[SolidQ]

1.39x over stock 6800xt in RT

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

Spoiler