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

 

[poke01]

If AMD wanted to really compete then they would have released RDNA 4, 6 months ago, at whatever price they could

I bet FSR4 wasn’t yet fully polished for release

 

[marees]

Status of the competition:

• good at $749
• not worth the money at $849
• a bad joke at $899 and above

https://twitter.com/x/status/1892221952409965008

In raster performance vs 4070 ti super:

• Alan Wake 2 11%,
• Black Myth Wukong 13%,
• Cyberpunk 2077 9%,
• Dying Light 2 16%,
• F1 24 3%,
• Horizon Forbidden West 7%,
• Indiana Jones and the Great Circle 6%,
• Kingdom Come Deliverance II 13%,
• Stalker 2 7%,
• Star Wars Outlaws 6%.

In Ray Tracing performance vs 4070 ti super:

• Alan Wake 2 13%,
• Black Myth Wukong 14%,
• Cyberpunk 2077 11%,
• Dying Light 2 15%,
• F1 24 5%,
• Indiana Jones and the Great Circle 13%,
• Star Wars Outlaws 4%

ASUS PRIME GeForce RTX 5070 Ti Video Card Review - Page 7 of 7

In our ASUS PRIME GeForce RTX 5070 Ti Review we'll explore performance vs RTX 4070 Ti SUPER, RTX 4070 Ti, RX 7900 XT, Raster, RT, DLSS, and Overclocking.

www.thefpsreview.com

If 9070xt matches a 4070 ti super on RT then AMD can go for MSFP of $625-$650 for MBA cards & $700+ for AIB cards

 

[gdansk]

software says no to early launches.

 

[marees]

What stops Jensen from launching 6 months ago as well?

Whatever is stopping them now from putting the 5070 on shelves right now

 

[Dezii90]

Whatever is stopping them now from putting the 5070 on shelves right now

Same reason why amd sandbagging 9070xt launch now. Depends on what the other person has

 

[Kepler_L2]

Do you know what a loss leader is? A loss leader is a product that loses money. AMD isn't losing money on every GPU sold.

Vega was sold at a loss.

 

[Thunder 57]

Vega was sold at a loss.

Because of HBM? Not being competitive? What about Polaris?

 

[adroc_thurston]

What about Polaris?

Also at a very slight loss for 8GB SKUs (but net profit for 4GB ones iirc).

Because of HBM? Not being competitive?

both, duh.

 

[beginner99]

It's the truth of how client discrete graphics works.
Either you win, or you don't.

it's not. halo doesn0t matter and if the 9070 XT would be $479 street price, they woul dfly of the shelves like no tomorrow. at $749 however, not so much. So price matters, halo matter jack shit.

 

[adroc_thurston]

halo doesn0t matter

Yeah it does.
Only thing that ever does.

and if the 9070 XT would be $479 street price, they woul dfly of the shelves like no tomorrow

It would force Nvidia price cuts and there you go.

 

[beginner99]

It would force Nvidia price cuts and there you go.

not if they lack supply. the low price must go with high supply.

 

[adroc_thurston]

not if they lack supply. the low price must go with high supply.

ramping the supply for non-HBM GPUs is beyond trivial.

 

[Dezii90]

They can go with safe option price where Nvidia doesn't care to discount = decent profits or they can milk short term since Nvidia cant reply, more profits, people disappointed or Discount heavily, go after marketshare, people happy, reviewers happy, but leaving profits on the table. Nvidia might take 2-3 months to reply with stock. I bet they can discount to a level where it can get them 25% marketshare and nvidia wouldnt care. My predication $629 9070xt, going for marketshare

 

[Timorous]

Making less profit is not losing money. Keeping a diversified product portfolio is important to a healthy business that can withstand market changes.

Given the low volumes of dGPU and the presence of console products, APUs, Strix Halo and the Instict accelerators I think AMD have a pretty diversified GPU landscape without really needing to care much about dGPU and the fact is of all the segments dGPU for gamers is not that profitable vs plenty of other products they currently sell.

AMD are also supply limited to a degree so it is not like dGPU is purely additive to everything they are doing. To really supply more than their current market share they need more supply from TSMC and if they are getting more supply from TSMC why not use that on other more profitable products than gaming dGPU that are also supply constrained like MI parts or 9800X3D etc.

Make the actual business case for it that would persuade Lisa to allocate the resource. I don't think it can be done tbh. Maybe AMD are better off creating some great new feature, shoving it into the next gen consoles and making it available for the PC ports so that NV are missing a feature. The problem is I bet if AMD did that they would be crucified just like they were when Starfield did not have DLSS because people thought AMD had told them not to include it when it seems it was just Bethesda being Bethesda and doing the minimum. So rather than whatever the feature is being a tick for AMD it would be a negative because they didn't share the toys with NV.

 

[PJVol]

AMD/ATI's main problem isn't halo/no halo or pricing, it's consistency. NV has had much more consistent execution from generation to generation. Every time AMD starts to gain a little momentum, they fall flat on their face.

Yeah, that's weird because they definitely have talented HW engineers, although things not always went smooth on the software side.

I'm not sure, but I thought I saw a link on this forum (please refer me, if you've seen it) to a blog post by a former AMD Radeon driver team member who was talking about the ATI legacy, and said what a nightmare it was to integrate the old display driver code. Something like a huge chunks of uncommented, "spaghetti" code )

 

[beginner99]

AMD/ATI's main problem isn't halo/no halo or pricing, it's consistency. NV has had much more consistent execution from generation to generation

Good point. AMD just does to much experiments with little to no gain. HBM and chiplets being the most recent 2 examples. HBM is gone, chiplets are gone, what was the point? a simple monolith with normal vram in both cases would have cost less R&D and performed more or less similar. in fact fury would have been better as not capped with 4 gb vram.

 

[Timorous]

Good point. AMD just does to much experiments with little to no gain. HBM and chiplets being the most recent 2 examples. HBM is gone, chiplets are gone, what was the point? a simple monolith with normal vram in both cases would have cost less R&D and performed more or less similar. in fact fury would have been better as not capped with 4 gb vram.

If chiplets can work, especially multiple compute chiplets, then you can scale without needing to manufacture huge dies that AMD is averse to building. Ultimately it will happen eventually and you need to start somewhere.

 

[yuri69]

If chiplets can work, especially multiple compute chiplets, then you can scale without needing to manufacture huge dies that AMD is averse to building. Ultimately it will happen eventually and you need to start somewhere.

This is a throwback to 2012 - exactly the same arguments were heard during the HD 7970 X2 reveal.

 

[PJVol]

If chiplets can work, especially multiple compute chiplets, then you can scale without needing to manufacture huge dies that AMD is averse to building.

From what I understand, the power cost of moving data between chiplets is still an issue for gaming-specific workloads.

 

[marees]

If chiplets can work, especially multiple compute chiplets, then you can scale without needing to manufacture huge dies that AMD is averse to building. Ultimately it will happen eventually and you need to start somewhere.

AMD thinks too far ahead without looking at execution practicalities

For ex:

1. Intel introduced 'multi-core' before AMD
2. Intel introduced APU with IGP before AMD
3. Packaging bugs & packaging capacity constraints have thrown a huge spanner in AMD's chiplet strategy

 

[poke01]

From what I understand, the power cost of moving data between chiplets is still an issue for gaming-specific workloads.

chiplets are just not it for gaming

 

[Timorous]

This is a throwback to 2012 - exactly the same arguments were heard during the HD 7970 X2 reveal.

Sure thing, and the point is still true. We see what it allows in the CPU space. Same is true of the GPU space and you can see that with MI300.

From what I understand, the power cost of moving data between chiplets is still an issue for gaming-specific workloads.

I am sure there are lots of hurdles to overcome but if you can make a 250mm part and then have it scale and 3 of them can compete with a 750mm part then you gain so much because you can compete in every tier while only putting the R&D into a single design. Once the hurdle is cleared it will be the way to go just like it is the way to go for CPUs.

AMD thinks too far ahead without looking at execution practicalities

For ex:

1. Intel introduced 'multi-core' before AMD
2. Intel introduced APU with IGP before AMD
3. Packaging bugs & packaging capacity constraints have thrown a huge spanner in AMD's chiplet strategy

AMD looking far ahead is what gave them Ryzen. To take market share from NV they need a silver bullet and being the 1st to fix the issues with chiplets and getting that actually working could be that silver bullet since it would be far more cost effective to build that super halo part than designing a 750mm die that they can only sell at consumer margins since NV have also got a huge advantage in the professional space with CUDA. The only way NV make it work is because they can put it into pro cards at huge margins and then some can go to a 5090.

 

[PJVol]

chiplets are just not it for gaming

Oh... didn't know that, thanks!

 

[beginner99]

If chiplets can work, especially multiple compute chiplets, then you can scale without needing to manufacture huge dies that AMD is averse to building. Ultimately it will happen eventually and you need to start somewhere.

but why experiment with expensive tech on cheap consumer cards? why not do that with the pro / datacenter cards? first and trickle down? nvidia is fine without needing this.

 

[madtronik]

Good point. AMD just does to much experiments with little to no gain. HBM and chiplets being the most recent 2 examples. HBM is gone, chiplets are gone, what was the point? a simple monolith with normal vram in both cases would have cost less R&D and performed more or less similar. in fact fury would have been better as not capped with 4 gb vram.

No gain? Where do you think the HBM and chiplets that MI300 abundantly uses come from? Oh! And X3D processors are just an offshoot of server experiments with stacked cache.