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

 

[beginner99]

Performance gap won't really be much.

Why? Slower RAM, less CUs, lower clocks. so could easily be a 15% gap

Then another 15% gap to the 9060 XT. and since N44 is pretty much half of N48, it would add up nicely to N44 based cards.

9070 XT	100 %
9070	85 %
9060 XT	70 %
"9060" (full n44)	50 %

 

[Keller_TT]

I mean personally I'm all for that being $500 but no, maybe the non-xt but certainly not the XT IF the performance rumors are true and performance is higher than expected at 5070 Ti levels.

Should AMD take their own ideal execution and roadmap as guiding standard or should they keep pricing based on Jensen's greed?
If AMD have to be somebody in the dGPU market, they should bring that wow factor to mid-range and bring an AMD identity - mindshare, and market share will follow.

Before 5070 Ti, they have to think about the supposed 5060 Ti with 16GB and NV not making 4060 Ti performance mistake again.

 

[del42sa]

https://videocardz.com/newz/amd-rad...ice-similar-to-the-rx-7800-xt-and-rx-7900-gre

some discussion with Azor about RDNA4

 

[Gideon]

The fact that this is named 70 series & 80 series makes me think that MBA boards won't go above $550
The non-xt is likely to come slightly under 7800xt at $480

Since board partner cards have 3x 8-pin they could be priced at $600

As for the further cut down 12gb card, it is likely to be launched very late for $380

I'm also pretty sure that unlike the top model, where they will go all out, they will heavily power limit the 16GB non-XT (just like they did with RX 6800)

If AMD have to be somebody in the dGPU market, they should bring that wow factor to mid-range and bring an AMD identity - mindshare, and market share will follow.

Well they also can't overdo it, or Nvidia will just adjust the pricing and midrange cards before release (as they did wit RTX 4080 12GB)

 

[SolidQ]

I'm also pretty sure that unlike the top model, where they will go all out, they will heavily power limit the 16GB non-XT

I think it's similar to 5700 with less CU/5700XT

 

[jpiniero]

(as they did wit RTX 4080 12GB)

They did, if you think about it. The 5070 is the same price as the 4070's price cut and the 5070 Ti is cheaper than the 4070 Ti Super.

 

[Gideon]

They did, if you think about it. The 5070 is the same price as the 4070's price cut and the 5070 Ti is cheaper than the 4070 Ti Super.

Yeah but it doesn't stop them from doing it again, if AMD does something crazy like offering 5070 performance at $400 and 5070 Ti at $550. If it's too late to change the initial products they can just ride it out for a few quarters and then introduce a 5070 SUPER with 16GB or something along those lines

 

[Keller_TT]

I think it's similar to 5700 with less CU/5700XT

AMD's target is moving on from 7700 XT and 7800 XT/7900 GRE with the 9070 series. A well done RDNA4 would mean 9070 shading the 7900 GRE, and 9070 XT likewise w.r.t 7900 XT.
My prior estimates tally with this.
220W 9070, 265W 9070XT MBA. If so, $400 & 500 or $450, $550?

AIBs could shape v/f up to 310W for the XT and go +$100 over MBA.

 

[SolidQ]

$450, $550?

Unless AMD will stay with same practice. like 6800/6800XT, 7700Xt/7800XT, 7900XT/7900XTx to upselll 9700XT
Still possible like 450-500


Wccftech

AMD Plans To "Disrupt" The Mainstream GPU Markets With RDNA 4 SKUs, Claiming That They'll Be Available In The $300-$1,000 Range​

Then they quote AMD
AMD remains focused on delivering value for money. When we announce RDNA4, we’ll introduce a powerful graphics card—not a $300 card, but also not a $1,000 card.

 

[blckgrffn]

They did, if you think about it. The 5070 is the same price as the 4070's price cut and the 5070 Ti is cheaper than the 4070 Ti Super.

I think this is a good observation and says a lot, actually, about how Nvidia seeing their own launch lining up. The 5070 at $550 is not some sort of gift (lmao) nor the 5080 at $1k, etc.

They are admitting right to our face they expect it to be somewhat iterative as a launch. This doesn’t seem to be what AMD was expecting but 🤷‍♂️

The last launch was epic partly because they moved from a mid process node at Samsung to the mature node at TSMC. Without that, it seems like a 1080ti to 2080ti type move was at least a possibility?

 

[Keller_TT]

Unless AMD will stay with same practice. like 6800/6800XT, 7700Xt/7800XT, 7900XT/7900XTx to upselll 9700XT
Still possible like 450-500

The 9070 vs 4070S would be like 7800 XT vs 6800 XT I guess. So, that might prevent them doing silly upsell tactics. I think the 5070 will slot in for 4070 Ti.

9070 with +4 GB VRAM vs 4070S, if it launches for $429 max, AMD will have two winners on their hands.

 

[Mahboi]

I've been out of tech for a fair bit since I'm writing a book, only poking my head really quickly...have our leaker boys gotten fooled? The official news online is nothing like the leaks I remember from 6 months ago.
Is it time to try to sell away my 7900 xt lol?

 

[Keller_TT]

The 9070 vs 4070S would be like 7800 XT vs 6800 XT I guess. So, that might prevent them doing silly upsell tactics. I think the 5070 will slot in for 4070 Ti.

9070 with +4 GB VRAM vs 4070S, if it launches for $429 max, AMD will have two winners on their hands.

Might be getting ahead of myself here. If NV don't drop the ball this time with 5060 Ti (supposedly with 16 GB), that could be a 4070 raster and ~ 4070S RT for $450.
So, interesting how the 9070 will be priced.

 

[blckgrffn]

I've been out of tech for a fair bit since I'm writing a book, only poking my head really quickly...have our leaker boys gotten fooled? The official news online is nothing like the leaks I remember from 6 months ago.
Is it time to try to sell away my 7900 xt lol?

Hope the book is going well!

7900XT is probably safe until next gen atm. You might net out like $100 and get fancy FSR for your troubles?

I'd just keep enjoying what you have.

I still might consider returning my XTX but that's a different road than selling. As in I just walk it into the local UPS Store and I kept the boxes and all that. 20 minutes of work.

Played some more games on it and I do like and appreciate how much faster it is at 3440*1440. It's a big lift from the 6800 that was clearly more comfortable at 2560*1440. That extra bit pushed it below 60 fps frequently at settings I liked on HD2 (surprisingly heavy game, probably an optimization thing) and even Wonderlands which is crazy smooth now.

I upgraded my whole rig for Borderlands 3 so I guess this is no surprise. Hopefully I am set for BL4!

 

[Ranulf]

Helldivers2 is one of those games that really seems to push hardware limits for whatever reason. It definately pushes the 4790k hard which is the min cpu for it.

 

[Mahboi]

Very nice, thanks for the complete writeup!
Yeah I'm only mildly gaming, the real reason to change would be for less wattage, and frankly, all I'd probably gain is ~250€ tops, over years of usage, and that's including watts & profit from the sale. I'll just stick to mine, plus after doing a little bit of AI with Amuse, I gotta say that this 20 Go VRAM buffer has some serious merits.

The book is going insanely well but unfortunately I am also going insane writing it, but that's kind of getting what I deserve for trying to write a full history of the USA.

 

[gaav87]

@adroc_thurston Hey adroc is no dual issue but super wide theoretically possible/viable ?

 

[adroc_thurston]

Hey adroc is no dual issue but super wide theoretically possible/viable ?

you're focusing way way too much on flops. Do not.

So, interesting how the 9070 will be priced.

It's not even the same perf class.

 

[Saylick]

you're focusing way way too much on flops. Do not.

Cache Rules Everything Around Me.

 

[reaperrr3]

If the two models are 16 GB, then the rumors of the 9700 Non XT being only a very slight cut are true. Which is dumb, but that's AMD for ya.

Well, the 7700XT was also only a slight cut, actually only 10% cheaper despite being ~20% slower.

But the 7700XT was in large part hampered by lower IF$- and mem bandwidth. The 9070 won't have that issue, so I expect the 9070 to be a bit closer to the XT than the 77XT was to 78XT.

Might be getting ahead of myself here. If NV don't drop the ball this time with 5060 Ti (supposedly with 16 GB), that could be a 4070 raster and ~ 4070S RT for $450.

Gonna be tough.
5070 at least has (presumably) the same amount of L2 and ROPs as the 4070Ti, plus higher mem bandwidth and a significant TDP increase vs. the 4070, so some of the SM deficit vs. the 4070Ti is probably offset by the higher bandwidth even before IPC and clocks come into play.

The 5060Ti on the other hand will have less ROPs, less L2 and less VRAM bandwidth than a 4070, and at least 10 SM less (as long as it remains GB206-based).
It's also rumored to get only 5W more TDP than the 4060Ti-16GB, which doesn't indicate much of a clockspeed increase, either.
Might be enough to catch the bandwidth-starved 7700XT in raster, but 4070/78XT will be tough.

From what we know so far, I'd expect the 9070 to be farther ahead of the 5060Ti than the 77XT was ahead of the 4060Ti-16G.

Unless Nvidia changes their original plan and makes the 5060Ti GB205-based with beefed-up specs, after all.

 

[MrTeal]

Well, the 7700XT was also only a slight cut, actually only 10% cheaper despite being ~20% slower.

But the 7700XT was in large part hampered by lower IF$- and mem bandwidth. The 9070 won't have that issue, so I expect the 9070 to be a bit closer to the XT than the 77XT was to 78XT.

Gonna be tough.
5070 at least has (presumably) the same amount of L2 and ROPs as the 4070Ti, plus higher mem bandwidth and a significant TDP increase vs. the 4070, so some of the SM deficit vs. the 4070Ti is probably offset by the higher bandwidth even before IPC and clocks come into play.

The 5060Ti on the other hand will have less ROPs, less L2 and less VRAM bandwidth than a 4070, and at least 10 SM less (as long as it remains GB206-based).
It's also rumored to get only 5W more TDP than the 4060Ti-16GB, which doesn't indicate much of a clockspeed increase, either.
Might be enough to catch the bandwidth-starved 7700XT in raster, but 4070/78XT will be tough.

From what we know so far, I'd expect the 9070 to be farther ahead of the 5060Ti than the 77XT was ahead of the 4060Ti-16G.

Unless Nvidia changes their original plan and makes the 5060Ti GB205-based with beefed-up specs, after all.

There's definitely room in the stack for a cut die to be a 5060 Ti, and it would solve the problem of weird VRAM amount by letting them use 12GB without having to wait for 3GB chips. It'd also give the 5060 Ti a bit of a reason to exist; the 4060 Ti was so hamstrung by low SM count it was barely faster than a 3060 Ti. GB206 5060 Ti would be two generations of very little performance increase in that class.
If 9070 is 16GB and say 54CU out of the 64, I'd agree the lead would grow. How does AMD price it though? The 7700XT is available in the US for the same price as the 8GB 4060 Ti with 20% better performance and 50% more VRAM, and still probably sells quite a bit worse. If they both launch at $400, even with a greater raster performance lead will it sell any better vs a 16GB 5060 Ti?

 

[soresu]

View attachment 114671
Cache Rules Everything Around Me.

White Room intensifies....

 

[insertcarehere]

Well, the 7700XT was also only a slight cut, actually only 10% cheaper despite being ~20% slower.

But the 7700XT was in large part hampered by lower IF$- and mem bandwidth. The 9070 won't have that issue, so I expect the 9070 to be a bit closer to the XT than the 77XT was to 78XT.

Gonna be tough.
5070 at least has (presumably) the same amount of L2 and ROPs as the 4070Ti, plus higher mem bandwidth and a significant TDP increase vs. the 4070, so some of the SM deficit vs. the 4070Ti is probably offset by the higher bandwidth even before IPC and clocks come into play.

Well the GPU-Z screenshot we have is showing 256bit GDDR6 at 20Gbps. To even get something with that backend to perform remotely like a 4080/XTX will require a ton of cache. Which is where the cuts can occur - 4070tiS had similarly mediocre cuts vs 4080 in terms of CUs/mem and it turned out that perf difference was bigger than expected for that reason as well. I presume any bin from 9070XT - > 9070 will be looking at the same.

 

[blckgrffn]

Well the GPU-Z screenshot we have is showing 256bit GDDR6 at 20Gbps. To even get something with that backend to perform remotely like a 4080/XTX will require a ton of cache. Which is where the cuts can occur - 4070tiS had similarly mediocre cuts vs 4080 in terms of CUs/mem and it turned out that perf difference was bigger than expected for that reason as well. I presume any bin from 9070XT - > 9070 will be looking at the same.

128MB IC back in play maybe? That’s bigger than any RDNA 3 implementation and maybe it’s “better” in some way, or the lower level caches grew a lot?

 

[SolidQ]

That’s bigger than any RDNA 3 implementation and maybe it’s “better” in some way, or the lower level caches grew a lot?

That what we don't know. Let's wait event