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

 

[tajoh111]

They're not short of cash.

It can be worthwhile, but it's also a very casino market and running expensive R&D programs to gamble is exactly the thing that almost killed AMD before.

AMD's cashflow is not particularly high at the moment(1.9 billion vs 25 billion in revenue) and something like 500 million in additional R and D quarterly would negatively affect their stock by turning their quarters into losses. Also their cash pile is not particularly large at the moment at 4.5 billion. So overspending as you mentioned could turn their fortunes into misfortunes quite quickly as you implied. Keeping their stock high is very important to AMD and this year has not been good to AMD in this regard(peak of 210 vs 125 now with AMD reaching a 52 week low recently).

At this point, AMD wants to prop up their stock price by making their financial look healthier(hence the layoffs and shuffling of resources). At this point AMD is likely saving any cash on hand for things like financing the purchase of an AI startup buyout or something like that.

Compare this to Nvidia's 55 billion and it's truly pretty poor gamble considering Nvidia's crazy cashflow at the moment which are among the best in the industry(70 billion in net profit this year vs 120 billion plus in revenue if you include the upcoming quarter).

AMD retreat from the highend is a sign of the times with how expensive it is to produce videocards in terms of cost to produce along with R and D expenditure. It's not a growing market either which does not really justify further investment. People expecting AMD to do what is best for them and not what is best for AMD just shows how delusional some buyers are.

Looking at this thread and the Intel battlemage thread, people expect companies without boatloads of money to sell cards at cost or a loss in the name of marketshare or bring up bogus data(TSMC is not giving Intel a 40% discount when their margins are 57% and their captital expenditure is 35 billion annually[more likely a rumor to help get Pat fired]) to create the illusion you can make relatively large dies graphic card at 250 dollars and still make a profit even though cost of production of silicon has quadrupled vs 10-12 years ago.

 

[adroc_thurston]

and something like 500 million in additional R and D quarterly would negatively affect their stock

A client GPU program (only part of it, they're making the IP anyway) does not cost 500mil quarterly.

Compare this to Nvidia's 55 billion and it's truly pretty poor gamble considering Nvidia's crazy cashflow at the moment which are among the best in the industry(70 billion in net profit this year vs 120 billion plus in revenue if you include the upcoming quarter).

That's a bubble thing.

At this point AMD is likely saving any cash on hand for things like financing the purchase of an AI startup buyout or something like that.

They don't need much more.

AMD retreat from the highend is a sign of the times with how expensive it is to produce videocards in terms of cost to produce along with R and D expenditure

no, they just don't like to gamble.

Keeping their stock high is very important to AMD and this year has not been good to AMD in this regard(peak of 210 vs 125 now with AMD reaching a 52 week low recently).

the market isn't rational so that's a wholly irrelevant metric.

 

[DaaQ]

AMD's cashflow is not particularly high at the moment(1.9 billion vs 25 billion in revenue) and something like 500 million in additional R and D quarterly would negatively affect their stock by turning their quarters into losses. Also their cash pile is not particularly large at the moment at 4.5 billion. So overspending as you mentioned could turn their fortunes into misfortunes quite quickly as you implied. Keeping their stock high is very important to AMD and this year has not been good to AMD in this regard(peak of 210 vs 125 now with AMD reaching a 52 week low recently).

At this point, AMD wants to prop up their stock price by making their financial look healthier(hence the layoffs and shuffling of resources). At this point AMD is likely saving any cash on hand for things like financing the purchase of an AI startup buyout or something like that.

Compare this to Nvidia's 55 billion and it's truly pretty poor gamble considering Nvidia's crazy cashflow at the moment which are among the best in the industry(70 billion in net profit this year vs 120 billion plus in revenue if you include the upcoming quarter).

AMD retreat from the highend is a sign of the times with how expensive it is to produce videocards in terms of cost to produce along with R and D expenditure. It's not a growing market either which does not really justify further investment. People expecting AMD to do what is best for them and not what is best for AMD just shows how delusional some buyers are.

Looking at this thread and the Intel battlemage thread, people expect companies without boatloads of money to sell cards at cost or a loss in the name of marketshare or bring up bogus data(TSMC is not giving Intel a 40% discount when their margins are 57% and their captital expenditure is 35 billion annually[more likely a rumor to help get Pat fired]) to create the illusion you can make relatively large dies graphic card at 250 dollars and still make a profit even though cost of production of silicon has quadrupled vs 10-12 years ago.

Have you not been parroting this since the interest/ loans were becoming due days? Back then it was CPU, now its GPU. It is tiring. I mean it is the same lengthy paragraphs of doom and gloom all the way to the OCN heydays.

If you are an insider, please keep your NDA intact, and spare us the analysis of your predictions from years ago rhetoric. Please.

It is obvious you are Intel/Nvidia through and through. Always some sort of AMD has done this, that, or the other wrong, in your opinion.

It is so tiring. Like 10 years tiring.

At least maybe just try and be somewhat constructive with your criticism from the past ten years. So it is TLDR.

 

[insertcarehere]

about Mi300

https://twitter.com/x/status/1871287937268383867

Easy to nod heads for an hour and then head back to wining and dining whomever, actually doing the hard work? From that article there's a legion of fundamental issues with their internal management that AMD needs to address to service this market in a manner that is satisfactory to the end customers there. Basic stuff like ensuring that ROCM is fully forward/backward compatible, or even doing matrix multiplication on a GPU without touching the graphics pipeline, y'know, pretty important for GPGPU work.

She should thank her stars that Intel is in such poor shape, because her seat should be plenty warm itself.

Have you not been parroting this since the interest/ loans were becoming due days? Back then it was CPU, now its GPU. It is tiring. I mean it is the same lengthy paragraphs of doom and gloom all the way to the OCN heydays.

If you are an insider, please keep your NDA intact, and spare us the analysis of your predictions from years ago rhetoric. Please.

It is obvious you are Intel/Nvidia through and through. Always some sort of AMD has done this, that, or the other wrong, in your opinion.

It is so tiring. Like 10 years tiring.

At least maybe just try and be somewhat constructive with your criticism from the past ten years. So it is TLDR.

That post certainly doesn't read like an Intel/NV homer, that last paragraph is anything but optimistic at Intel's GPU program moving forward.

Compare/contrast the market prominence of AMD/NV now versus the same 5 years ago is evidence enough of missed opportunities and bad execution on the former's part.

AMD retreat from the highend is a sign of the times with how expensive it is to produce videocards in terms of cost to produce along with R and D expenditure. It's not a growing market either which does not really justify further investment. People expecting AMD to do what is best for them and not what is best for AMD just shows how delusional some buyers are.

They haven't retreated from the high-end, the high-end has just become GPU-based accelerators. Of course there seems to be many things that need to be fixed for them to be taken seriously as a competitor...

 

[DaaQ]

Easy to nod heads for an hour and then head back to wining and dining whomever, actually doing the hard work? From that article there's a legion of fundamental issues with their internal management that AMD needs to address to service this market in a manner that is satisfactory to the end customers there. Basic stuff like ensuring that ROCM is fully forward/backward compatible, or even doing matrix multiplication on a GPU without touching the graphics pipeline, y'know, pretty important for GPGPU work.

She should thank her stars that Intel is in such poor shape, because her seat should be plenty warm itself.


That post certainly doesn't read like an Intel/NV homer, that last paragraph is anything but optimistic at Intel's GPU program moving forward.

Compare/contrast the market prominence of AMD/NV now versus the same 5 years ago is evidence enough of missed opportunities and bad execution on the former's part.



They haven't retreated from the high-end, the high-end has just become GPU-based accelerators. Of course there seems to be many things that need to be fixed for them to be taken seriously as a competitor...

If you have been on OCN, then you should realize what I am alluding to here. It has been a constant berating of one companies practices above all others.

 

[insertcarehere]

If you have been on OCN, then you should realize what I am alluding to here. It has been a constant berating of one companies practices above all others.

I don't give a damn about what was or wasn't said on OCN. AMD is all but exiting the consumer GPU space except for a token presence and is miles behind in the datacenter GPU space as well (arguably more so than consumer). It deserves plenty of beratement for its practices over the last couple of years.

 

[marees]

Taking RDNA 1 as a launch template (with 50% cumulative inflation):

1. $400 5700xt — 9070xt 16gb 64 CU ($600)
2. $350 5700 — 9070 16gb 56 CU ($525)
3. $280 5600xt — 9060xt 12gb 48 CU ($420)
4. $200 5500xt 8gb — 9050 xt 16gb 32 CU ($350?)
5. $170 5500xt 4gb — 9050 xt 8gb 32 CU ($300??)
6. OEM 5500 — 9040 xt 8gb? 28 CU ($250???)

 

[Kepler_L2]

Taking RDNA 1 as a launch template (with 50% cumulative inflation):

1. $400 5700xt — 9070xt 16gb 64 CU ($600)
2. $350 5700 — 9070 16gb 56 CU ($525)
3. $280 5600xt — 9060xt 12gb 48 CU ($420)
4. $200 5500xt 8gb — 9050 xt 16gb 32 CU ($350?)
5. $170 5500xt 4gb — 9050 xt 8gb 32 CU ($300??)
6. OEM 5500 — 9040 xt 8gb? 28 CU ($250???)

9070 XT would be beyond DOA at $600. Even at $500 it's a tough sell vs $600 5070.

 

[adroc_thurston]

Even at $500 it's a tough sell vs $600 5070.

Ehhh, the shill factor gap will be smaller than ever. Unless Battaglia somehow persuades the kids into doing WSL2 CUDA installs.

 

[SolidQ]

Unless Battaglia somehow persuades the kids into doing

Don't worry, he will found few pixels advantage.
But for real, if RT gap will be still big, then nothing to help AMD

 

[adroc_thurston]

But for real, if RT gap will be still big, then nothing to help AMD

RT isn't super-relevant, but the kids crave ML upscaler nuggjes and they'll get some.

 

[deasd]

Just for fun and maybe off topic. I almost forget and barely remember the account All_the_watts deleted himself before.

Phoenix leaks got removed altogether with All_The_Watts twitter account deletion.

https://twitter.com/All_The_Watts

I don't delete my account just for testing but can the deleted account be found back? IIRC so many old post during 2022 got removed.

 

[SolidQ]

RT isn't super-relevant

more and more games come with RT, so became more relevant.

kids crave ML upscaler nuggjes and they'll get some.

if it's beat Xess then fine, if not big fail

All_the_watts deleted himself before.

That was other acc.
Anyway interesting he's delete post about perfomance/price, maybe in real games is different story?

 

[poke01]

more and more games come with RT, so became more relevant.

Yep it is relevant to both MS and Sony. Maybe not to the people here but it is to both console makers.

if it's beat Xess then fine, if not big fail

Lets see it, FSR4 should be more common than XeSS.

 

[SolidQ]

Maybe not to the people here but it is to both console makers.

on gaming forums a lot people care about RT. Here is different story

FSR4 should be more common than XeSS.

Yeah, but interesting one moment will it work on other vendor GPU's or it's like Xess with D4pa

Yep it is relevant to both MS and Sony.

That true Sony talking about it, but interesting one moment will PS6 beat six year card like 4090 in perfomance or gonna be big fail.
PS5 beat 1080ti, which 3 year difference

 

[adroc_thurston]

if it's beat Xess then fine, if not big fail

It doesn't have to, it just gotta have "AI" in the name.

but it is to both console makers.

It's so relevant MS hasn't made any DXR changes in 6 years. Be serious.

Yeah, but interesting one moment will it work on other vendor GPU's or it's like Xess with D4pa

Unlikely. For cheaper stuff there's FSR2.

 

[SolidQ]

Unlikely. For cheaper stuff there's FSR2.

So seems users with RX 6xxx are screwed

It doesn't have to, it just gotta have "AI" in the name.

but need, otherwise every reviewers will speak about bad quality

 

[adroc_thurston]

So seems users with RX 6xxx are screwed

Yea.

but need, otherwise every reviewers will speak about bad quality

ML upscaling doesn't struggle with quality usually. It's perf that might be poo.

 

[SolidQ]

ML upscaling doesn't struggle with quality usually

PSSR struggle, but one moment it's devs trouble or pssr, still hard to say

 

[GodisanAtheist]

Anyone saying RT/Upscaling/Frame Gen don't matter at this stage in the game with regards to the consumer DGPU market has completely lost the plot.

Again, it's not about whether you care for the tech, the public has spoken and it matters.

Nvidia is basically a raging inferno creating its own weather at this point...

 

[adroc_thurston]

Anyone saying RT/Upscaling/Frame Gen don't matter at this stage in the game with regards to the consumer DGPU market has completely lost the plot.

Again, it's not about whether you care for the tech, the public has spoken and it matters.

The shill points would move elsewhere.
Remember when everyone was a streamer and NVENC was the #1 thing to look out for?

Nvidia is basically a raging inferno creating its own weather at this point...

Bingo! But they've been really, really good at it for 25 years.

 

[SolidQ]

Bingo! But they've been really, really good at it for 25 years.

Interesting AMD software team is old? Seems they out of ideas, they just mostly follow the NV

NVENC was the #1 thing to look out for?

Still there for Twitch. Twitch are allergic to move for AV1

The shill points would move elsewhere.

That what AMD team need aggresive marketing.
As we see AMD got with RDNA3 some market share in 2023 and without HALO, they just fallen to DLSS/RT and Super series

 

[adroc_thurston]

Seems they out of ideas, they just mostly follow the NV

They always did that. They're brutalism in hweng form.
No frills. Just the basics. From the very R300 that more or less started it all.

Still there for Twitch. Twitch are allergic to move for AV1

Obviously but no one really gives a damn about the encoder block anymore since DLSS™ is a thing to look out for.

 

[poke01]

It's so relevant MS hasn't made any DXR changes in 6 years. Be serious.

MS will follow Sony, next gen xbox will copy PS6 in terms of GPU/AI feature sets.

6 years ago RT/AI upscaling was barely relevant and it wasn't

NVENC was the #1 thing to look out for?

quick sync is much better than NVENC.

The shill points would move elsewhere.

That’s the point, make it so good the shills can’t move the goal posts. For that AMD needs to think outside the box and stop following Nvidia.

simply put the Radeon teams need to be more like Ryzen teams. Be the leader instead of following the carrot.

 

[marees]

Interesting AMD software team is old? Seems they out of ideas, they just mostly follow the NV

When NV was releasing CUDA, AMD was acquiring ATI (& planning APU fusion etc.)

Check this thread:

https://twitter.com/x/status/1871367209060319557