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

 

[Win2012R2]

That’s not ever changing

Not with the current defeatist attitude, that's for sure.

Maybe they should get onto Kickstarter to finance proper YUGE chip that will be guarantee to beat Nvidia's best by at least 30%, get bonus if it's 50%.

 

[marees]

Not with the current defeatist attitude, that's for sure.

Maybe they should get onto Kickstarter to finance proper YUGE chip that will be guarantee to beat Nvidia's best by at least 30%, get bonus if it's 50%.

I am not sure that focussing on client (gaming) gpu is a good idea. Broadcom has 5x the market capitalization of AMD. If AMD is not careful then Broadcom will eat their lunch.

 

[SolidQ]

I am not sure that focussing on client (gaming) gpu is a good idea.

But they can try with future UDNA create gaming an AI chip, if gamers not going too much buy, then AI companies gladly will buy.
Just need create:
good media engine
good upscaling
strong RT/PT
fix softwares like Blender/other 3d applications and Adobe etc
make strong RoCm and AI
after that more and more people will hear positive about Radeon's, like it's was with Ryzen, now Ryzen gaining market share

 

[adroc_thurston]

future UDNA

No such thing.

after that more and more people will hear positive about Radeon's, like it's was with Ryzen, now Ryzen gaining market share

The only way to get that is the chainsaw. Only the chainsaw matters.
The shill points ebb and flow, ever changing. The chainsaw is eternal.

 

[SolidQ]

The shill points ebb and flow, ever changing.

shill, not shill, they need cross own red line and create, what i'm above write. Youtube bloggers will speak about positives. That how marketing working. People hear positive, people looking into product.
I know people in real life, that was negative about AMD, but hearing Ryzen is good, then change opinions and buying it.

Mostly now when reading different forums/real life people talking about like it
AMD can't do RT
AMD dont have upscaling
AMD driver is bad
AMD media quality is bad
AMD in 3D applicaitons like Blender 5-7 times slower, Adobe crash
only mostly negative, need turn it into positive

 

[adroc_thurston]

shill, not shill, they need cross own red line and create, what i'm above write. Youtube bloggers will speak about positives. That how marketing working. People hear positive, people looking into product.

Only the Chainsaw can be positive. No amount of good enough poverty parts can revive the brand.

but hearing Ryzen is good

Ryzen IS the Chainsaw. They ship really cool halo parts years before Intel can respond because it's an offshoot of an even more cool and halo server roadmap.
Ffs there's almost a 5y gap between AMD and Intel shipping large LLC client parts.
No such thing is possible in client dgfx with reasonable budget.

AMD can't do RT
AMD dont have upscaling
AMD driver is bad
AMD media quality is bad
AMD in 3D applicaitons like Blender 5-7 times slower, Adobe crash
only mostly negative, need turn it into positive

All goalposts. Easily shifted, moved, changed and warped.
You can't shift a 3kilobuck hundreds of CUs part like that.

 

[SolidQ]

They ship really cool halo parts

Zen1 was slower in games, but still gaining market share because Cores/Price

No amount of good enough poverty parts can revive the brand.

Let's wait leaks about UDNA and what they cooked, currently we don't have any info about UDNA

 

[adroc_thurston]

but still gaining market share because Cores/Price

At dismal margins. Ryzen wasn't a real thing before Zen3 won it all.
Vermeer won no questions asked and Vermeer-X kinda sealed the deal.

Let's wait leaks about UDNA and what they cooked, currently we don't have any info about UDNA

no such thing as UDNA.
There's RDNA5, and it has 3 parts and no halo.
Baby, token market presence with no ambition to take any heads whatsoever. Ergo, they're dead.

 

[SolidQ]

Ryzen wasn't a real thing before Zen3 won it all.

That true, but directions was right, now they need to do same with Radeon step by step, don't need halo for now just make strong 6/7/8 series. Just put right price and give nice Upscaling(that main thing for now), next RT, next software etc

You can't shift a 3kilobuck hundreds of CUs part like that.

What is problem? AI companies will buy it.

 

[adroc_thurston]

That true, but directions was right

The direction was right because a year before they shipped another kinda halo part, a 16c big core client CPU. Something Intel could not touch.
It didn't win gaming but it won everything else and by a large margin.

now they need to do same with Radeon step by step

They need a 3-4 kilobuck USD part that kills.

Just put right price and give nice Upscaling(that main thing for now), next RT, next software etc

not. a. chainsaw.

What is problem? AI companies will buy it.

None of the GDDR DC sticks have usable TCO. No one buys them.

 

[SolidQ]

not. a. chainsaw.

first software fixed need with low/mid parts, after fix you can create HALO, because people will know, software is good, we can buy 3k part.

 

[adroc_thurston]

first software fixed need with low/mid parts

Goalpost stuff. Worthless.

after fix you can create HALO

You can't, no one's gonna greenlight it anymore.

because people will know, software is good, we can buy 3k part.

You don't need anything when you're on top.

 

[SolidQ]

You don't need anything when you're on top.

if popular programs like Blender working much slower, than NV at same tier cards, even top won't help

Goalpost stuff. Worthless.

I feel like for all those years none guy at AMD analyze situation or they just ignore it.

You can't, no one's gonna greenlight

Can, if create positive with low/mid parts

it anymore.

People changes in all companies, so we can't speak about "anymore". Maybe next CEO will come adn say create HALO.

 

[adroc_thurston]

if popular programs like Blender working much slower, than NV at same tier cards, even top won't help

A miserable dismal share of client TAM uses Blender on yearly basis.

I feel like for all those years none guy at AMD analyze situation or they just ignore it.

It's all narrative goalpost stuff. Goalies are shifted easily when you're on top.

Can, if create positive with low/mid parts

You, in fact, can not! Ryzen won because AMD consistently shipped parts Intel couldn't. All sauced in chipet magick™ giving rather rickety parts an aura of technological superiority.
Bigger, meaner sticks. For more money.
Threadripper is an even more acute case of brand building, it managed to eliminate Intel from HEDT which no AMD part previously ever could.
Even in P4 era Intel had cool and performant EE's like Gallatin. Now? Not so much.

People changes in all companies, so we can't speak about "anymore". Maybe next CEO will come adn say create HALO.

Lisa is forever. She is eternal.

 

[yuri69]

There's RDNA5, and it has 3 parts and no halo.

Frankly, designing 3 dGPU chips on 3nm is still horribly expensive given the expected revenue.

 

[adroc_thurston]

Frankly, designing 3 dGPU chips on 3nm is still horribly expensive given the expected revenue.

Gotta keep the foot in the market, full exit is always a lot more expensive.

 

[yuri69]

Gotta keep the foot in the market, full exit is always a lot more expensive.

Full exit means 0 chips. A single "x600" chip should do it.

 

[vanplayer]

LOL.

https://videocardz.com/newz/amd-radeon-rx-9070-xt-expected-to-boost-up-to-3-1-ghz-260-330w-tbp-depending-on-a-model

It’s worth noting that the so-called “leaked” 3DMark scores reported by some sites are not accurate. The 3DMark software does not display correct names for unreleased graphics cards this early. So please keep that in mind.

 

[Tuna-Fish]

3.1 would mean ~= 7900XT on compute, but no more. And 20% less bandwidth.

 

[gaav87]

What clocks and power does your OC’ed 6800XT achieve?

20555score stock 2350mhz ~283w
23500score 2650mhz 340w

Average for TS Graphics on 3dmark is 19k... that's a decent OC.

Nah my 9800x3d increases the graphic score by quite alot vs my 5800x3d for example idk why.

 

[gaav87]

LOL.

https://videocardz.com/newz/amd-radeon-rx-9070-xt-expected-to-boost-up-to-3-1-ghz-260-330w-tbp-depending-on-a-model

Like i said yesterday 2.8G + early drivers =~23k TS score
Release drivers 24k+
3x 8pin = 26k ~7900xt gaming
600-700euro (depending on version)

 

[Tup3x]

Like i said yesterday 2.8G + early drivers =~23k TS score
Release drivers 24k+
3x 8pin = 26k ~7900xt gaming
600-700euro (depending on version)

At this point, I doubt drivers will increase performance much. Somewhat made sense when RDNA2 launched but not this time.

 

[adroc_thurston]

A single "x600" chip should do it.

That's basically a full exit. Need to keep most segments staked in.

 

[Saylick]

Like i said yesterday 2.8G + early drivers =~23k TS score
Release drivers 24k+
3x 8pin = 26k ~7900xt gaming
600-700euro (depending on version)

Allegedly, according to the leak, it performs within 5% of a 4080 in games. What those games are? Who knows.

 

[SolidQ]

Allegedly, according to the leak, it performs within 5% of a 4080 in games

because games and 3d mark is different story. 3D mark 06 was last in good 3D mark for me

At this point, I doubt drivers will increase performance much

First restults Navi31 were also lower