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

 

[adroc_thurston]

So they actually delayed the cards to make them more expensive. Wow.

It was pretty obvious.

Looks like AMD chose to get ripped a new one in reviews and lose all the ability to regain marketshare.

They do not want MSS at expense of margins.

AMD needed a clear perception of higher value to have people regain interest in their brand

Nope. What they need is a 300CU 800W halo part.

Is there any IPC gain if their new 64 CUs need more power than 96 older CUs plus all mcm associated expences and 50% more memory stuff

Nope. It's lower power.

 

[PJVol]

It's lower power.

Why such a ridiculous psu requirements then?

 

[adroc_thurston]

Why such a ridiculous psu requirements then?

I assume they diablotek-proofed it.

 

[ToTTenTranz]

They do not want MSS at expense of margins.

Bulls"$#.

Are you saying they had no margins on the 9070 series before they decided to delay the cards to hike the price?

 

[Tuna-Fish]

My bet is that the OC models have high PSU requirements because they are juiced to the gills to get that last few % of perf, for better competitive positioning.

 

[coercitiv]

What's wrong with these AIBs having the psu requirement of 850-900W ?

Sapphire 6800XT Nitro+ 350W TBP, min 850W PSU
Sapphire 7900XT Nitro+ 368 TBP, min 750W PSU

Let's get a very rough feeling for transient loads on 6800XT vs. 7900XT

Spoiler: Spikes 20ms

So RDNA3 has tamed power spikes to around ~15-20% of TBP @ 20ms, but RDNA2 could easily spike over 50% TBP for the faster SKUs while something like RX 6800 was happy with just ~20%. If we assume RDNA4 is pushed hard to compete, then 350W TBP and 500W+ spikes are quite possible.

Napkin math time:
Intel Hot-Rod CPU - 150W average, 50W+ extra for spikes
Radeon Hot Wine GPU 2025 - 350W TBP, 150W extra for spikes
Low power PC components for gamers mindful of the planet - 50W
RGB tax because the planet is doomed anyway - 50W

AVG power -> 600W
20ms Transients -> 800W

 

[lucasworais]

Bulls"$#.

Are you saying they had no margins on the 9070 series before they decided to delay the cards to hike the price?

Why have good margins when you clearly have an opportunity to have even better margins? That’s business basics.
The volume probably wouldn’t change that much anyway.

 

[GTracing]

I feel for those trying to build kids gaming pc's on the cheap

The low end market actually isn't too bad. You can get an rx 7600 for $260 from newegg right now. I guess it depends on what you consider "cheap".

 

[In2Photos]

So now pricing for the 9070XT is up around where the 7900XT was? And performance is about the same? What's the point? AMD truly will have become Nvidia with that logic.

 

[maddie]

Bulls"$#.

Are you saying they had no margins on the 9070 series before they decided to delay the cards to hike the price?

I agree.

Why have good margins when you clearly have an opportunity to have even better margins? That’s business basics.
The volume probably wouldn’t change that much anyway.

Throws away the stated goal of this generation and those following. Increasing marketshare in the highest volume mid part of the TAM.

I guess a few Execs will boast of increasing division margins in their next reports, ignoring the designed role of RDNA4 to be a cheap, mean, marketshare increasing weapon.

I wrote some weeks ago and repeat now. They are FOOLS pretending to do strategy while resorting to cheap tactics for victory.

 

[linkgoron]

Why have good margins when you clearly have an opportunity to have even better margins? That’s business basics.
The volume probably wouldn’t change that much anyway.

@maddie is correct. You can't seriously claim that you're doing a "small die" or "affordable die" strategy, aiming for maximizing TAM and creating affordable graphics cards, and then turn around and price your cards at the same price as the competition. Obviously, you can do it - just don't expect anyone to take you seriously afterwards.

Either you have a strategy, or you don't. It looks like the actual strategy was "We thought that Nvidia was going to destroy us so we said some BS about affordable cards, and didn't even create a large die card because we thought that Nvidia is this huge unstoppable machine". Now, I always thought that the TAM talk was nonsense because AMD just wasn't competitive, but they still officially SAID those thing, how they prefer midrange cards or whatever. Now, it looks like this gen is actually a dud from Nvidia and RDNA4 is competitive so they're just going to backtrack on everything and just continue doing business as usual.

 

[PJVol]

So RDNA3 has tamed power spikes to around ~15-20% of TBP @ 20ms, but RDNA2 could easily spike over 50% TBP for the faster SKUs while something like RX 6800 was happy with just ~20%. If we assume RDNA4 is pushed hard to compete, then 350W TBP and 500W+ spikes are quite possible.

What makes you think that rdna4 spikes are somehow worse than with rdna3? Not to mention that for a decent psu short current spikes have long ceased to be an issue. Furthermore, starting from rdna3 AMD have implemented some additional PM logic to deal with fast di/dt's, like Cac/EDC throttlers aimed at both on-chip and VR longevity.
I tend to partly agree with adroc on "overprovisioning", but also don't rule out the case where they push the clocks way above the linear part of the VF curve.

 

[reaperrr3]

for a decent psu

Don't underestimate how many people lack exactly that.

Look up the Hardware Unboxed YT videos about the exploding Gigabyte PSUs from 2021.
And that happened with ~60% loads.

Pretty sure some cheap 750-850W PSUs out there aren't that much better/safer (probably even some 900W+ PSUs, for that matter).

 

[GTracing]

You can't seriously claim that you're doing a "small die" or "affordable die" strategy, aiming for maximizing TAM and creating affordable graphics cards, and then turn around and price your cards at the same price as the competition. Obviously, you can do it - just don't expect anyone to take you seriously afterwards.

I completely disagree with this take. It is 100% possible to make a play for more market-share while still only matching your competitor's price.

If AMD makes a play to achieve a moderate amount of market-share, say 35% this gen. That will increase the supply, and thus lower the price for both AMD's and Nvidia's offerings. Pricing their GPUs $200 below what they're worth will just cause the vast majority of them to be scalped; it doesn't benefit consumers.

 

[coercitiv]

What makes you think that rdna4 spikes are somehow worse than with rdna3?

I could speculate on a number of reasons, but my example was less about what I think and more about what AIBs are telling us with their PSU ratings. The alternative would be higher TBP.

 

[Hail The Brain Slug]

I completely disagree with this take. It is 100% possible to make a play for more market-share while still only matching your competitor's price.

If AMD makes a play to achieve a moderate amount of market-share, say 35% this gen. That will increase the supply, and thus lower the price for both AMD's and Nvidia's offerings. Pricing their GPUs $200 below what they're worth will just cause the vast majority of them to be scalped; it doesn't benefit consumers.

Scalpers will try to scalp them no matter what. They've learned they can cause or majorly contribute to the very shortage that profits them.

 

[PJVol]

Don't underestimate how many people lack exactly that.

Yeah, that's true. Although those units incapable of coping with fast spikes just engage OCP/OVP and don't explode?

 

[coercitiv]

I completely disagree with this take. It is 100% possible to make a play for more market-share while still only matching your competitor's price.

That would require AMD to have the better overall solution, clearly beating Nvidia in both hardware and software. Some say it would also require a winning flagship.

Pricing their GPUs $200 below what they're worth will just cause the vast majority of them to be scalped; it doesn't benefit consumers.

Scalping only works when supplies are low or demand is through the roof. If anything, AMD pricing close to Nvidia would require them to throttle production, which would encourage scalping. Did someone say price fixing?! Shut up!

 

[gaav87]

Frank reponse to the 9070xtx 32gb comedy gold.
Nvidia should respond to the Always Marketing Disaster with
"Always Melting Dude"

 

[jpiniero]

@maddie is correct. You can't seriously claim that you're doing a "small die" or "affordable die" strategy, aiming for maximizing TAM and creating affordable graphics cards, and then turn around and price your cards at the same price as the competition. Obviously, you can do it - just don't expect anyone to take you seriously afterwards.

Well, assuming the real price of the 5070 Ti and 5080 is $849 and $1299.. this is still quite a bit cheaper.

 

[PJVol]

Always Marketing Disaster

Just hilarious

 

[GTracing]

Scalpers will try to scalp them no matter what. They've learned they can cause or majorly contribute to the very shortage that profits them.

No, if there's enough stock that scalping isn't proffitable then they won't scalp

 

[gdansk]

People are always mad about pricing.
But they weren't going to buy Radeon anyway, so who cares.

But AMD needs to think about reviews. If they don't counter the biggest feature of Blackwell (fake MSRP technology) they're in for a rough time.

 

[GTracing]

Scalping only works when supplies are low or demand is through the roof. If anything, AMD pricing close to Nvidia would require them to throttle production, which would encourage scalping. Did someone say price fixing?! Shut up!

They can lower prices once stock normalizes.

 

[ToTTenTranz]

Why have good margins when you clearly have an opportunity to have even better margins? That’s business basics.
The volume probably wouldn’t change that much anyway.

If you think people will always pay whatever price tag AMD decides, why not just price the 9070 XT at $1000 then?
They're losing the opportunity to have even higher margins than that. Just think of how much money they're leaving on the table by not pricing the cards higher.



They could on the other hand realize their brand perception, marketshare and mindshare are at an all time low and the competition has never been stronger in how much they can manipulate developers into adopting exclusive features.

Undercutting them by 10% is going to do the exact same thing it did for the past 10 years which is to suffocate the brand even more.
It's the very definition of insanity.