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

 

[branch_suggestion]

40% faster than GRE is the claim from AMD apparently

Hopefully that means the price is in the same realm.

If true, it should basically match the 7900XTX, at least up to 1440p. Maybe it falls a little behind at 4K as the bandwidth requirements go up. We'll see.

Well TPU has the XTX 37% faster than the GRE at 4K.
Therefore it is a 16GB N31 at lower power with far better RT and low precision accel using much less Si.
Aka the best uplift of any GPU architecturally since Maxwell.

 

[Kepler_L2]

yeah, it's 7900XTX plus or minus 5%.
Do you know the MBA wattage?

304W is the official spec but MBA won't be released

 

[adroc_thurston]

Hopefully that means the price is in the same realm.

Man I admire your optimism.

Aka the best uplift of any GPU architecturally since Maxwell.

no RDNA2 was better I think, didn't crank power up.

304W is the official spec but MBA won't be released

wait they teased MBAs and canned them? Impressive.

 

[MrTeal]

If true, it should basically match the 7900XTX, at least up to 1440p. Maybe it falls a little behind at 4K as the bandwidth requirements go up. We'll see.

It makes sense, that puts it right about at 5070 Ti raster performance, and at a $699 MSRP it'd be Nvidia - $50.

 

[branch_suggestion]

Man I admire your optimism.

$649 would be the limit of the same realm, not too crazy to think.

no RDNA2 was better I think, didn't crank power up.

I mean it was similar PPC to RDNA1, and roughly 2x area for 2x perf which is good, but around the same as RV770 to Cypress.
Power was very good, but overall PPA was muted due to IC vs a 384b bus.
Both RDNA1 and 3 were flawed to an extent for sake of comparison, I think overall Maxwell was more impressive than RDNA2.

 

[gdansk]

$699 MSRP it'd be Nvidia - $50.

The Radeon strategy. But this time... it'll work!? AMD can't stop themselves.

 

[Elfear]

It makes sense, that puts it right about at 5070 Ti raster performance, and at a $699 MSRP it'd be Nvidia - $50.

Unfortunately, it does make sense. If the 9070XT sells for real MSRP, it would be a $900-1,000 5070Ti against a $700 9070XT. For that kind of difference, I think a lot of buyers would go for the AMD card (I know I would). If it's only a $50 price difference though, AMD is going to be a tough sell.

 

[branch_suggestion]

It makes sense, that puts it right about at 5070 Ti raster performance, and at a $699 MSRP it'd be Nvidia - $50.

5-10% ahead stock for stock iso-power it seems.
Very similar Si area used for both, and NV has a big membw advantage.
Both RT and ML perf advantages for NV seem largely muted too.
So AMD wins this gen if things bear out on a technical level, but if they built the big boy it would've been a real victory.

 

[adroc_thurston]

$649 would be the limit of the same realm, not too crazy to think.

Indeed but sanity sucks.

Both RDNA1 and 3 were flawed to an extent for sake of comparison, I think overall Maxwell was more impressive than RDNA2.

I guess.

If the 9070XT sells for real MSRP, it would be a $900-1,000 5070Ti against a $700 9070XT

NV will ramp volumes soon enough.
They just had a late prod kickoff for Blackwell since the thing is rife with issues on all levels.

 

[GodisanAtheist]

$649 would be the limit of the same realm, not too crazy to think.

I mean it was similar PPC to RDNA1, and roughly 2x area for 2x perf which is good, but around the same as RV770 to Cypress.
Power was very good, but overall PPA was muted due to IC vs a 384b bus.
Both RDNA1 and 3 were flawed to an extent for sake of comparison, I think overall Maxwell was more impressive than RDNA2.

- I think RDNA 2 gets brownie points for sort of coming out of nowhere. Everyone expected 6800XT = 3070 let's go home thanks to AMD's abysmal track record and suddenly BAM competitive stack.

Maxwell was a helluva improvement over Kepler, but all the same Nvidia was already pulling away from AMD at that point.

 

[gaav87]

40% faster than GRE is the claim from AMD apparently

Good.

 

[adroc_thurston]

I think RDNA 2 gets brownie points for sort of coming out of nowhere

No.

Everyone expected 6800XT = 3070

No one with a brain did.

Maxwell was a helluva improvement over Kepler

So was RDNA2, it was a major PPW and PPB improvement with no additional die area, on the same node.

 

[GodisanAtheist]

No.

-Yes.

No one with a brain did.

-That's a lot of people.

So was RDNA2, it was a major PPW and PPB improvement with no additional die area, on the same node.

-Yep.

 

[Josh128]

40% faster than GRE is the claim from AMD apparently

Sauce??

 

[Josh128]

yeah, it's 7900XTX plus or minus 5%.
Do you know the MBA wattage?

Supposed to be 304W TBP for the 2970MHz MBA version of XT. 3.1GHz will probably be in the 360W+ range.

https://videocardz.com/newz/amd-radeon-rx-9070-xt-confirmed-as-304w-card-rx-9070-non-xt-is-220w

 

[Josh128]

It makes sense, that puts it right about at 5070 Ti raster performance, and at a $699 MSRP it'd be Nvidia - $50.

Yup. This is what we'll see, and that price will probably stick and may even increase due to the insane market conditions these days. Im calling $579 for non-XT version.

 

[adroc_thurston]

Sauce??

Me

3.1GHz will probably be in the 360W+ range.

It's nowhere near this steep.

 

[branch_suggestion]

So was RDNA2, it was a major PPW and PPB improvement with no additional die area, on the same node.

N23 was a really good gain vs N10 for sure, so much that AMD is about to do revision 3 of the same config.

Supposed to be 304W TBP for the 2970MHz MBA version of XT. 3.1GHz will probably be in the 360W+ range.

Lets see real world clocks first.

 

[adroc_thurston]

AMD is about to do revision 3 of the same config.

well, this time they're EVIL cu's.

 

[SolidQ]

interesting how 9600XT will compete vs 5060ti(shoud be one of most selling card)
12 gb (AMD) vs 5060ti 16 not much to do here

 

[branch_suggestion]

well, this time they're EVIL cu's.

7700XT should be doable for N44, extrapolating from N48 and historical scaling.

interesting how 9600XT will compete vs 5060ti(shoud be one of most selling card)
12 gb (AMD) vs 5060ti 16 not much to do here

Do we know if 60Ti is cut GB205 or full GB206?
Not that it matters, 3/4 N48 clears both.

 

[SolidQ]

full GB206?

TPU saying it's GB206

 

[Josh128]

Me

It's nowhere near this steep.

Doubt, but I hope you are correct about that.

 

[GodisanAtheist]

If 9070XT and 9070 are both fully enabled N48 with different clock/power/voltage limits for stacking, then 9600XT is virtually guaranteed to be a cut N48 and AMD will be able to dial it in to outperform 5060Ti no problemo.

N44 is where things get interesting again IMO.

 

[gaav87]

Doubt, but I hope you are correct about that.

No doubt.