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]

If the problem is the memory shouldn't it be worse at 4K?

New cachemem setup carries the bulk of the 4k perf.
You need moar membw to triage the outliers.

 

[leoneazzurro]

Can they make that clamshell 32Go VRAM 9070 XT have GDDR7?

As there are 3Gb GDDR7 chips available (see RTX5090 Mobile), they could even do 24Gbytes without clamshell, if N48 could support it.

 

[DAPUNISHER]

Doesn't matter if 5070 is unobtanium or only available at or above 9070XT prices.

We needed Nvidia to deliver to bring us better AMD prices. They let us down.

To quote Toms, of all places:

"And finally, what about the RX 9070 versus the RTX 5070, both nominally priced at $549? If you're at all good at math and were paying attention above, you'll already know that the 9070 comes out ahead. It's 12% faster at 4K, 8% faster at 1440p, and 4%/7% faster at 1080p. There are five games where the 5070 manages any lead at all, with Space Marine 2 being the biggest margin of victory and the only one where the 5070 leads at 4K. In general, though, the RX 9070 is clearly better for rasterization performance at native resolution."

I don't know where Jarred is testing but his Space Marine 2 numbers looked badly nerfed. Maybe he is mixing up 40 and 50 series results? Because his long lost cousin Steve Walton has Blackwell broken in the game on every card he tests. How the hell is a 5070ti faster than the 4080 super at 1440 Ultra?


It has to be where the testing is being done?

 

[inquiss]

If the person I talked to doesn't include it in their review, it should be perfectly fine for me to post the screenshot here anyway methinks. It's something anyone could try on N48.

"Nudge*

 

[jpiniero]

Again, that's doesn't make 9070 a good product. Having a product exist only to act as a poor value, to make another product in your lineup look better in comparison is just dumb.

That's why I made the comment about managing supply. The 9070 only exists really to absorb the dies that can't be 9070 XT.

 

[Ranulf]

I don't know where Jarred is testing but his Space Marine 2 numbers looked badly nerfed. Maybe he is mixing up 40 and 50 series results? Because his long lost cousin Steve Walton has Blackwell broken in the game on every card he tests. How the hell is a 5070ti faster than the 4080 super at 1440 Ultra?

View attachment 118973
It has to be where the testing is being done?

Yeah, I saw someone with oddball numbers on SM2 last week's reviews. Maybe it was TPU but I can't remember. I just remember noticing it because I thought SM2 favored AMD a bit because of vram.

 

[Timorous]

I don't know where Jarred is testing but his Space Marine 2 numbers looked badly nerfed. Maybe he is mixing up 40 and 50 series results? Because his long lost cousin Steve Walton has Blackwell broken in the game on every card he tests. How the hell is a 5070ti faster than the 4080 super at 1440 Ultra?

View attachment 118973
It has to be where the testing is being done?

That is the 9070 at TPU, not the 9070 XT

 

[DAPUNISHER]

Yeah, I saw someone with oddball numbers on SM2 last week's reviews. Maybe it was TPU but I can't remember. I just remember noticing it because I thought SM2 favored AMD a bit because of vram.

Without showing where they test and the duration, the info is too easily manipulated. Trust me bro ain't cutting it.

That is the 9070 at TPU, not the 9070 XT

Bruh, look at the rest of the cards compared to the HUB chart FFS.

 

[Mopetar]

Not nearly. Price is really the only win. They still a step behind in RT and AI scaling.

For OG Ryzen it was similar. It didn't win on performance, but it did offer affordable 8-core options that blew Intel out of the water on cost for that performance.

I think it's an apt comparison. It's not better than Blackwell behind the 5070 Ti by a few % in raster, but if the MSRP holds it's effectively the same performance at 66% (or less) of the cost since you can't buy a 5070 Ti at MSRP and even the cards listed at $900 or more are sold out. Looking at eBay listings the market price is about $1,000 but those are only listings from China. Everything else is above $1,100.

If AMD iterates on this and catches (or overtakes) Nvidia in a few generations then it really would make it comparable to Ryzen.

 

[adroc_thurston]

If AMD iterates on this and catches (or overtakes) Nvidia in a few generations then it really would make it comparable to Ryzen.

uh, no.
AMD's not gonna build any leadership products.

 

[Josh128]

Yeah they would.

They'd need double the perf of the 9070XT just to match the 5090, so >650W. Thats no bueno.

 

[Timorous]

Without showing where they test and the duration, the info is too easily manipulated. Trust me bro ain't cutting it.

Bruh, look at the rest of the cards compared to the HUB chart FFS.

4000 looks similar, 7000 and 5000 do a lot better at TPU. Could just be different scenes.

Here the XT is nearly 50% ahead of the 7800XT with a very similar config.

If the 9060XT 16GB can do the same over the 7600XT then it should be 7700XT performance.

As for a cut 9060 if it was the same relative cut as the 9070 is to the 9070 XT then it would in the 4060Ti ballpark however with 8GB Vram it would have cases where it is truly awful.

A 9060 12GB on a 96bit bus may be more like a 3060Ti but performance would have fewer cliff edge cases and it would have longer legs imo.

 

[gdansk]

They'd need double the perf of the 9070XT just to match the 5090, so >650W. Thats no bueno.

Double 9070 would be close enough. And that's only 56 CU @ 220W. Of course, scaling is worse. And so on.
The reason they cancelled it is probably because:
1) it would be massive
2) it would be expensive
3) they assumed 5090 would be even better than it turned out to be

 

[Kronos1996]

I’m thoroughly whelmed. Navi 48 has much worse PPA vs GB203. It’s a 5% small die but 20-25% slower. All I can conclude is that they lost a lot of PPA efficiency by just doubling Navi 44. It should be in-between GB203 + GB205 for this performance level. Closer to 300-330mm2. Also screwed power efficiency. It’s an impressive achievement vs Navi 31/32 but second place is still second place.

 

[adroc_thurston]

they assumed 5090 would be even better than it turned out to be

It's just this.

 

[SolidQ]

Germany inc VAT

• Asus PRIME Radeon RX 9070 16GB - 739,00
• Asus TUF Radeon RX 9070 16GB - 789,00
• Asus PRIME Radeon RX 9070 XT 16GB - 799,00
• Asus TUF Radeon RX 9070 XT 16GB - 879,00
• Sapphire Nitro+ Radeon RX 9070 XT - 879,00
• Sapphire Pure Radeon RX 9070 XT - 789,00
• Sapphire Pulse Radeon RX 9070 XT - 689,00

Asus is drunk

Navi 48 has much worse PPA vs GB203. It’s a 5% small die but 20-25% slower

we need to see results like same memory speed/bandwidth

It's just this.

RDNA5 use GDRR7? or btm use still use 6?

 

[adroc_thurston]

Navi 48 has much worse PPA vs GB203

no.

It’s a 5% small die but 20-25% slower.

no.

Also screwed power efficiency.

they didn't, it's sits right there in the blurb with everyone else.

 

[DAPUNISHER]

Could just be different scenes.

That is precisely what I wrote.

Now, which scene? That's the whole problem here. Unlike the 2 written reviews with the inflated numbers that I now doubt are representative of the game's performance on the hardware. Steve explains the issue is with Blackwell and Nvidia has not fixed it yet. He threw out the data and retested after the latest patch, and it is still broken.The onus is on the written reviewers to prove their work is accurate. We should not be speculating about it, we should be able to see exactly where they tested.

 

[jpiniero]

no

The model Compubase used (the Sapphire Pure) was 20% slower than the 5080 on average.

But yeah... the results are all over the place. If that's because of the membw, it's because of the membw.

 

[SolidQ]

Deep Dive

AMD's RDNA4 Architecture (Video)

Hello you fine Internet folks,

chipsandcheese.com

 

[marees]

Germany inc VAT

• Asus PRIME Radeon RX 9070 16GB - 739,00
• Asus TUF Radeon RX 9070 16GB - 789,00
• Asus PRIME Radeon RX 9070 XT 16GB - 799,00
• Asus TUF Radeon RX 9070 XT 16GB - 879,00
• Sapphire Nitro+ Radeon RX 9070 XT - 879,00
• Sapphire Pure Radeon RX 9070 XT - 789,00
• Sapphire Pulse Radeon RX 9070 XT - 689,00

Asus is drunk



we need to see results like same memory speed/bandwidth


RDNA5 use GDRR7? or btm use still use 6?

My feeling is that AMD has given discount vouchers only to the cheap cards to bring them closer to the launch "MSRP"

No point giving discount to costlier models (from ASUS, Gigabyte etc.)

So for example, if there is a worldwide launch stock of 200000 cards then only 10% 20000 could get AMD's discount voucher

 

[gdansk]

I’m thoroughly whelmed. Navi 48 has much worse PPA vs GB203. It’s a 5% small die but 20-25% slower. All I can conclude is that they lost a lot of PPA efficiency by just doubling Navi 44. It should be in-between GB203 + GB205 for this performance level. Closer to 300-330mm2. Also screwed power efficiency. It’s an impressive achievement vs Navi 31/32 but second place is still second place.

Using 1440p TPU data RTX 5080 is 9% better performance per area. While using, in theory, 20W more power and 50% more memory bandwidth.

 

[moonbogg]

The cards are fine, but no one is buying these if the 5070 and 5070ti are suddenly in stock at msrp. Especially the 5070 vs the 9070. They are similar in raster and the 5070 is faster in ray traced games. That's definitely what people will buy because AMD priced matched Nvidia on that one. The 5070TI is easily superior to the 9070XT and if available at only $150 more, that's the card to get without batting an eye. If these cards sell, it's only due to current market conditions. AMD wasn't aggressive enough IMO. The ray tracing deficit and general Nvidia brand power is killing the mindshare completely.
The 9070 should be $400 and the 9070XT should be $500.
How are these cards better than the 7900 XT and XTX besides price? They could have just cut prices on those and got the same results: more sales for Nvidia.

 

[adroc_thurston]

Using 1440p TPU data RTX 5080 is 9% better performance per area. While using, in theory, 20W more power and 50% more memory bandwidth.

That's not better perf per area, that's "we paid for 30Gbps GDDR7".

 

[blckgrffn]

Using 1440p TPU data RTX 5080 is 9% better performance per area. While using, in theory, 20W more power and 50% more memory bandwidth.

Hey you with the numbers, why do you have to step in here? That was a very punchy take they had there and now its sorta meh. We can't have flame wars with this type of data based discussion!



I think its fine - and these other parts we are comparing too are like $1000-$1800 parts atm? If we can buy them near MSRP even for one day (please) it'll be sorta nuts.