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

 

[marees]

Transistor density of n44

https://twitter.com/x/status/1930267617006366890

 

[blckgrffn]

Wow, check out this Ars endgame summary:



Compared to how these would go even a couple years ago, I think the perceived impenetrable fortress of NVidia is best because "move goalpost here" magic is not nearly as strong as it once was.

Which is just healthy for the market. We can want three players, but having at least two players is better, even if a huge historical advantage is still in play.

 

[branch_suggestion]

So GB206 is about 10% smaller vs Navi 44 while being slightly more efficient and having better RT. All else being a wash. I guess Nvidia still has a slight PPA advantage, GDDR7 can’t account for all of that.

Well N44 has the full uncore unlike GB206.
And of course G6 vs G7, AMD has lower BOM which is what really matters.

 

[SolidQ]

good review with a lot games.

1%

 

[Heartbreaker]

Wow, check out this Ars endgame summary:

View attachment 124883

Compared to how these would go even a couple years ago, I think the perceived impenetrable fortress of NVidia is best because "move goalpost here" magic is not nearly as strong as it once was.

Which is just healthy for the market. We can want three players, but having at least two players is better, even if a huge historical advantage is still in play.

FSR4 really helps level the playing field. DLSS was the #1 reason I went NVidia for an RTX 4070.

Today I would have much less concern choosing an RX 9060 over a 5060 Ti if I was in the market.

It's not only the improved quality of FSR4, but it's the fact that it's now DL based and can be trained to improve at a faster rate than a human coded solution is likely to improve.

If I had chosen an RX 6800 XT or 7800 XT, I'd be a little annoyed that I had missed out on FSR4.

 

[inquiss]

It's a cheaper 7700 XT at a lower price with more VRAM and better upscaling support.

That's not awful, on paper. But I think buyers will consider DLSS worth the price difference that shows up in practice.

Course they will. And when AMD is on par with that it will be something else. Maybe nvenc again yeah nvenc

 

[GodisanAtheist]

Problem with the current market is undersupply, and we've entered a topsy-turvey world where AMD no longer brings NV prices down, but NV brings AMD prices up.

I'd be shocked if the 9060XT remains less than $400 in the real world and immediately destroys its own value proposition.

However, AMD will be able to ride the initial positive reviews and still sell what they make to lower info or desperate consumers who have nowhere else to turn. It's the opposite of AMD's usual strategy of putting its worst possible foot forward for launch reviews and it seems to be paying dividends for them.

 

[Kronos1996]

Well N44 has the full uncore unlike GB206.
And of course G6 vs G7, AMD has lower BOM which is what really matters.

True, I was surprised to see 16 PCIE lanes.

 

[coercitiv]

If I had chosen an RX 6800 XT or 7800 XT, I'd be a little annoyed that I had missed out on FSR4.

I bought the 6800 XT for less than $550 in equivalent US pricing and have already enjoyed the card for more than two and a half years. Today we're being offered FSR4 on hardware that is way too close in price/performance. I have zero regrets.

 

[Hans Gruber]

Techspot's review of the RX 9600XT says that it's made on N4P. Is this true?

 

[Rigg]

Techspot's review of the RX 9600XT says that it's made on N4P. Is this true?

That's what the TPU database says. It also says N44 has a 23% transistor density advantage over GB206.

 

[Thunder 57]

You know why you came here, here's the verdict! /s

View attachment 124866
View attachment 124867

I loved the part where he said it will probably cost more like $400, and at that price he would go for the $420 5060 Ti 16GB, while failing to mention that's nearly a $500 card right now.

 

[igor_kavinski]

Techspot's review of the RX 9600XT says that it's made on N4P. Is this true?

TPU DB lists the same: https://www.techpowerup.com/gpu-specs/radeon-rx-9060-xt-8-gb.c4251

 

[Hans Gruber]

That's what the TPU database says. It also says N44 has a 23% transistor density advantage over GB206.

If it's true (N4P). I would expect to see power consumption numbers inline with the 4060 cards from Nvidia. I guess you could compare them to the 5060/5060ti as well. The 4060 cards were on the 4N node which should be equal to or slightly inferior to N4P for power consumption and performance.

I am thinking a stock RX 9600XT should have 100-130w power consumption under gaming conditions. I was seeing 150-180w which would be N5 like power numbers.

If you look at the Intel B580 on N5. The power consumption numbers are pretty amazing going from N6 on ARC to N5 on Battlemage. N4P is a pretty significant leap in power consumption efficiency over N5. The RX 7600XT was on N6 as well.

 

[SolidQ]

Fabio and 40 games at 1440p

 

[Rigg]

If it's true (N4P). I would expect to see power consumption numbers inline with the 4060 cards from Nvidia. I guess you could compare them to the 5060/5060ti as well. The 4060 cards were on the 4N node which should be equal to or slightly inferior to N4P for power consumption and performance.

I am thinking a stock RX 9600XT should have 100-130w power consumption under gaming conditions. I was seeing 150-180w which would be N5 like power numbers.

If you look at the Intel B580 on N5. The power consumption numbers are pretty amazing going from N6 on ARC to N5 on Battlemage. N4P is a pretty significant leap in power consumption efficiency over N5. The RX 7600XT was on N6 as well.

GB206 is a smaller and less transistor dense die than N44. Based on that, It appears that Nvidia still has a significant architectural advantage. They're getting slightly more performance, from a slightly smaller die, on a less efficient/transistor dense node variant, while using slightly less power. I'm no expert in this by any means though. Someone who actually knows WTF they're talking about would have to chime in here to properly respond to this.

 

[igor_kavinski]

They're getting slightly more performance, from a slightly smaller die, on a less efficient/transistor dense node variant, while using slightly less power.

Probably because their AI powered design tools are wayyy ahead of anything AMD has managed to develop. Not that surprising since Nvidia got a huge head start in making sense of AI and they were able to attract and retain the best AI talent in the world thanks to their already dominant position.

 

[branch_suggestion]

NV has better voltage binning, they kinda need to as they make many more bins than AMD.
Check fixed framerate benchmarks for power, they trade blows just fine but AMD juice stock voltage way more to get as many higher bins as possible.

 

[soresu]

Might it be that they see things differently to some here?

Yes, they see cross marketing opportunities.

I get what you are trying to say, but it's naive at best to say that these people are there for the stunning conversation rather than the very obvious motive of making their own online presence more visible.

 

[soresu]

So GB206 is about 10% smaller vs Navi 44 while being slightly more efficient and having better RT. All else being a wash. I guess Nvidia still has a slight PPA advantage, GDDR7 can’t account for all of that.

Yeah but the extra cost of GDDR7 is non trivial, so the area advantage is significantly mitigated anyway.

 

[marees]

What is this controversy around split embargoes for the 9060xt ?
Any idea ??

 

[coercitiv]

What is this controversy around split embargoes for the 9060xt ?
Any idea ??

Not sure what you mean by split embargoes, but AMD tried to limit who gets to review the 8GB cards. HUB reported they had to ask for the 8GB card (as in AMD told them they need to request it) while other outlets like LTT and Hardware Canucks requested and did not receive 8GB SKUs.

See this exchange between HUB and HC as reference. LTT speaks about it at the begining of their review.

 

[igor_kavinski]

Strong words!

 

[marees]

Not sure what you mean by split embargoes, but AMD tried to limit who gets to review the 8GB cards. HUB reported they had to ask for the 8GB card (as in AMD told them they need to request it) while other outlets like LTT and Hardware Canucks requested and did not receive 8GB SKUs.

See this exchange between HUB and HC as reference. LTT speaks about it at the begining of their review.

Check this (this is about 16gb)

https://twitter.com/x/status/1930505111740023102

 

[marees]

Not sure what you mean by split embargoes, but AMD tried to limit who gets to review the 8GB cards. HUB reported they had to ask for the 8GB card (as in AMD told them they need to request it) while other outlets like LTT and Hardware Canucks requested and did not receive 8GB SKUs.

See this exchange between HUB and HC as reference. LTT speaks about it at the begining of their review.

I am referring to 16gb. Check this

https://twitter.com/x/status/1930505111740023102