Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[dzoni2k2]

Not really. It might be embarrassing if AMD achieved that through pure architecture and smart compression tech (eg. a reverse of 256-bit Maxwell vs. 512-bit Hawaii), but having a boatload of cache on die is merely the other obvious solution to the same problem.

If anything is going to be embarrassing for Nvidia, it's the perf/w difference.

I don't think AMD just traded wider bus for bigger die size (more cache). That cache is probably crucial for the perf/W you are seeing.

 

[dr1337]

That would mean the same board is used for more than one GPU though, which would be strange. The GPU is either HBM, or GDDR6. Because the package has to change.

This is most likely just a dev board that was being used to test something else, or they were going to hand place the RAM chips (which can be done).

I mean yeah but don't the linux drivers for sienna chinchillid aka navi 21 mention both GDDR6 and HBM controllers? The rumor came out a while back, before the GPU specs were added to the drivers and the 256 bit bus came to light. I've tried searching it but all I find are articles about the current leaked specs. It was definitely talked about further down in this thread though.

The kapton tape makes it seem so deliberate. Nobody cares that much about some open pads especially if they're going to hand place surface mount components or just do some quick testing. But then again maybe thats a clue that its bait.

 

[HurleyBird]

I don't think AMD just traded wider bus for bigger die size (more cache). That cache is probably crucial for the perf/W you are seeing.

A part of it, sure, but XSX has great perf/w on a 384-bit bus with (almost certainly) much less cache than desktop parts.

 

[zinfamous]

RADV ACO Lands NGG Geometry Shader Support - Phoronix

www.phoronix.com

NGG Appears To Be In Good Shape For RDNA 2 / RX 6000 Series - Phoronix

www.phoronix.com

NGG paths working for both RDNA1 and RDNA2.

Radeon Linux Driver Seeing "MALL" Feature For Big Navi - Phoronix

www.phoronix.com

Memory Access on Last Level is also ready for N21.

I mean, because previous history of AMD talking about a lot of great features that...never worked or of course required devs to utilize them, and they never did, so they just never worked.

I want this stuff to be true, but you know, all we have to go on is actual history right now. I'm just hoping it's different this time. But it needs to be that way, before it is.

 

[Stuka87]

I mean yeah but don't the linux drivers for sienna chinchillid aka navi 21 mention both GDDR6 and HBM controllers? The rumor came out a while back, before the GPU specs were added to the drivers and the 256 bit bus came to light. I've tried searching it but all I find are articles about the current leaked specs. It was definitely talked about further down in this thread though.

The kapton tape makes it seem so deliberate. Nobody cares that much about some open pads especially if they're going to hand place surface mount components or just do some quick testing. But then again maybe thats a clue that its bait.

Yes, but my thinking is that it will be a market segmentation thing. GDDR6 for gaming, HBM for enterprise. BUT, they could certainly use one board for developing both.

 

[zinfamous]

Double crown = trumping Intel's gaming dominance and Nvidia's for the record....The way I meant it.

It's pretty much unthinkable that they'd pull off a double whammy in the same year, but it's not like 2020 has stayed true to the norm.

The internet forums and article comment sections would go heavily into rage mode!

well, what do you mean by "dominance"? AMD still doesn't have that against Intel, and won't through next year, either. I do believe they have all the performance to claim performance crown....but dominance, imm, is market. AMD is nowhere close, and I think Intel isn't going to lie down and take it without pulling every dirty trick, including sweeping that leg, to keep their market wins. (AMD is only just now creeping into the mobile space again--as their performance has still been ~lacking in that area)

Heck, AMD has had the performance wins in most categories that matter for 2 years now over Intel, but that still hasn't won them much mind share and market share in all that time. This is only now starting to happen.

Knocking out nVidia this year, next year, the year after? lolno.

 

[zinfamous]

Those leaked photos seems to belongs to some AIB OC model (beefed up VRM, higher PCB etc.).

Photo of reference card for comparison

huh, so if that's the reference card, I'm glad they've taken cooling seriously.

 

[dzoni2k2]

huh, so if that's the reference card, I'm glad they've taken cooling seriously.

That looks way too beefy for a reference card.

 

[Stuka87]

huh, so if that's the reference card, I'm glad they've taken cooling seriously.

I would almost bet money that its a dev board. You can even see a standoff that its mounted too.

 

[DDH]

That looks way too beefy for a reference card.

That's AMDs new reference design

 

[Vope45]

https://twitter.com/x/status/1319373361789755392

Looking good but 5700xt is on par with 2080 in FireStrike. Judging from this score alone the XT is at least probably on par with 3080.

 

[Kenmitch]

Looking good but 5700xt is on par with 2080 in FireStrike. Judging from this score alone the XT is at least probably on par with 3080.

Possibly. Without knowing how RDNA2 does in FireStrike relative to it's gaming performance we have no idea in the end. Probably why they threw the salty bone out there.

 

[soresu]

I want this stuff to be true, but you know, all we have to go on is actual history right now. I'm just hoping it's different this time. But it needs to be that way, before it is.

In this case according to the driver devs its fundamentally necessary to make mesh shaders work.

 

[Veradun]

It's the same production line, just with additional EUV steps, so no its not going to create additional capacity. People aren't using the EUV step on 7N for some reason.. probably low yields?

I'd say new design rules for a short lived process while one can use N6 if not migrating to N5 is a clear tombstone for N7+

 

[TESKATLIPOKA]

And what If I agree with your views on Firestrike Ultra?

What If I, myself consider it to be outdated, and not - real world gaming benchmark?

Have you thought that I was simply making fun of this whole launch, and leaks?

Actually I haven't. From your post It was not clear for me and I am not a mind-reader.

 

[Vope45]

Train derailed again for the nth time ?

https://twitter.com/x/status/1319507480993353728

 

[TESKATLIPOKA]

Train derailed again for the nth time ?

https://twitter.com/x/status/1319507480993353728

I find It veird. In DirectX 11 Fire strike It's faster than even RTX 3090, but in DirectX 12 Time Spy It's between RTX 3070 and RTX 3080.
DirectX 11 was AMD's Achilles heel not DirectX 12.

 

[uzzi38]

I find It veird. In DirectX 11 Fire strike It's faster than even RTX 3090, but in DirectX 12 Time Spy It's between RTX 3070 and RTX 3080.
DirectX 11 was AMD's Achilles heel not DirectX 12.

I've said it before, I'll say it again, Time Spy is excessively memory bound. Especially graphics test 2

It's pretty much the worst case scenario for an 80CU GPU with a 256 bit bus.

 

[Shmee]

Curious results there for sure. Could potentially get better DX12 results later on with more optimized drivers? Looks like overall the performance is fairly promising, and could possibly improve. While on it's weaker points now, it is not up with a 3080, it should still be quite a bit faster than a 2080Ti / 3070, which is not a bad improvement.

 

[dzoni2k2]

We need some gaming benchmark leaks. All this scores are useless.

 

[Shmee]

Yeah, and as far as a possible 256bit bus, I wonder how much that will hurt ETH peformance. I wonder if AMD did this on purpose to discourage miners?

 

[reb0rn]

ETH and all memory mining is limited by memory BUS primary so yeah its will be slower then 5xxx...
Only algo being GPU which are mostly asic will be good for 6xxx so RIP mining

 

[TESKATLIPOKA]

I've said it before, I'll say it again, Time Spy is excessively memory bound. Especially graphics test 2

It's pretty much the worst case scenario for an 80CU GPU with a 256 bit bus.

Maybe It won't be such a problem in games.

Timespy uses DX12 while firestrike uses DX11. First gen Navi vs Turing showed the same kind of scaling:
5700XT:
Firestrike Ultra = 6655
Timespy Extreme = 4223

2070 Super:
Firestrike Ultra = 6012
Timespy Extreme = 4771

(avg GPU score from 3Dmark's search)

tweet

 

[HurleyBird]

Yeah, and as far as a possible 256bit bus, I wonder how much that will hurt ETH peformance. I wonder if AMD did this on purpose to discourage miners?

I guarantee you that limiting mining performance is that last thing on anyone's mind when it comes to architecture and the immutable characteristics of each die.

 

[Vope45]

Yeah, and as far as a possible 256bit bus, I wonder how much that will hurt ETH peformance. I wonder if AMD did this on purpose to discourage miners?

Some ground work for MCM maybe ? Similar to zen + to zen 2.