Question Speculation: RDNA3 + CDNA2 Architectures Thread

[uzzi38]

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 is much bigger than from RDNA1 to RDNA2. We should expect many big improvements in GFX11. 🤔" / Twitter

 

[Dribble]

Lastly, consider that AMD's problem in the server space isn't performance, it's mindshare. And mindshare is force multiplicative across markets. Winning in one market increases your mindshare in others. If Ryzen were performing like Bulldozer, it would be more difficult to market RDNA, and so on.

It's not mindshare.

The biggest reason most servers stayed Intel was probably production volume - AMD can't make enough chips. Behind that there is cost to switch - if you are a company and have spent big $$$ verifying your intel cpu's work reliably, then it takes more then parity or a even a slight performance lead for AMD to make it worth doing all that for some AMD setup and making the switch, and you have to believe that AMD will stay ahead because you don't want to have to switch back again in another few years. Behind that there is cost - Intel will have completely different prices available to customers they might lose. Behind that there is stability - not so much of the cpu's but the motherboards.

I don't think "Mindshare" really exists, a better word would be momentum. "If it ain't broken don't fix it" - a company will be conservative with changes to hardware as they if it goes wrong then the minor cost savings of the change are far outweighed by the damage done to the business. Hence as long as Intel keep pumping out reliable offerings then why take the risk?

 

[Kaluan]

For what it's worth, some new BVH stuff in RDNA3 that some may have missed:

https://twitter.com/x/status/1571164858917261315

 

[Aapje]

Given that I don't think AMD management is stupid anymore, my bet is that they're either sandbagging on GPUs (a bigger GCD or multi-GCD product is coming, but has been well hidden), or they tried to do multi-GCD and failed.

I completely disagree. They are making a huge change by switching to chiplets. It's a huge extra risk to also adopt multi-GCD, not in the least because of potential driver and game support issues. Also, SLI was notoriously unreliable due to timing differences. There is a big risk to run into similar issues at a very late stage of the design process or even while making the design ready for production.

If I were AMD, I would also start by splitting out the IO/L3 cache and then experiment with multi-GCD on near-finalized RDNA3 silicon. Then if that goes nearly perfectly, I could see a next year refresh, but it's more likely that we'll see this for RNDA4.

 

[Kepler_L2]

I think there's a good chance that the extreme clocks only apply to the 5nm GPUs, that is, N31 and N32.

Nope.

 

[maddie]

Nope.

It seems the clock increases depend on the rearchitecturing (simplifying) of the CU circuitry and not so much on the node. N5, by itself, delivers ~10% speed increase over N6 % ~13% over N7.

N33 being such a small die will allow some ability to catch up to the larger N31 & N32. Clock frequency, signal trace length and all that stuff.

Looking forward to a deep dive analysis one day.

 

[leoneazzurro]

It seems the clock increases depend on the rearchitecturing (simplifying) of the CU circuitry and not so much on the node. N5, by itself, delivers ~10% speed increase over N6 % ~13% over N7.

N33 being such a small die will allow some ability to catch up to the larger N31 & N32. Clock frequency, signal trace length and all that stuff.

Looking forward to a deep dive analysis one day.

IIRC it was +20% performance (speed increase) at iso power vs N7, and -40% power at iso performance.
This for the standard process, thing is, we don't know about some "secret sauce" AMD engineers added to the standard process, in a similar way to what happens for the CPU.

 

[Aapje]

we don't know about some "secret sauce" AMD engineers added to the standard process, in a similar way to what happens for the CPU.

Yuck.

 

[maddie]

IIRC it was +20% performance (speed increase) at iso power vs N7, and -40% power at iso performance.
This for the standard process, thing is, we don't know about some "secret sauce" AMD engineers added to the standard process, in a similar way to what happens for the CPU.

Got it from this.

 

[leoneazzurro]

Launches November 3

Scott Herkelman on Twitter: "Join us on November 3rd as we launch RDNA 3 to the world! More details to come soon! #RDNA3 #AMD https://t.co/oftq1Fjrgt" / Twitter

 

[Karnak]

Launch or announcement? Probably the latter I'd say.

 

[leoneazzurro]

Got it from this.
View attachment 67805

yeah, that is very confusing. It should be +15% perf at ISO power and -30% power at ISO perf vs. N7, but it depends on the N5 variant chosen. Also the ratio from N6 to N5 is more difficult to check. But this is assuming standard TSMC processes, while in both AMD and NV case there is a significant customization effort.

 

[leoneazzurro]

Launch or announcement? Probably the latter I'd say.

They say launch, maybe availability is not too distant from the launch date.

 

[RnR_au]

They say launch, maybe availability is not too distant from the launch date.

It was rumoured back in late August that availability would be on the 15th Nov.

 

[biostud]

It was rumoured back in late August that availability would be on the 15th Nov.

Yeah, it would be weird if they missed the winter sales.

 

[Kaluan]

Yeah, it would be weird if they missed the winter sales.

There's almost no chance of that happening. The only big question is if they'll have sub $900 SKUs on shelves before nVidia does.

 

[jpiniero]

There's almost no chance of that happening. The only big question is if they'll have sub $900 SKUs on shelves before nVidia does.

The rumors sounded like only N31. It's a safe bet that all the N31 models will be more than that.

 

[Saylick]

So Nvidia charges $899 for the 4080 12 GB, which uses a ~300mm2 N4 die and is overpriced to begin with, while AMD is going to show up with ~300mm2 of N5 + 225mm2 of N6. How much extra money does it cost for the N6 chiplets and the extra 12 GB of RAM? Seems like it's entirely possible that we could see the non-V-cache version of the 7900XT under $1200. If it ties the 4090 in raster and gets within 75% of the RT performance, I think that's a compelling product. More importantly, who would buy the 4080 16 GB if it costs the same as the 7900XT and loses to it?

 

[jpiniero]

So Nvidia charges $899 for the 4080 12 GB, which uses a ~300mm2 N4 die and is overpriced to begin with, while AMD is going to show up with ~300mm2 of N5 + 225mm2 of N6. How much extra money does it cost for the N6 chiplets and the extra 12 GB of RAM? Seems like it's entirely possible that we could see the non-V-cache version of the 7900XT under $1200. If it ties the 4090 in raster and gets within 75% of the RT performance, I think that's a compelling product. More importantly, who would buy the 4080 16 GB if it costs the same as the 7900XT and loses to it?

I think it's going to be faster in raster... I think the real question is will it be $1599 or will it be more.

 

[Saylick]

I think it's going to be faster in raster actually... I think the question is will it be $1599 or will it be more.

No way AMD asks $1499 for the 7900XT, meaning I would really like to buy one and that's a little more expensive than I'd like. 😅

 

[jpiniero]

No way AMD asks $1499 for the 7900XT, meaning I would really like to buy one and that's a little more expensive than I'd like. 😅

Heh. I've mentioned this before but I do have doubts about how high AMD can go with pricing regardless of performance. But there's only one way to find out...

 

[HurleyBird]

I think it's going to be faster in raster...

I think so too. And it's possible it won't even be close, if the close to 4GHz rumors are true. Might actually beat Ada in raw TFLOPs while all the fixed function stuff is running at a much higher frequency.

 

[TESKATLIPOKA]

N31 won't be clocked at 4GHz, but supposedly only ~3.5-3.6GHz.
I made a table with supposed specs.

	RX 6950 XT (N21)	N31	Difference
WGP(CU)	40(80)	48(96)	+20%
Frequency(boost)	1890 (2310 MHz)	? (~3600 MHz)	+56%
Shaders	5120	12288	+140%
TMUs	320	384 ? (8 per WGP)	+20%
ROPs	128	128 ?	+0%
Memory	256 bit	384 bit	+50%
Bandwidth	576GB/s (GDDR6 18GHz)	960GB/s (GDDR6 20GHz)	+67%

Shaders saw by far the greatest increase, the rest will benefit mostly from 56% higher clocks.
I am not sure If N31 will be capable to tie with Ada102 in raster.

 

[maddie]

N31 won't be clocked at 4GHz, but supposedly only ~3.5-3.6GHz.
I made a table with supposed specs.

	RX 6950 XT (N21)	N31	Difference
WGP(CU)	40(80)	48(96)	+20%
Frequency(boost)	1890 (2310 MHz)	? (~3600 MHz)	+56%
Shaders	5120	12288	+140%
TMUs	320	384 ? (8 per WGP)	+20%
ROPs	128	128 ?	+0%
Memory	256 bit	384 bit	+50%
Bandwidth	576GB/s (GDDR6 18GHz)	960GB/s (GDDR6 20GHz)	+67%

Shaders saw by far the greatest increase, the rest will benefit mostly from 56% higher clocks.
I am not sure If N31 will be capable to tie with Ada102 in raster.

Don't think you should compare WGP(CU). They have the same name but seem to be different creatures.

 

[DisEnchantment]

I am not sure If N31 will be capable to tie with Ada102 in raster.

Big, but not so big leap in raster for Ada it looks, ~1.5x It will comfortably be more than a tie there. Other things to be seen.
Just clock speeds alone 2.3 GHz --> 3.5 GHz = +52% for N31

But what I find amusing or rather interesting is that Scott Herkelmann being confident of putting out a reveal tweet just hours before Ada reveal. Seems they are also eager to show their stuffs. But whatever, we'll see.

I can imagine AMD drumming around points like below already
Efficiency. (but honestly those 4 slot cards look atrocious)
PCIe G5
Smart Access Memory 2.0
Smart Access Storage --> They already hinted about this during Ryzen reveal a new mode of operation build from grounds up for Direct Storage with Zen4 and AM5, which is an even more optimized way of doing Direct Storage
Smart Shift 3.0?

Rest is up in the air.

 

[Frenetic Pony]

Welp, looks like AMD is winning the hardware race this round, Nvidia's "2x if you squint right at 3 actual titles and we don't mention the resolution they're running at please don't notice it's actually below 2x" is surprisingly unimpressive for the huge leap they made in process node. I was honestly expecting an actual, practical 2x/

That being said the frame intertweening in DLSS3 is cool, but not novel, I'm sure we'll see it from Intel and AMD in due time. But it'll still be a "differentiator" for Nvidia, at least in the mind of fans, for a while.