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

 

[GodisanAtheist]

I'd never really count NV out until we have head to head benchmarks, but yes given NV's hype slides were weirdly underwhelming and it does look like AMD has about as good a shot at convincingly taking the crown this round as its had in... well for about as long as I can remember.

Since the R300 maybe?

Given RDNA 2 was far more competitive in practicality than it would appear on paper, if the trend keeps up this round we might see Nvidia actually lose the Halo spot.

Not holding my breath though.

 

[IEC]

To me the most striking part about this is the stark difference between the AMD of today vs yesteryear.

Remember, when Vega 10 used up the bulk of the transistors to enable more clock speed and failed miserably?
I wouldn't be surprised if Navi 21 is actually the first chip where AMD used machine-learning tools extensively to extract so much clock headroom on the same node as the predecessor.

I'm pretty sure this "leaker" is the one who has trolled Videocardz.com and the like before with fake leaks. So I don't believe the 4GHz rumor one bit.

 

[Saylick]

I'm pretty sure this "leaker" is the one who has trolled Videocardz.com and the like before with fake leaks. So I don't believe the 4GHz rumor one bit.

That was FCL. This leaker is HXL.

I know... There's too many names lol

 

[Frenetic Pony]

I'm pretty sure this "leaker" is the one who has trolled Videocardz.com and the like before with fake leaks. So I don't believe the 4GHz rumor one bit.

GHZ is be limited by power/thermal density (chip get too hot even with big cooler). And power leakage is nonlinear, a fancy word for "clockspeed makes power usage, and thus heat, go up exponentially". So "almost 4ghz" is bullshit, a 50% power efficiency gain might, might translate to a 25% increase in clockspeed. The increased wattage of cards doesn't help much because AMD will still be running their smaller cards (6600 and watnot) at the highest clockspeeds their coolers can handle.

Still, a 20-25% (maybe 30ish with a crazy cooler?) increase in clockspeed can only be good for prices. Nvidia's huffing their own farts while they can with those 4xxx prices (do they just, not care there's a major demand crash?) and hopefully AMD can bring them down to size.

 

[Saylick]

GHZ is be limited by power/thermal density (chip get too hot even with big cooler). And power leakage is nonlinear, a fancy word for "clockspeed makes power usage, and thus heat, go up exponentially". So "almost 4ghz" is bullshit, a 50% power efficiency gain might, might translate to a 25% increase in clockspeed. The increased wattage of cards doesn't help much because AMD will still be running their smaller cards (6600 and watnot) at the highest clockspeeds their coolers can handle.

Still, a 20-25% (maybe 30ish with a crazy cooler?) increase in clockspeed can only be good for prices. Nvidia's huffing their own farts while they can with those 4xxx prices (do they just, not care there's a major demand crash?) and hopefully AMD can bring them down to size.

My expectation is that AMD can get sustained 3 GHz with 350W or less. How high it clocks after that is anyone's guess. Maybe with 450W+ N31 can briefly hit 3.6 GHz or higher, but I wouldn't bet that it can do 3.5 GHz sustained without some serious cooling and power requirements.

 

[moinmoin]

For N5 based Zen 4 CPUs the announced base frequencies are significantly higher than previous gens/nodes which indicates AMD managed to move the efficiency infliction point in the v/f curve significantly upward as well. It's not unthinkable that AMD applied that knowledge to the N5 based RDNA3 chips right away.

 

[RnR_au]

GHZ is be limited by power/thermal density (chip get too hot even with big cooler). And power leakage is nonlinear, a fancy word for "clockspeed makes power usage, and thus heat, go up exponentially". So "almost 4ghz" is bullshit, a 50% power efficiency gain might, might translate to a 25% increase in clockspeed. The increased wattage of cards doesn't help much because AMD will still be running their smaller cards (6600 and watnot) at the highest clockspeeds their coolers can handle.

Still, a 20-25% (maybe 30ish with a crazy cooler?) increase in clockspeed can only be good for prices. Nvidia's huffing their own farts while they can with those 4xxx prices (do they just, not care there's a major demand crash?) and hopefully AMD can bring them down to size.

Didn't AMD have some innovative L1 cache sharing patent a while back? Reduced the perf/power needs by eliminating redundant calculations?

Has this been rolled out yet?

Just saying that power improvements could be possible with RDNA3 that is in addition to moving to N5.

 

[TESKATLIPOKA]

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

It's actually 2.52GHz vs 3.5-3.6GHz = +39-43% and Nvidia tends to understate the actual boost.

Why do you think raster will improve by so little? Specs tell us something else.
Ada 102 vs 3090 Ti
SM count: 128 / 84 = +52%
Shader(FP32) count: 16384 / 10752 = +52%
FP32+INT32 count: 16384+8192 / 10752 = +129%
TMU count: 512 / 336 = +52%
ROP count: not sure
Base frequency (boost): 2230(2520) / 1560(1860) = +43%(+35%)
Then there is Shader execution reordering which supposedly improves gaming performance by 25%.
And this is only the cutdown Ada with 450W TDP.

In comparison, N31 saw massive increase in Shaders(+140%) and TFlops(+292%), the rest improves mainly by the 52-56% higher clock and some modest 20% increase in count at best.

P.S. Can someone explain to me, why we even now mostly compare in pure raster? Isn't It time to compare with RT on at least for highend?
Let's be honest, If I buy N31 or Ada102 then I want the best eye candy and If I disable RT I am actually not playing with the best visuals on. I would disable It only If FPS drops massively or activate DLSS or FSR.

edit: Ada in the end didn't separate INT32/FP32 units.

 

[TESKATLIPOKA]

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

What It is then? A unicorn?
It has 2x more shaders and other stuff was doubled or at least beefed, but 1x RDNA3 WGP shouldn't perform as 2x RDNA2 WGPs.

 

[Saylick]

MLID alleges, via a source, that AMD will launch 2 SKUs this year: 7900XT and 7800XT.

 

[beginner99]

MLID alleges, via a source, that AMD will launch 2 SKUs this year: 7900XT and 7800XT.

Yeah i think it was pretty clear no new midrange in 2022.

 

[TESKATLIPOKA]

N33 is supposedly mobile first, so my bet for announcement is CES 2023.

 

[Kepler_L2]

MLID alleges, via a source, that AMD will launch 2 SKUs this year: 7900XT and 7800XT.

Nope.

 

[DisEnchantment]

MLID alleges, via a source, that AMD will launch 2 SKUs this year: 7900XT and 7800XT.

Nope.

May the Fsource be with you

 

[Saylick]

Nope.

Full N31 and a cut-down N31? Or no?

 

[Timorous]

So with what we have seen the 4090 is about 1.6x the 3090Ti in raw raster performance.

At 4k the 3090Ti is about 1.17x the 6900XT (used TPU charts for ease) making the 4090 about 1.87x the 6900XT. If the top N31 is 2x the 6900XT performance then it is going to be the performance leader by a margin, if it is closer to 2.5x as earlier rumours suggested then it will be 1 or 2 tiers ahead.

In RT RDNA 2 drops around 60-70% fps so N31 with no relative RT improvement and 2x raster perf would be around 1.35x more performance with RT on vs the 6900XT with RT off. It would also make it faster than the 3090Ti with RT on. With 2.5x more performance N31 has RT perf way ahead of the 3090Ti. If AMD improve the performance drop from turning RT on then it gets even better meaning iy could actually compete with a 4090 in RT as well.

Lots of ifs there but AMD being outright faster in both raster and RT is not actually off the table based on what NV have showed.

This DLSS3 also makes me think back to the HWUB blacklist issue a while ago. Imagine if media outlets agreed to NVs terms and started having DLSS3 with interpolated frames included in the standard performance graphs. It would so easily skew the real picture and I think this will be telling ad to which outlets are worth reading going forward.

 

[leoneazzurro]

Nope.

More SKUs? In effect Angstronomics was hinting at more parts. Stacked cache too?

 

[DisEnchantment]

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

Can someone explain to me, why we even now mostly compare in pure raster?

Why do you think raster will improve by so little? Specs tell us something else.

Going by their own published scores buried deep in all that PR speak. (e.g. Valhalla, Division etc.)

 

[coercitiv]

P.S. Can someone explain to me, why we even now mostly compare in pure raster? Isn't It time to compare with RT on at least for highend?

Because RT is a wildly moving target, both in terms of computation requirements and perceived benefits. Here's Nvidia showing you how enabling next gen RT looks like, press F if you're getting tessellated concrete vibes:



Ignore RDNA for a second and think how you should compare 3000 vs. 4000 series in this context.

 

[jpiniero]

Going by their own published scores buried deep in all that PR speak. (e.g. Valhalla, Division etc.)

Yep. I suspect nVidia has a severe memory bandwidth issue that maybe AMD can handle better.

Full N31 and a cut-down N31? Or no?

That's what I expect. Probably just called 7900 XT and 7900. Or 7900 XT and 7950 XT.

N33 is supposedly mobile first, so my bet for announcement is CES 2023.

I think it's going to be even later than that for desktop. Maybe you will see N32 at CES. But if you look back, N22 was March the following year and N23 was August.

 

[Tigerick]

I will give it a shot, no more 7700XT cause AMD will rename it to 7800 to match 4080 12GB

Model Name	RDNA3 Die	VRAM	Estimated SRP
RX 7800	Navi32 7168 CU	3 MCD 12GB 192-bit GDDR6	$599 - $799
RX 7800XT	Navi32 7680 CU	4 MCD 16GB 256-bit GDDR6	$799 - $999
RX 7900	Navi31 10240 CU	5 MCD 20GB 320-bit GDDR6	$999 - $1,199
RX 7900XT	Navi31 10752 CU	6 MCD 24GB 384-bit GDDR6	$1,299 - $1,499

 

[jpiniero]

I will give it a shot, no more 7700XT cause AMD will rename it to 7800 to match 4080 12GB

Model Name	RDNA3 Die	VRAM	Estimated SRP
RX 7800	Navi32 7168 CU	3 MCD 12GB 192-bit GDDR6	$599 - $799
RX 7800XT	Navi32 7680 CU	4 MCD 16GB 256-bit GDDR6	$799 - $999
RX 7900	Navi31 10240 CU	5 MCD 20GB 320-bit GDDR6	$999 - $1,199
RX 7900XT	Navi31 10752 CU	6 MCD 24GB 384-bit GDDR6	$1,299 - $1,499

AMD should be able to justify more than the 4080s for Navi 32.

 

[Karnak]

Pretty sure they won't because of the small N5 die size and the tiny MCDs. I'd say full fat N31 will be somewhere around $1149. Still an increase over the 6900XT but below AD103.

And if Nvidia's benchmarks are true then I'm also pretty sure N31 will be faster in raster than AD102. Honestly no idea about RT perf though, could literally be anything.

 

[jpiniero]

And if Nvidia's benchmarks are true then I'm also pretty sure N31 will be faster in raster than AD102. Honestly no idea about RT perf though, could literally be anything.

It's possible that even N32 could be close to the 4090 in raster. Depends on how well the IF$ in the MCD does to handle memory bandwidth saving.

 

[Timorous]

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.

2x perf of 6900XT beats 4090 with a bit of margin (10% or so) 2.5x 6900XT beats the 4090 by 2 performance tiers.

RT is more questionable but if AMD can get the cost down to a 40% fps reduction instead of 60%-70% then I think that will run the 4090 close without DLSS3.