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

 

[eek2121]

The +30% frequency slide seems to indicate that happens when you give up any efficiency gain. So if you want +54% performance/watt increase you have to back down on the frequency.

That would seem to imply that the headroom is there. Maybe AMD is artificially limiting performance of partner cards prior to launch? Just a guess.

It would be neat to see a partner card hit 2.8-3.0ghz. I would be curious as to what the performance would be.

 

[TESKATLIPOKA]

The +30% frequency slide seems to indicate that happens when you give up any efficiency gain. So if you want +54% performance/watt increase you have to back down on the frequency.

Yes, that makes sense. It looks like N5 process advantage is included in that graph.
RX 6950XT has 2.1GHz as gaming frequency and the gaming frequency for RX 7900XTX is 2.3GHz for shaders and 2.5GHz for frontend.

RX 7900XT has very high TBP of 300W considering game clock is only 2GHz(shaders most likely and frontend ~2.15-2.2GHz?) and has 84CU so pretty similar to 6950XT. Based on that frequency/power chart, TBP should have been lower.

 

[maddie]

The +30% frequency slide seems to indicate that happens when you give up any efficiency gain. So if you want +54% performance/watt increase you have to back down on the frequency.

If you get + 30% performance at iso power, isn't that an efficiency gain, just not as much as at iso frequency? I'm really surprised at the small Δ between them.

 

[eek2121]

I do wish we could get some benchmarks soon, rather than right at launch. I hate the launch day review rush.

I am only even considering upgrading my GPU because of one reason: power efficiency. At a lower TGP, the 7900XTX looks to be faster than my 3090. My only reservation is RT performance: will it be competitive with the 3090 (non-Ti)? If so, I will upgrade assuming the GPU fits in my case and is relatively quiet.

 

[TESKATLIPOKA]

AMD RDNA 3 GPU Architecture Deep Dive: The Ryzen Moment for GPUs
Link
An article about RDNA3.

Only 17.4% IPC improvement? Isn't that a bit too underwhelming? Even Ampere had ~25-30%. And for this small increase, they used 165% more transistors per WGP?

Ok. Here is an official slide about boost and gaming clock, and It looks like the lower clock of shaders is the official clockspeed.

 

[DisEnchantment]

And for this small increase, they used 165% more transistors per WGP?

They used it to add myriad of other features. Compute operations are improved in many respects.
You can see in slide 6


This is very good for ROCm and Co.

 

[DisEnchantment]

Slide 3


Slide 6



Slide 20

Three slides in the launch presentation indicating ~3GHz clocks are enough to convince me RDNA3 clocks are met.
Like I said before, corpo types want to put +50% energy efficiency because AMD is pro planet yada yada, therefore they cap the RDNA3 clocks at 2.3 Game clock.

This product was designed to be manufacturable with a cost target and they rightly target the sub 1000USD.
Reminds me of Rick Bergman days.

An AIB card with ~3 GHz factory OC will be there you can bet.

There is also the Rage Mode OC.


Lets wait for AIB cards to see factory OC clocks

 

[maddie]

AMD RDNA 3 GPU Architecture Deep Dive: The Ryzen Moment for GPUs
Link
An article about RDNA3.

Only 17.4% IPC improvement? Isn't that a bit too underwhelming? Even Ampere had ~25-30%. And for this small increase, they used 165% more transistors per WGP?
View attachment 71138

Ok. Here is an official slide about boost and gaming clock, and It looks like the lower clock of shaders is the official clockspeed.

265% transistors per unit area (+ 165%)
154% transistors per WGP (+ 54%)
117.4% IPC (+ 17.4%)

That is a huge transistor density improvement.

Really + 17.4% IPC for + 54% transistors

 

[PJVol]

There is also the Rage Mode OC.

+6% power limit? )

 

[eek2121]

I didn’t see this one posted here. Found on Reddit. It has the 7900XT benchmarks.

 

[Stuka87]

Just because 7900 doesn't hit 3Ghz, doesn't mean the smaller chips won't. Their comment doesn't really state that ALL RDNA3 chips will hit 3Ghz.

 

[Stuka87]

I didn’t see this one posted here. Found on Reddit. It has the 7900XT benchmarks.

Thats from AMD's announcement live stream.

 

[moinmoin]

Only 17.4% IPC improvement? Isn't that a bit too underwhelming?

Does "Enhanced CU delivers approximately 17.4% architectural improvement clock for clock" refer to the improvement of the CU, the WGP or the whole card?

The text says enhanced CU, the slide shows a WGP, and we refer to the IPC of the complete product...

 

[leoneazzurro]

It depends a lot on how it is measured. The increase in perf relative to the 6950XT is relative to the whole card, so it could be fair to say that this may represent the contribution of the CU improvements to the total performance (the other should be the increase in CU number, clocks, and memory subsystem).

 

[TESKATLIPOKA]

They used it to add myriad of other features. Compute operations are improved in many respects.
You can see in slide 6
View attachment 71139

This is very good for ROCm and Co.

Sorry, I am blind. Transistor count in WGP increased by only 54%, the 165% increase was transistors/mm2.

265% transistors per unit area (+ 165%)
154% transistors per WGP (+ 54%)
117.4% IPC (+ 17.4%)

That is a huge transistor density improvement.

Really + 17.4% IPC for + 54% transistors

Yeah, I noticed my mistake when I was replying to DisEnchantment.
If I think how I at the beginning were guessing the improvement will be at least 50% and now It's only 17.4% I feel embarrassed.

Does "Enhanced CU delivers approximately 17.4% architectural improvement clock for clock" refer to the improvement of the CU, the WGP or the whole card?

The text says enhanced CU, the slide shows a WGP, and we refer to the IPC of the complete product...

They mention architectural improvement clock for clock, so I think they really mean for the whole GPU -> IPC. N33 will be great for comparison purposes.
But based on what eek2121 posted, the improvement is 53%, 51%, 67% and 47% in four different games.
Theoretical performance improvement: 100*1.174*1.2*2300/2100= 154%
If I exclude Cyberpunk 2077, then It's in line with the rest.

 

[eek2121]

Thats from AMD's announcement live stream.

Someone said that on reddit. It wasn’t on the livestream.

Livestream slide deck is here: https://www.anandtech.com/Gallery/Album/8202

 

[eek2121]

A bit of info surrounding RT, Note that AMD is claiming up to 1.8X RT performance on “heavy RT workloads” as per this slide:

 

[eek2121]

Sorry, I am blind. Transistor count in WGP increased by only 54%, the 165% increase was transistors/mm2.


Yeah, I noticed my mistake when I was replying to DisEnchantment.
If I think how I at the beginning were guessing the improvement will be at least 50% and now It's only 17.4% I feel embarrassed.


They mention architectural improvement clock for clock, so I think they really mean for the whole GPU -> IPC. N33 will be great for comparison purposes.
But based on what eek2121 posted, the improvement is 53%, 51%, 67% and 47% in four different games.
Theoretical performance improvement: 100*1.174*1.2*2300/2100= 154%
If I exclude Cyberpunk 2077, then It's in line with the rest.

Cyberpunk 2077 is rather poorly optimized IMO. Wish they had tested Spiderman Remastered.

 

[TESKATLIPOKA]

Three slides in the launch presentation indicating ~3GHz clocks are enough to convince me RDNA3 clocks are met.
Like I said before, corpo types want to put +50% energy efficiency because AMD is pro planet yada yada, therefore they cap the RDNA3 clocks at 2.3 Game clock.

This product was designed to be manufacturable with a cost target and they rightly target the sub 1000USD.
Reminds me of Rick Bergman days.

An AIB card with ~3 GHz factory OC will be there you can bet.

Based on the frequency/power chart you posted, RDNA3 gain 30% higher frequency at ISO power or 50% lower power consumption at ISO frequency.
RX 6950XT has 335W TBP and from that ~250-275 W is sorely for GPU, right?
If I made an RDNA3 GPU with comparable specs to RX 6950XT, I should shave off 125-137.5 W, the new TBP would be only 198-210W.
We have a pretty comparable RX 7900XT with 84CU and 2GHz for gaming and 2.4GHz for boost clocks, yet It has a 300W TBP.
Based on that chart It should have lower TBP or higher clocks, yet It doesn't have either. Why? Maybe these rumors about a problem with design are true.
I have to wonder If 450W would be enough for 3GHz OC models.

 

[Stuka87]

Someone said that on reddit. It wasn’t on the livestream.

Livestream slide deck is here: https://www.anandtech.com/Gallery/Album/8202

That slide deck is incomplete. Its actually missing a lot of what was shown. I don't have time to track down where in the original slide deck it was. But I know that chart is what many review places were using to estimate performance versus the 4090. As we know how the 4090 performs versus the 6950, and that shows how the 7900XTX performs vs the 6950.

 

[Karnak]

I didn’t see this one posted here. Found on Reddit. It has the 7900XT benchmarks.

Not gonna lie the so called 7"9"00XT looks like poor value. On it's own but also compared to the XTX.

Naming doesn't fit IMO. Should've been 7800XT and 7900XT, although I think AMD can't go below that $899 price tag without sacrificing margins and that's why it's called 7900XT and not 7800XT.

 

[TESKATLIPOKA]

Just because 7900 doesn't hit 3Ghz, doesn't mean the smaller chips won't. Their comment doesn't really state that ALL RDNA3 chips will hit 3Ghz.

True, but the question is what is the ceiling for RDNA3? They say 30% higher clocks but compared to what?
RDNA2 ended at 2815 MHz for 6nm 6500XT excluding the OC models.
For 7nm the highest clock was 2635 MHz for 6650XT excluding OC models.
If you apply 30% higher clocks to them, then that is 3426-3660 MHz.
Can N32 clock to 3.3GHz? Can N33 clock to 3.4GHz? If they manage ~3GHz then that's only ~15% better frequency compared to N22(N23).
Then there is still at what TBP.

 

[Kaluan]

That would seem to imply that the headroom is there. Maybe AMD is artificially limiting performance of partner cards prior to launch? Just a guess.

It would be neat to see a partner card hit 2.8-3.0ghz. I would be curious as to what the performance would be.

Well, reference with its ~366W of usable sustained power certainly won't go too far, but customs with up to ~510W certainly may surprise. I hope we have both reference and custom on review/launch day. All custom AIB listings have [redacted] in the clock specifications. Not sure if this is common behavior in pre-launch listings.

I do wish we could get some benchmarks soon, rather than right at launch. I hate the launch day review rush.

I am only even considering upgrading my GPU because of one reason: power efficiency. At a lower TGP, the 7900XTX looks to be faster than my 3090. My only reservation is RT performance: will it be competitive with the 3090 (non-Ti)? If so, I will upgrade assuming the GPU fits in my case and is relatively quiet.

A bit of info surrounding RT, Note that AMD is claiming up to 1.8X RT performance on “heavy RT workloads” as per this slide:

View attachment 71147

The "up to 1.8-1.84x" claim is weird, since even going by their own claims/slides, we can clearly see a "2x" case in Dying Light 2:




At least in this title ("heavy RT"?), RX 7900 may be quite a bit faster than your 3090.

 

[Saylick]

Based on the frequency/power chart you posted, RDNA3 gain 30% higher frequency at ISO power or 50% lower power consumption at ISO frequency.
RX 6950XT has 335W TBP and from that ~250-275 W is sorely for GPU, right?
If I made an RDNA3 GPU with comparable specs to RX 6950XT, I should shave off 125-137.5 W, the new TBP would be only 198-210W.
We have a pretty comparable RX 7900XT with 84CU and 2GHz for gaming and 2.4GHz for boost clocks, yet It has a 300W TBP.
Based on that chart It should have lower TBP or higher clocks, yet It doesn't have either. Why? Maybe these rumors about a problem with design are true.
I have to wonder If 450W would be enough for 3GHz OC models.

According to Igor, he estimates the GPU power as 230W for the 3090, which has a 350W TBP. If we apply a similar approach to N31, I suspect we'll get a similar GPU power as well. Perhaps 250W or so?

Either way, it does seem like we should get some kind of power consumption reduction with 2300 MHz boost clocks. We'll just have to see what reviews show as the true sustained game clock. If it can do 2.5 GHz reliably within the 355W reference TBP, I think 20% higher clocks for likely 85% of the power is on the money, especially considering that 85% needs to be scaled back up by 20% to account for 20% more CUs.

 

[TESKATLIPOKA]

I think the worst heavy RT game for AMD is Cyberpunk 2077. Control with RT enabled is also pretty bad for AMD from the games TPU tested.