RDNA3 Efficiency [Spinoff from RDNA4]

[CarstenS]

a) N31 isn't more efficient than AD102/103 at any point so guru3d chart hits the fucking trashcan (did you really dump the data without any sanity checks?).
b) View attachment 10936
iso config (same WGP count same membw) against more reasonably clocked N21 part (6800 runs ~1v versus .9v for 7800XT here) PPW bump is 15% aka the shrink.
Worse, this is CP2077, one of the two good cases for RDNA3.
Juice 1v into 7800XT and watch the ppw delta evaporate.

Did you find some data proving *your* point in the meantime? Because that's what you do in a discussion: Make a point and then deliver data to back that up. Doing the opposite has another name tagged onto it and I'm sure you don't want to do what *that* name implies.

 

[Arun]

Every single voltage that has been mentioned here is wayyyyyyyy above TSMC N4 Vmin! The exact number probably depends on how you implement/power your SRAM arrays + power circuitry (and associated cost trade-offs), and I'm not really expecting AMD/NVIDIA desktop cards to run at e.g. 0.65v, but there's obviously still an opportunity to go lower than 0.9v if they were willing to sacrifice perf/$ for the sake of perf/W! From my limited personal experience being involved in competitive analysis of this kinf of thing, I find it's more reliable to compare notebook SKUs which are highly optimised for power to see what the chips are capable of (best case vs best case) as opposed to desktop SKUs where focus on cost vs power is highly variable. Unfortunately it's probably impossible to find the same quality of data for notebook GPUs.

BTW - I'm currently comparing AD102 vs H100 power efficiency and v/f curves, and I'm a bit surprised at just how much higher AD102 clocks at the same voltage (but H100 supported Vmin is way lower). A lot of this comes down to physical implementation as opposed to the high-level architecture, so it's hard to dissociate the two to figure out why a product's perf/W isn't as good as you might expect as an outsider.

 

[LordEC911]

Did some digging since I primarily use TPU #s quite often to refresh my memory on past generations and performance comparisons.

TPU PPW improvements using their numbers-

gtx980 over gtx780 (guideline since Maxwell 2 is praised so highly)
1080p thru 4k
1.45x - 1.47x ppw

7970 over 6970 (crazy good)
1680x1050 thru 2560x1600
1.47x - 1.56x ppw

Hawaii skipped (not the focus)
(no real improvement)

Fury X over 290x (dang good but overshadowed by Maxwell 2)
1080p thru 4k
1.33x - 1.43x ppw

Vega64 over FuryX (outlier)
1080p thru 4k
1.17x - 1.17x ppw

5700xt over Vega64 (crazy good)
1080p thru 4k
1.59x - 1.56x ppw


Looking closer at RDNA2 and RDNA3 data, found some weird stuff in there and ran the numbers myself since they completely changed up their testing/review methodology between RDNA1 and RDNA2.

6900xt over 5700xt (avg)
techpowerup launch#s
1080p - 1.124x ppw
1440p - 1.316x ppw
4k - 1.493x ppw
avg all res - 1.311x ppw

6900xt over 5700xt (good)
comparison during 6950xt review
1080p - 1.237x ppw
1440p - 1.371x ppw
4k - 1.490x ppw
avg all res - 1.366x ppw

6900xt over 5700xt (good)
real launch #s correcting their "weird" data
1080p - 1.249x
1440p - 1.381x
4k - 1.505x
avg all res - 1.378x ppw

6900xt "weird" launch data
perf diff between 6900xt and 3060ti at
1080p avg 1.406x perf
1440p avg 1.502x perf
4k avg 1.564x perf

7 out of 23 (30%) results heavily favors the 3060ti
throw out weird outliers at 1080p 6900xt performs 1.542x better
evenly throw out all outliers at 1080p 6900xt performs 1.480x better
Roughly, RDNA2 performance hindered by ~8-15%.


7900xtx over 6900xt (good)
TPU's "efficiency" chart
1.351x ppw

7900xtx over 6900xt (avg)
using TPU's #s for CP2077
1080p - 1.190x ppw
1440p - 1.300x ppw
4k - 1.397x ppw
avg all res - 1.296x ppw

7900xtx over 6900xt (outlier?)
using all of TPU's #s
1080p - 1.113x ppw
1440p - 1.163x ppw
4k - 1.246x ppw
avg all res - 1.174x ppw


Some pretty big discrepancies with TPU's #s on RDNA2 and RDNA3 depending on what exact numbers they are using behind the scenes that they don't give you.
RDNA1 does drop off a bit at 4k, but RDNA2 not being significantly better than RDNA1 at 1080p was mainly due to the games TPU was using at the time limiting RDNA2's perf.

Edit-
Found a recent TPU review where they have more results in their "efficiency chart"

6900xt over 5700xt
1.46x ppw

7900xtx over 6900xt
1.411x ppw


Their own numbers shows 7900xtx uses ~17% more power on average than 6900xt.
So to reach that ppw improvement, 7900xtx needs to perform 65% better than a 6900xt.
CP2077 @ 4k, 7900xtx is 86% faster than 6900xt.
The difference at 4k across all games, 7900xtx is 47% faster than 6900xt.

 

[trinibwoy]

It's an outlier because they missed their publicly stated PPW targets.

Kinda difficult to follow this convo with all of the flowery & distracting language but I think this quoted bit was the original thesis. But RDNA3 can't be an outlier purely because of missed targets since targets are arbitrary relative to other generations. To prove it's an outlier we can just look at empirical data, i.e. how does RDNA3's improvement in perf/w compare to prior generations.

Easier said than done though as there were so many variables over the last few generations.

Using TPU numbers:

Vega 64, 14nm, 300W (GlobalFoundries w/ HBM)
5700 XT, 7nm, 225W (+56% perf/watt)
6700 XT, 7nm, 230W (+20% perf/watt)
7800 XT, 5nm, 260W (+45% perf/watt)

Doesn't look like an outlier to me.

 

[Bondrewd]

Doesn't look like an outlier to me

You picked the least efficient RDNA2 in no less arbitrary manner so why bother?
Try Phoenix1 (or really anything iso WGP count).

(the AMD metric is biggest config to biggest config anyway, it's Vega10 to Navi10 to Navi21 to Navi31 to Navi48 and so on and so forth).

 

[DegustatoR]

For RDNA3 to be an "outlier" it must land into a position which isn't characteristic of where AMD GPUs usually are in comparison to competition.
From that point of view RDNA3 is anything but an "outlier".
You could argue that RDNA2 was somewhat of an "outlier" but this had a lot to do with the competition as well.

I honestly stopped following the above "conversation" so what exactly is everyone arguing? That RDNA4 will magically do good simply because RDNA3 is an "outlier"? Real world doesn't work this way. This whole discussion seems completely pointless.

 

[trinibwoy]

You picked the least efficient RDNA2 in no less arbitrary manner so why bother?

I picked cards from all generations with similar power consumption to limit the number of variables.

 

[LordEC911]

I picked cards from all generations with similar power consumption to limit the number of variables.

Doesn't that add more variables since you aren't taking into account how far they may have pushed the envelope to compete?
Picking the highest performing bin of the highest ASIC at launch somewhat evens out that variable.

 

[trinibwoy]

Doesn't that add more variables since you aren't taking into account how far they may have pushed the envelope to compete?
Picking the highest performing bin of the highest ASIC at launch somewhat evens out that variable.

Picking the highest performing bin of every generation will result in comparing vastly different sized ASICs with significantly different power consumption. So no, that's not better. You said it yourself "how far they may have pushed". That's a completely unknowable and arbitrary variable.

 

[Bondrewd]

Picking the highest performing bin of every generation will result in comparing vastly different sized ASICs with significantly different power consumption

I can stuff 4 different RDNA2 chops into 230W, but only one die goes to 330W.
It's a bad comparison point since 6800 (n21 chop) is eons more efficient than 6750XT, both are 250W.

That's a completely unknowable and arbitrary variable.

jamming 1.2v is not really arbitrary

 

[LordEC911]

Picking the highest performing bin of every generation will result in comparing vastly different sized ASICs with significantly different power consumption. So no, that's not better. You said it yourself "how far they may have pushed". That's a completely unknowable and arbitrary variable.

Size and power consumption doesn't matter, since we are reducing it down to ppw improvements.
"Pushing the envelope" of a specific ASIC/bin might be somewhat arbitrary but not unknowable when you look at the competetive landscape and the general numbers from history.

 

[Albuquerque]

MOD MODE: I've resurrected a bunch of the deleted stuff from the RDNA4 thread, at least the parts which seemed relevant to efficiency and RDNA3 things. Keep it on topic, please and thank you.

 

[trinibwoy]

Size and power consumption doesn't matter, since we are reducing it down to ppw improvements.
"Pushing the envelope" of a specific ASIC/bin might be somewhat arbitrary but not unknowable when you look at the competetive landscape and the general numbers from history.

If nothing else matters how do you choose which SKUs on which to compare PPW improvements?

I can stuff 4 different RDNA2 chops into 230W, but only one die goes to 330W.
It's a bad comparison point since 6800 (n21 chop) is eons more efficient than 6750XT, both are 250W.

jamming 1.2v is not really arbitrary

6800 xt is 5% more efficient than the 6700 xt (again according to TPU). Not sure how that changes the narrative around RDNA 3. None of the numbers point to it being particularly deficient vs its predecessors.

 

[Bondrewd]

6800 xt is 5% more efficient than the 6700 xt (again according to TPU).

That's a 300W part.
I said 6800.

The gap between the most and the least efficient RDNA2 is 38%, and, funnily enough, they're both 250W TBP!

None of the numbers point to it being particularly deficient vs its predecessors.

Do I have to poast operating voltages again?
Drawing full desktop power at .95v is pretty bad.

 

[LordEC911]

If nothing else matters how do you choose which SKUs on which to compare PPW improvements?

The highest performing part, which I mentioned in a previous post, since it simplifies the comparison.
Again, the whole discussion caught my eye because some are saying AMD "lied" about ppw improvements.

 

[trinibwoy]

The gap between the most and the least efficient RDNA2 is 38%, and, funnily enough, they're both 250W TBP!

6750 xt is an overclocked fully enabled Navi 22, 6800 is a massively chopped and underclocked Nav 21. What's the punchline and what does it mean for RDNA 3 perf/watt? Are you implying we should use the chopped 6800 as representative of RDNA 2?

Do I have to poast operating voltages again?
Drawing full desktop power at .95v is pretty bad.

It's bad according to who? We're looking at delivered results here and RDNA 3's improvements in delivered perf/watt seem perfectly in line with other generations.

The highest performing part, which I mentioned in a previous post, since it simplifies the comparison.
Again, the whole discussion caught my eye because some are saying AMD "lied" about ppw improvements.

Got it. Agree to disagree then. "Highest performing part" isn't in itself a valid signpost for comparing perf/watt across generations especially when those generations targeted very different performance and power levels.

 

[Bondrewd]

Are you implying we should use the chopped 6800 as representative of RDNA 2?

Yes! Because it's literally the only RDNA2 part aggressively-ish downclocked like that.

We're looking at delivered results here

You can deliver PPW imporvements by downclocking even Vega10, doesn't mean it's a viable part.

RDNA 3's improvements in delivered perf/watt seem perfectly in line with other generations.

The devil is in details.
Running low volts at desktop power is bad since the parts that ran low volts to begin with (mobile stuff) get free of charge power bumps across the stack.

 

[LordEC911]

Got it. Agree to disagree then. "Highest performing part" isn't in itself a valid signpost for comparing perf/watt across generations especially when those generations targeted very different performance and power levels.

Yeah, I guess we will have to disagree because I see your comparison showing numbers even farther from reality. You own definition says you can't compared Vega64 to Navi10. So Navi10 has to be compared to Polaris? But not any of the normal ones, it has to be RX590 (because that is a good comparison). Navi10 had 1.786x ppw improvements following your selection process.

I guess for some people the consensus is RDNA2 had below average ppw improvements, RDNA3 had amazing ppw improvements even though no one ever touted it... as well as Maxwell 2's ppw improvements being invalidated due to the parts being compared.

 

[trinibwoy]

I guess for some people the consensus is RDNA2 had below average ppw improvements, RDNA3 had amazing ppw improvements even though no one ever touted it... as well as Maxwell 2's ppw improvements being invalidated due to the parts being compared.

No-one has said any of those things.

 

[LordEC911]

No-one has said any of those things.

No? The RDNA2 vs RDNA3 thing was implied earlier in the thread with people posting ppw comparisons.
Like you mentioned, again, here.

Kinda difficult to follow this convo with all of the flowery & distracting language but I think this quoted bit was the original thesis. But RDNA3 can't be an outlier purely because of missed targets since targets are arbitrary relative to other generations. To prove it's an outlier we can just look at empirical data, i.e. how does RDNA3's improvement in perf/w compare to prior generations.

Easier said than done though as there were so many variables over the last few generations.

Using TPU numbers:

Vega 64, 14nm, 300W (GlobalFoundries w/ HBM)
5700 XT, 7nm, 225W (+56% perf/watt)
6700 XT, 7nm, 230W (+20% perf/watt)
7800 XT, 5nm, 260W (+45% perf/watt)

Doesn't look like an outlier to me.

The Maxwell 2 comment was me exaggerating to point out the issue with selective comparisons.