NVidia Ada Speculation, Rumours and Discussion

[CarstenS]

It is more efficient than Ampere at 450W.

Ampere was well past it's sweet spot: https://www.pcgameshardware.de/Gefo...nders-Edition-kaufen-Benchmarks-1392446/3/#a4
"Im Falle der Founders Edition rechnen wir mit 455 Watt Durchschnittsverbrauch, womit die RTX 3090 Ti auf miserable 4,55 Wpi kommt. Zum Vergleich, die Geforce RTX 3090 erreicht ordentliche 3,78 Wpi bei 350 Watt Leistungsaufnahme, während die RTX 3080 10GB mit 325 Watt auf 4,03 Wpi kommt. Die Radeon RX 6900 XT LC rangiert mit 345 Watt bei 3,86 Wpi, während die RX 6800 XT mit rund 295 Watt und 3,72 Wpi knapp die Effizienzkrone abholt."
(in short: They correlate their performance index with their measured avg. power consumption across a couple of games, arriving at a "watts per index point" (WPI), where lower is better. The portion about 3090 Ti (Ampere at 450 watts) translates to "the RTX 3090 FE clocks in at an abysmal 4,55 WPI. For comparison, the RTX 3090 [Ampere at 350 watts] achieves respectable 3,78 WPI."

 

[troyan]

4090 will be >50% more efficient than a 3090 while using 100W more. For example the 4080 (320W, 30W less) is nearly 60% more efficient in the leaked Timespy score: https://videocardz.com/newz/nvidia-geforce-rtx-4080-16gb-reaches-3-0-ghz-in-3dmark-timespy

An this card clocks to 3000MHz instead of ~1800MHz.

 

[Kaotik]

Look at the small "max" upper left to the 616 watts. I know it's hard to read, but looking at the graph it cannot be min or even avg. View attachment 7151
If explanation is needed: If you reset the counters in GPU-z, for a new avg. etc., the graph also resets. The graph shows very low values in the left-hand portion. So, if 615,8 watts was an avg., current draw would be much much higher, which in turn is highly unlikely. It is further corroborated by the clearly current value of GPU Chip power draw right under it.

If you meant, that it's not just a a short term spike - where's your proof or at least indication that it isn't?

Check my edit which was there before you posted

 

[CarstenS]

Check my edit which was there before you posted

Which it wasn't when I started the post. But good to see we agree here.

 

[Clukos]

Funny, how the marketing narrative catches on so quickly. Suddenly, people (not just here in the thread) do think that 450W is actually not so bad. Smooth move, Nvidia.

We'll see how much it is worth when AMD launches Navi31 and we can compare them both at equal power draw.

Really it just depends on the perf you get for it. Power draw can be fixed by undervolting anyway.

 

[xpea]

Really it just depends on the perf you get for it. Power draw can be fixed by undervolting anyway.

Correct. it all depends where you stand in the voltage power curve. RTX 6000 (Full AD102 with 48GB of ECC GDDR6) has a TDP of 300W only

 

[CarstenS]

Correct. it all depends where you stand in the voltage power curve. RTX 6000 (Full AD102 with 48GB of ECC GDDR6) has a TDP of 300W only

Righty-right, and that makes a comparison of Ada and Ampere at 450 watts pretty meaningless, when that's so clearly way, way over the sweet spot.

 

[DegustatoR]

Righty-right, and that makes a comparison of Ada and Ampere at 450 watts pretty meaningless, when that's so clearly way, way over the sweet spot.

Why meaningless? The sole fact that Ampere is over its sweet spot at 450W doesn't mean that Lovelace is over it at the same wattage.
And even then the comparison is still valid since the power is the same.

 

[Henry swagger]

Correct. it all depends where you stand in the voltage power curve. RTX 6000 (Full AD102 with 48GB of ECC GDDR6) has a TDP of 300W only

Plus the 3090 using power hungry gddr6x.. 3070 vs 3070ti tdp difference just by adding inefficient memory

 

[DegustatoR]

Plus the 3090 using power hungry gddr6x.. 3070 vs 3070ti tdp difference just by adding inefficient memory

3070 vs 3070Ti power difference is from a different power/voltage curve and relaxed silicon quality requirements.

 

[PSman1700]

Per watt Ada gives more performance than Ampere.

 

[CarstenS]

Per watt Ada gives more performance than Ampere.

Hopefully, with TSMC N4 vs. Samsung 8 and a new architecture. But I also hope, this holds true not only in selected benchmarks, were DLSS3 FG or some other Ada-specific features come into play, but also in general raster perf.

 

[PSman1700]

Hopefully, with TSMC N4 vs. Samsung 8 and a new architecture.

Thats usually how performance per watt increases happen yes.

But I also hope, this holds true not only in selected benchmarks, were DLSS3 FG or some other Ada-specific features come into play, but also in general raster perf.

That will indeed be interesting to see full benchmarks and reviews.

 

[troyan]

Hopefully, with TSMC N4 vs. Samsung 8 and a new architecture. But I also hope, this holds true not only in selected benchmarks, were DLSS3 FG or some other Ada-specific features come into play, but also in general raster perf.

Zen4 is less efficient than Zen3 with 100W more power. So it doesnt look so easy to improve efficiency over previous products.

 

[Subtlesnake]

Zen4 is less efficient than Zen3 with 100W more power. So it doesnt look so easy to improve efficiency over previous products.

Zen 4 delivers 95% of its performance at Zen 3 power levels. If the same is true for Ada at 350W vs 450W then it makes the case stronger for lower power products.

 

[Kaotik]

Zen4 is less efficient than Zen3 with 100W more power. So it doesnt look so easy to improve efficiency over previous products.

That's the negative side of competition - the need to push the envelope further and further despite diminishing returns.
Using Eco-mode (105W or 65W depending on default TDP) barely hurts them in real world. In Cinebench you can go down to 65W on 7950X and still run circles around 12900K and 5950X in nT

 

[CarstenS]

Zen4 is less efficient than Zen3 with 100W more power. So it doesnt look so easy to improve efficiency over previous products.

Exactly why it makes 0 sense to compare products that are beyond their sweet spot in the V/F-curve with products that comfortably sit there.
Oh, and btw: R9 7950X vs. 5950X with both @105/142 watts is 170 vs. 157 CBR23 nT per Watt with the fatter AM5 plattform to carry. Comparing just PPTs at 142 watt, R9 7950X is 45% faster (per watt PPT).

Same for Ampere vs. Ada - makes no sense to compare on in it's sweet spot with the other way outside of it.

Zen 4 delivers 95% of its performance at Zen 3 power levels. If the same is true for Ada at 350W vs 450W then it makes the case stronger for lower power products.

Exactly.

 

[PSman1700]

Advances seem not bad at all, far from it.

 

[troyan]

Zen 4 delivers 95% of its performance at Zen 3 power levels. If the same is true for Ada at 350W vs 450W then it makes the case stronger for lower power products.

Its stock vs. stock. It was AMD's decision to release such a product stack. Every released Zen4 product is less efficient than their preprocessor.

RTX4080 and 4090 have much higher clocks while using less power. A 4080 12GB has the same performance numbers as a 3090TI while AD104 is only half the size.

 

[Subtlesnake]

Its stock vs. stock. It was AMD's decision to release such a product stack. Every released Zen4 product is less efficient than their preprocessor.

RTX4080 and 4090 have much higher clocks while using less power. A 4080 12GB has the same performance numbers as a 3090TI while AD104 is only half the size.

The thread topic is not Zen 4 and how efficient the product stack is at stock speeds. We're in a GPU thread evaluating what the stock power levels should be in the first place. In that context, the example of Zen 4 only reinforces the claim that vendors are clocking their products far beyond where they need to be to provide meaningful benefits.