Nvidia Post-Volta (Ampere?) Rumor and Speculation Thread

[DegustatoR]

I dont necessarily agree with Wolfram overall but hes correct in terms of transistor counts. 5700xt is 10.3 billion. The 2070 super is 13.6 billion.

It's not that simple.
2060 is a salvage part based on a cut down TU106 die which in full is a 2070 card. So in case of 2060 it's not correct to say that it's using the same amount of transistors as a full TU106 chip has.
2070 Super is also a salvage part based on a cut down TU104 die which in full is a 2080 Super card. It does feature 13.6 bln transistors but not all of them are active in a 2070 Super card.
What 2070 Super has in common with 5700XT on the other hand are the number of FP32 ALUs - it's 2560 in both cases - and the rated memory bandwidth figure - 448GB/s for both.

So comparing a 5700XT to 2070S does make a lot of sense. But you also have to account for the fact that 2070S sustained clocks are higher than these of 5700XT and also for the fact that 2070S does have a lot of additional features missing from 5700XT and some of them do in fact cost a lot in transistor complexity.

 

[trinibwoy]

I dont necessarily agree with Wolfram overall but hes correct in terms of transistor counts. 5700xt is 10.3 billion. The 2070 super is 13.6 billion.

No he’s not correct in terms of transistor count. How can you compare counts in any sensible way without adjusting for features that don’t exist on Navi?

 

[techuse]

No he’s not correct in terms of transistor count. How can you compare counts in any sensible way without adjusting for features that don’t exist on Navi?

Im not saying you can or should, but from the angle hes coming at of just looking at aggregate performance on todays games relative to transistor count, the 5700xt should be compared to TU106.

 

[CarstenS]

Look at what Degustator said: 2070S is carrying even more dead transistors wrt to general gaming perf than just RT cores and Tensors: The inactive 512 ALUs.

Apart from pricing and product positioning, one should compare RX 5700 XT and RTX 2080 Super, if one would want to compare how much gaming performance each vendor crams into each mm².

edit: abstracting from product level and going by the highest ALU count with the highest clock frequency offered for that count across gaming, professional and mobile, TU106 fares comparatively worse than TU104 and 102 at 17,4 TFLOPS/mm², whereas 104 and 102 are at 20,5 and 21,6 respectively.

Navi 10 OTOH is comfortably ahead at 40,4, but trailing Vega 20 with 43,2.

Best-in-class for older generations where Polaris 30 with 30,7 and GP104 with 28,9.

 

[DavidGraham]

Im not saying you can or should, but from the angle hes coming at of just looking at aggregate performance on todays games relative to transistor count, the 5700xt should be compared to TU106.

Factor in Ray Tracing and AI upscaling in that picture, and the 2070 will completely annihilate the 5700XT in performance.

 

[CarstenS]

AI upscaling has nothing to do with performance in games, sorry. Performance is work done per time unit. Upscaling is effectively lowering the amount of work done.

 

[trinibwoy]

Sounds like HPC Ampere is a 128CU, another chip just showed up, with 7936 CUDA cores.

https://twitter.com/x/status/1234861928230653955

How many async engines does Turing have? Nvidia doesn’t talk much about it’s global thread management implementation.

 

[DavidGraham]

New AI upscaling has nothing to do with performance in games, sorry. Performance is work done per time unit. Upscaling is effectively lowering the amount of work done.

For now at least , AI can also be used later to increase resolution (DLSS2X?), accelerate certain algorithms .. etc.

 

[CarstenS]

For now at least , AI can also be used later to increase resolution (DLSS2X?), accelerate certain algorithms .. etc.

Yes, and I'm not saying, those transistors are completely useless now and forever. But I think especially current techniques like DLSS - while they might be useful for some users - should not be employed in any kind of performance comparisons outside their own world.

 

[techuse]

For now at least , AI can also be used later to increase resolution (DLSS2X?), accelerate certain algorithms .. etc.

I think its likely that by the time any of those see anything close to widespread use, TU106 will be too slow to use them.

 

[Bondrewd]

Yes, and I'm not saying, those transistors are completely useless now and forever. But I think especially current techniques like DLSS - while they might be useful for some users - should not be employed in any kind of performance comparisons outside their own world.

DLSS is and forever will be stillborn.
2070 does do RTRT (albeit slowly) and VRS, but both are gimmicks until they reach consoles, which is soon™.

 

[troyan]

There are none console supporting VRS and RTRT unlike more than 20 million Turing buyers and millions of Geforce Now user.

 

[Bondrewd]

There are none console supporting VRS and RTRT unlike more than 20 million Turing buyers and millions of Geforce Now user.

They are all irrelevant.
Your bottom line is consoles.
It's all neat snazzy gimmicks until your bottom line supports it.

Oh wait fuck initial 9th gen year at least will be crossgen too.
This makes it even worse.

 

[DavidGraham]

They are all irrelevant.
Your bottom line is consoles.

The avg user keeps his card approx 4 years, this means buyers of current Navi cards will be left in the dust 2 years into the console cycle. Why would anyone buy any navi card now, knowing that future navi cards will support important visual/performance features that currents cards lack?

Also with DLSS, current Turing cards can have longer lives as it boosts performance of current and future titles.

 

[Bondrewd]

Wasn't it an AMD thing to sell you on the future ™ features?
Oh how the times have changed.

The avg user keeps his card approx 4 years, this means buyers of current Navi cards will be left in the dust 2 years into the console cycle. Why would anyone buy any navi card now, knowing that future navi cards will support important visual/performance features that currents cards lack?

Because Turing is woefully underpowered for actually running RTRT stuff, or will be such in 2-3 years when the bottom line adapts to the new consoles.

Also with DLSS, current Turing cards can have longer lives as it boosts performance of current and future titles.

Yeah 2 working implementations@1.5 years.
Hard pass.

 

[DavidGraham]

New Wasn't it an AMD thing to sell you on the future ™ features?

These -right now- are currently working features, no longer future fairy tales stuff.

Because Turing is woefully underpowered for actually running RTRT stuff, or will be such in 2-3 years when the bottom line adapts to the new consoles.

That's purely conjecture on your part, we'll have to see AMD's RT performance first before you come to this conclusion.

Yeah 2 working implementations@1.5 years.

Your loss, it takes time to develop working tech, but once they are developed and well established they become a force to be reckoned with.

 

[Bondrewd]

These -right now- are currently working features, no longer future fairy tales stuff.

No they're not lmao.
Less implementations than I can count with my fingers.

That's purely conjecture on your part, we'll have to see AMD's RT performance first before you come to this conclusion.

Uh, N2x silicon is more than alive.

Your loss, it takes time to develop working tech, but once they are developed and well established they become a force to be reckoned with.

Lmao.
Didn't you say nV is selling you currently working tech not two paragraphs ago?

 

[techuse]

There arent any reasons to believe DLSS will see any uptick in adoption IMO.

 

[DavidGraham]

Less implementations than I can count with my fingers.

10 RTX games + 2 VRS games, with more coming this year. If I count demos and small indie projects there will be more.

Uh, N2x silicon is more than alive.

How does it's RT performance compare to Turing?

New There arent any reasons to believe DLSS will see any uptick in adoption IMO.

I disagree, with the new version, I think there will be.

 

[Bondrewd]

10 RTX games

Out of which only Control is anywhere near worth mentioning.
1.5years, 10 games, one looks anything decent.
Wowsers truly makes me shell out $1200 for 2080ti.

How does it's RT performance compare to Turing?

Real good shit.