Next GeForce series won't be Volta but Pascal-Refresh

[Vya Domus]

Reminds me about that similar toy money craziness few years ago.

I guess there's no need to upgrade from my 970 yet.

Meh a 970 is still plenty fast really , just watched a couple of videos from Digital Foundry where they take a 970 and run games at 4K on it.

 

[Ruru]

Meh a 970 is still plenty fast really , just watched a couple of videos from Digital Foundry where they take a 970 and run games at 4K on it.

Yup no problems with 1080p/60.

 

[R-T-B]

Kepler actaully is a more versatile compute card than Maxwell, because it has double precision compute that is decent. Maxwell's "advance" came by neutering that, because only a handful of compute applications use it. Those that do, perform far worse on Maxwell than Kepler.

 

[qubit]

Kepler actaully is a more versatile compute card than Maxwell, because it has double precision compute that is decent. Maxwell's "advance" came by neutering that, because only a handful of compute applications use it. Those that do, perform far worse on Maxwell than Kepler.

Didn't NVIDIA fix this deficiency for Pascal?

 

[Nuckles56]

Didn't NVIDIA fix this deficiency for Pascal?

I believe that was only for the GP100 based cards, the GP102 and lower are still neutered for fp64 I believe

 

[qubit]

Yeah, that sounds about right.

 

[TheGuruStud]

Plot twist: Volta IS Paxwell with all the compute packed on and that slight shrink, so they get a couple dies per wafer. Consumer ver has all of that stripped, so it's just Pascal in the end.

 

[Deleted member 67555]

Good... What Nvidia did with this generation was spectacular to say the least.... They deserve to get paid for awhile.

 

[R-T-B]

Didn't NVIDIA fix this deficiency for Pascal?

No. It's worse if anything.

 

[Nokiron]

Kepler actaully is a more versatile compute card than Maxwell, because it has double precision compute that is decent. Maxwell's "advance" came by neutering that, because only a handful of compute applications use it. Those that do, perform far worse on Maxwell than Kepler.

It's versatile, sure. But general compute is so much worse that you don't want it either way. If you are using loads that benefit from using double precision you use a Titan or Quadro, not a 780 Ti. (IF you do, that is basically self-flagellation)

Maxwell and Pascal runs FP64 at 1:32
780/780 Ti runs FP64 at 1:24 (Titan runs at 1:3)
Fermi runs FP64 at 1:8

(Which means a GTX 580 runs FP64 just as fast as a GTX 780)

Fermi>Maxwell/Pascal>>>>Kepler

No. It's worse if anything.

No, it's exactly the same rate.

 

[R-T-B]

It's versatile, sure. But general compute is so much worse that you don't want it either way. If you are using loads that benefit from using double precision you use a Titan or Quadro, not a 780 Ti. (IF you do, that is basically self-flagellation)

Maxwell and Pascal runs FP64 at 1:32
780/780 Ti runs FP64 at 1:24 (Titan runs at 1:3)
Fermi runs FP64 at 1:8

(Which means a GTX 580 runs FP64 just as fast as a GTX 780)

Fermi>Maxwell/Pascal>>>>Kepler

No, it's exactly the same rate.

Yes, a little research confirms all this. I still don't see how you think the removal of FP64 in maxwell improves things though, it doesn't frankly (for compute anyways). All it does is boost one type of compute at the expense of another, but I will let this go.

 

[Nokiron]

Yes, a little research confirms all this. I still don't see how you think the removal of FP64 in maxwell improves things though, it doesn't frankly (for compute anyways). All it does is boost one type of compute at the expense of another, but I will let this go.

I didn't and have never said that. It is not really relevant though since you absolutely do not want to run FP64 on a 780 Ti (as a hobby, testing or something, sure).

But as I said before, Kepler is a lousy architecture for compute. FP64 is it's only redeeming factor but only on the Titan and Quadro.

 

[Duality92]

But guise, remember that Pascal is a Maxwell refresh (and improvement), so even if this is true, it doesn't mean we can't expect cards with more performance. (like we saw from maxwell to pascal, even though the refresh status)

 

[Vya Domus]

The jump from 28nm to 16nm was huge and allowed for larger dies at higher clocks. The jump to 12nm is not so big.

 

[Vayra86]

Plot twist: Volta IS Paxwell with all the compute packed on and that slight shrink, so they get a couple dies per wafer. Consumer ver has all of that stripped, so it's just Pascal in the end.

And Pascal is actually just Maxwell on a smaller node and with updated voltage regulation.

We come full circle and in 2020 we conclude that we've been served ancient tech for the better half of ten years.

 

[TheoneandonlyMrK]

For me the question is, with an evolved effecient design such as volta ,they of course evolve what they have for a new node so to me it will be a refined ,refreshed Pascal on a new node but , what clocks are they aiming for , this generation clocks high at stock and can be pushed quite high frequently to 2ghz ,a really awesome achievement given the complexity of the chip.
So what next ,???2Ghz @stck , efficiently ,even a refresh could actually be something awesome for nvidia.

 

[RejZoR]

More shaders, slower clocks. They have the headroom to do that, even in terms of power consumption. Clearly, clocking past 2GHz doesn't seem realistic for factory GPU's to be honest.

 

[Vayra86]

More shaders, slower clocks. They have the headroom to do that, even in terms of power consumption. Clearly, clocking past 2GHz doesn't seem realistic for factory GPU's to be honest.

Agreed. I don't see clocks going much higher out of the box than they are now - 1.6-1.8 Ghz it will be and GPU Boost handles the headroom available anyway. The reasons Nvidia could improve clocks were architectural tweaks and 16nm and those can't happen again.