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

[PSman1700]

RTX is two years old by mid 2020, maybe by then a new gpu releases.

 

[Bondrewd]

New Do we even know if Ampere is a successor to Volta or Turing?

It needs to be a successor for both.

 

[DavidGraham]

Do we even know if Ampere is a successor to Volta or Turing?

Turing is effectively a Volta Successor. So Ampere is naturally a Turing Successor.

 

[Pinstripe]

Turing is effectively a Volta Successor. So Ampere is naturally a Turing Successor.

Turing is a derivative of Volta, specifically made for gaming.

With Ampere, we don't know yet whether it's only a Volta successor, or an architecture suited for HPC, AI and gaming combined.

 

[DavidGraham]

Turing is a derivative of Volta, specifically made for gaming.

I disagree, Turing is an evolution of Volta, it improved on it in Tensor functionality and compute (separate FP32, INT32), the only reason NVIDIA never cared to release big Turing for HPC like Volta, is that they are both manufactured at 12nm, there will be little difference between the two.

 

[Kaotik]

Also worth noting is that at least Samsung has stated that they're starting NVIDIA 7nm production in 2020, would be quite tight fit to launch in H1, no matter what the name is

 

[Bondrewd]

Also worth noting is that at least Samsung has stated that they're starting NVIDIA 7nm production in 2020, would be quite tight fit to launch in H1, no matter what the name is

Late June is still technically H1, so all good!

 

[Deleted member 13524]

Also worth noting is that at least Samsung has stated that they're starting NVIDIA 7nm production in 2020, would be quite tight fit to launch in H1, no matter what the name is

Agreed. Similar to how we saw GDDR5Xproduction starting in Q1 2016 with Pascal releasing mid Q2 in very limited quantities, I doubt it'll be any different for nvidia's GPUs made on Samsung's 7nm EUV.

 

[PSman1700]

New nvidia gpus for gaming most likely will be late 2020 or something.

 

[mczak]

I disagree, Turing is an evolution of Volta, it improved on it in Tensor functionality and compute (separate FP32, INT32).

The separate int/fp units are there in volta as well.
There is very little difference between turing and volta as far as the compute cores are concerned. The tensor functionality improvement is more of a tweak. Although there's more changes on the 3d graphics side. Calling it an evolution though seems justified to me though.

 

[techuse]

The separate int/fp units are there in volta as well.
There is very little difference between turing and volta as far as the compute cores are concerned. The tensor functionality improvement is more of a tweak. Although there's more changes on the 3d graphics side. Calling it an evolution though seems justified to me though.

What changes on the 3d graphics side? I thought the only changes were the RT cores and the slight tensor tweaks.

 

[mczak]

What changes on the 3d graphics side? I thought the only changes were the RT cores and the slight tensor tweaks.

Variable rate shading comes to mind, as well as task/mesh shaders.
There's a bunch of other (smaller) gl/vulkan extensions listed for turing too.

 

[techuse]

Variable rate shading comes to mind, as well as task/mesh shaders.
There's a bunch of other (smaller) gl/vulkan extensions listed for turing too.

I thought these were also possible on Volta, just no reason for implementation.

 

[PSman1700]

Geforce 3000 series coming first half of 2020, Intel's GPU's next year too. Great times ahead

 

[nnunn]

Geforce 3000 series coming first half of 2020, ...

Good timing, given Sammy going wide and fat with HBM2: 12 layer TSV

 

[PSman1700]

Good timing, given Sammy going wide and fat with HBM2: 12 layer TSV

They promise more performance for less money.... will see about that. Maybe they have to this time around as competition from AMD is much stronger now, and possibly Intel that enters the market too.

 

[Frenetic Pony]

Good timing, given Sammy going wide and fat with HBM2: 12 layer TSV

That's guaranteed to be for high end corporate customers.

Ampere, assumedly being for games, will probably go more down the road of 1660/5 ti. Which is to say stripping out AI/corporate sim silicon in order to charge lower prices. As a guess I'd say tensor cores will just be removed completely, and fp16 will find its way into normal compute units like AMD has done. "DLSS" might just be dropped quietly alongside it.

That being said is there any confirmation Ampere is actually for gaming, and not another corporate only release like Volta? Nvidia might wait until holidays next year to release any consumer focused cards, as I'd expect their corporate stuff has both plenty of demand and much higher profit margins.

 

[CarstenS]

That's guaranteed to be for high end corporate customers.

Ampere, assumedly being for games, will probably go more down the road of 1660/5 ti. Which is to say stripping out AI/corporate sim silicon in order to charge lower prices increase margins. As a guess I'd say tensor cores will just be removed completely, and fp16 will find its way into normal compute units like AMD has done. "DLSS" might just be dropped quietly alongside it.

That being said is there any confirmation Ampere is actually for gaming, and not another corporate only release like Volta? Nvidia might wait until holidays next year to release any consumer focused cards, as I'd expect their corporate stuff has both plenty of demand and much higher profit margins.

Fixed that for you.

 

[Malo]

Rys, please return the strike-tag...

Click on the + icon, wait don't it'll explode!

 

[ninelven]

As a guess I'd say tensor cores will just be removed completely, and fp16 will find its way into normal compute units like AMD has done. "DLSS" might just be dropped quietly alongside it.

DLSS is and always has been dumb as hell, but tensor cores can potentially be quite useful for consumers in general....