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

[CarstenS]

With chips (well, cards) already selling in the 5-digit-range and the hype for so-called AI still going strong, I doubt that cost would be a prohibiting factor. Especially if density scaling is really north of 1.75× compared to 7nm and you have a good harvesting scheme. Willingness for risk might be a larger factor.

 

[PSman1700]

https://www.pcgamesn.com/nvidia/ampere-33-tflops-gpu

16TF @ 1100mhz, 27TF @1800mhz. 30+TF probably within a year or two?

 

[techuse]

https://www.pcgamesn.com/nvidia/ampere-33-tflops-gpu

16TF @ 1100mhz, 27TF @1800mhz. 30+TF probably within a year or two?

No chance for any consumer GPU.

 

[Kaotik]

https://blogs.nvidia.com/blog/2020/03/02/gtc-san-jose-online-event/

NVIDIA cancels GTC due coronavirus, replaced by online event

 

[PSman1700]

No chance for any consumer GPU.

No idea, but around 16TF for a 3070/3080 isn't far fetched. If not higher then that.

 

[DegustatoR]

No chance for any consumer GPU.

In a year or two? Why not?

 

[PSman1700]

In a year or two? Why not?

We are already shy of 14TF with the 2080Ti (high end), and 16.5TF with a RTX Titan (enthusiast). Both are consumer products, even if the latter is very expensive. These 2018 products will be over two years old by late 2020, three to four years by 2021/2022.

A 20TF Turing @7nm would be a monster already. Anything close to that with ampere is even more impressive.

 

[techuse]

No idea, but around 16TF for a 3070/3080 isn't far fetched. If not higher then that.

In a year or two? Why not?

33 TFLOPs is what i was referring to.

 

[Remij]

We are already shy of 14TF with the 2080Ti (high end), and 16.5TF with a RTX Titan (enthusiast). Both are consumer products, even if the latter is very expensive. These 2018 products will be over two years old by late 2020, three to four years by 2021/2022.

A 20TF Turing @7nm would be a monster already. Anything close to that with ampere is even more impressive.

2080Ti's also boost far higher than 1545Mhz. With a slight overclock the cards boost to 1900-2000Mhz. So it's more around the 16-17TF range for the 2080Ti as it is. The 3080Ti is going to be an absolute beast. I'm sure it will be over 20TF at this point.

 

[DegustatoR]

33 TFLOPs is what i was referring to.

Don't see why this isn't possible in two years on 5nm.

 

[Rootax]

They can have a modest teraflop increase, but better performances... BTW, how do they measure RT cores performances ? With flops too, or some kind of ray per seconds metric ?

 

[techuse]

Don't see why this isn't possible in two years on 5nm.

I dont think we will have a high end 5nm Nvidia GPU in 2 years.

 

[CarstenS]

I dont think we will have a high end 5nm Nvidia GPU in 2 years.

… any more or yet?

 

[techuse]

… any more or yet?

Yet. 2 years from now will be at most 1.5 years from their first 7nm GPU. Just seems far from likely.

 

[trinibwoy]

I’m expecting 4 years or more of 7nm Before we see 5nm parts.

 

[CarstenS]

I don't.

 

[DegustatoR]

4 years of N7+ wouldn't make any sense considering TSMC plans, yeah.

 

[AlNom]

if the price is right...

 

[trinibwoy]

4 years of N7+ wouldn't make any sense considering TSMC plans, yeah.

Lol some people never learn.

 

[psurge]

Isn’t Hopper rumored to use chiplets? Presumably that’s not just to get around reticle limits, but also to enable large transistor counts without having to wait for a process to be mature enough to manufacture large monolithic dies at acceptable yield, and to reduce the number of increasingly expensive tapeouts required for a product generation. Assuming that can be made to work, maybe the 5nm schedule will be more aggressive than one might otherwise expect?