nvidia GTC 2016:Nvidia Pascal

[Prysin]

like i predicted, this was a rather meh show.

The info, and the number of topics, were really few and thin.

[Samfisher]

With the introduction of the P100, I think I can safely say that all the rumoured specs are wrong.  The P100 has 3584 CUDA cores, and has 56 SM units.  Big Pascal has 60 SM units, so that's a further 256 CUDA cores.  The Pascal architecture uses 64 CUDA cores per SM unit compared to 128 CUDA cores per SM unit for Maxwell.

 

This means Big Pascal with 60 SM units will top out at 3840 CUDA cores, each more powerful than a Maxwell CUDA core.  Based on teraflop numbers, the P100 is 43% faster in FP32 performance than GM200 (Titan X) which is the most important metric for game performance.

[Big-G]

On 6/4/2016 at 5:46 PM, larsi239 said:

I really hope the only reason Nvidia did not announce any new pascal gaming cards is that they did not want to harm the sales of their 900-series, and they want to wait with the announcement right until they are ready to ship the cards - crossing my fingers for computex. I mean, they probably could have said everything about pascal if they wanted to, they even had two pascal modules ready for the drive px 2 system they showed. 

No, it's because GTC is aimed at the HPC, AI and corporate product market... Computex is where the consumer stuff will get shown

 

15 minutes ago, Samfisher said:

With the introduction of the P100, I think I can safely say that all the rumoured specs are wrong.  The P100 has 3584 CUDA cores, and has 56 SM units.  Big Pascal has 60 SM units, so that's a further 256 CUDA cores.  The Pascal architecture uses 64 CUDA cores per SM unit compared to 128 CUDA cores per SM unit for Maxwell.

 

This means Big Pascal with 60 SM units will top out at 3840 CUDA cores, each more powerful than a Maxwell CUDA core.  Based on teraflop numbers, the P100 is 43% faster in FP32 performance than GM200 (Titan X) which is the most important metric for game performance.

Not entirely correct, the specs now given are for Big Pascal -> GP100/P100... They will not use a lower binned P100 for their HPC/IA market and plonk a bigger one in the Consumer market... Generally the one you see at GTC is basically what will end up as the Titan spec for the next Consumer line.

[Samfisher]

2 minutes ago, GidonsClaw said:

 

 

Not entirely correct, the specs now given are for Big Pascal -> GP100/P100... They will not use a lower binned P100 for their HPC/IA market and plonk a bigger one in the Consumer market... Generally the one you see at GTC is basically what will end up as the Titan spec for the next Consumer line.

Source that 56SM's are the new Big Pascal specs?

[Big-G]

6 minutes ago, Samfisher said:

Source that 56SM's are the new Big Pascal specs?

I'm not complaining about 56... I'm saying that what we see at GTC, is Big Pascal

http://www.anandtech.com/show/10222/nvidia-announces-tesla-p100-accelerator-pascal-power-for-hpc

https://devblogs.nvidia.com/parallelforall/inside-pascal/

[Samfisher]

4 minutes ago, GidonsClaw said:

I'm not complaining about 56... I'm saying that what we see at GTC, is Big Pascal

http://www.anandtech.com/show/10222/nvidia-announces-tesla-p100-accelerator-pascal-power-for-hpc

https://devblogs.nvidia.com/parallelforall/inside-pascal/

Nvidia's devblogs support what I said. 

 

Quote

GP100 achieves its colossal throughput by providing six GPCs, up to 60 SMs, and eight 512-bit memory controllers (4096 bits total). Each SM has 64 CUDA cores and four texture units, for a total of 3840 CUDA cores and 240 texture units.

P100 is 56 SMs and therefore isn't Big Pascal. P100 has 3584 CUDA cores.  GP100 will have 3840.  That's the Titan class GPU we will be getting for desktops.

[Big-G]

1 minute ago, Samfisher said:

Nvidia's devblogs support what I said. 

 

P100 is 56 SMs and therefore isn't Big Pascal. P100 has 3584 CUDA cores.  GP100 will have 3840.  That's the Titan class GPU we will be getting for desktops.

Ok, so let's say this is a binned GP100... shall we hazard a guess at the potential next Ti?

[Samfisher]

5 minutes ago, GidonsClaw said:

Ok, so let's say this is a binned GP100... shall we hazard a guess at the potential next Ti?

It's weird.  We don't have a solid precedent for what a Titan should be.  The original Titan had less cores than the 780Ti and only the Titan Black and the 780Ti were the full Kepler core.

 

Then on the other hand, the Titan X is the full Maxwell core while the 980Ti isn't.  If they go the 980Ti route, the P100 is probably what the Pascal x80Ti would look like.