Nvidia Volta Speculation Thread

[Kaotik]

Why would they change it when it was apparently good enough for a high end HPC focused GPU and Turing is far more consumer oriented? We know that the Tensors are different in Turing?

But is it a case of "it's good enough" or "it's better than anything else we have" or "it's good enough trrade off with current hardware"?

 

[CarstenS]

… or of constant evolution and field testing.

 

[Deleted member 2197]

We know that the Tensors are different in Turing?

Turing added some additional precision modes (INT8, INT4) for inferencing, and experimental 4-bit and 1-bit precision modes for low-precision math research.

 

[A1xLLcqAgt0qc2RyMz0y]

New Nvidia Tesla V100S

https://www.anandtech.com/show/15146/new-nvidia-gpu-variant-at-supercomputing-2019

 

[A1xLLcqAgt0qc2RyMz0y]

New Nvidia Tesla V100S for PCIe Official Specs:

https://www.nvidia.com/en-us/data-center/tesla-v100/#specs

Performance is 16-17% faster than the Original Tesla V100 for PCIe.

Memory bandwidth is up a whopping 26% (1134 GB/s vs 900 GB/s) which is probably the main reason for the 17% overall gain.

 

[Deleted member 2197]

Interesting the Tesla V100S PCI-e version has higher performance/specs than the original Tesla V100 NVLink version.

 

[DavidGraham]

Interesting the Tesla V100S PCI-e version has higher performance/specs than the original Tesla V100 NVLink version.

It's probably using the full die of GV100, ie the full 5376 cores.

 

[iMacmatician]

It's probably using the full die of GV100, ie the full 5376 cores.

The data sheet shows that it still has 5120 SPs.

 

[Andrew222333]

The point is they still increased it by 41.5% when the physical side only increased by 33% while also adding even more functionality and all within same TDP on a massive die and similar 16nm node (albeit latest iteration in TSMC typical fashion called 12nm).
They still do packed accelerated 2xFP16 math in V100 just like P100 btw.
You get 30TFLOPs FP16 and also the Tensor matrix function unit/cores, usually Tensor matrix will have more specific uses primarily towards Deep Leaning framework/apps (future it is in theory possible to use this with professional rendering-modelling, not talking about gaming though).
Those Tensor function units/cores can also be used for FP32 operations as well, so I think that works out around 2x faster with DL supported framework/apps.
Cheers

The semiconductor worlds is going into 18" wafers, this means more dies per wafer, better efficiency in fabs and better utilization of silicon. This die per wafer calculator is able to predict the number of dies per wafers and wafer lot. That is really good news because the price for end user will be lower.

 

[entity279]

The semiconductor worlds is going into 18" wafers [...]

Is it going? Not sure if we have any indication of that..