GV100 PCI-e Volta announced at ISC

[the54thvoid]

Nvidia has announced the PCI-e verion of it's V100 top end card. This will be the determinant for all consumer GV10X cards. So volta is arriving later this year, meaning the consumer version is most likely well on it's way as well.

http://www.anandtech.com/show/11559...ces-pcie-tesla-v100-available-later-this-year

Come on @R-T-B, get it while it's fresh. Need to balance those Vega headlines.

 

[FordGT90Concept]

The HBM2 bandwidth is double that of Vega so, 4 stacks instead of 2? That's a lot of CUDA cores but it is reflected in the relatively low clockspeed. More TFLOP per watt compared to Vega too. The advantages can be explained in part by TSMC 12nm versus GloFo 14nm. Still gives me a sad. I hope AMD is investing those Zen earnings into knocking Navi out of the park.

 

[Xzibit]

The HBM2 bandwidth is double that of Vega so, 4 stacks instead of 2? That's a lot of CUDA cores but it is reflected in the relatively low clockspeed. More TFLOP per watt compared to Vega too. The advantages can be explained in part by TSMC 12nm versus GloFo 14nm. Still gives me a sad. I hope AMD is investing those Zen earnings into knocking Navi out of the park.

AMD roadmap isnt doing 12nm. Vega 20 is on 7nm for 2nd half 2018. They do have Vega 10x2 for later this year. I suspect that's their answer instead of doing a half-node 12nm product. Unless things change.

 

[Vya Domus]

The compute performance metric is unimpressive , they have done a conversion that dose not benefit the consumer variants : trading clock speed for core count. That being said I do not believe this will be the final flagship consumer version , it dose not make sense. The yields for this silicon must be horrendous no way they will be selling it for less. A cut down version with higher clocks is more likely , maybe an entirely different chip. Nonetheless I do not expect anywhere near the same jump in performance that was from Maxwell to Pascal.

 

[Fluffmeister]

The compute performance metric is unimpressive , they have done a conversion that dose not benefit the consumer variants : trading clock speed for core count. That being said I do not believe this will be the final flagship consumer version , it dose not make sense. The yields for this silicon must be horrendous no way they will be selling it for less. A cut down version with higher clocks is more likely , maybe an entirely different chip. Nonetheless I do not expect anywhere near the same jump in performance that was from Maxwell to Pascal.

This is no more a consumer chip than the GP100 is. History tells us they strip away things like Tensor cores and double precision and all the other gubbins that aren't relevant to a consumer grade card, and you are left with a GV102, clocked far higher and that blows the doors off everything.

I suspect in the meantime a GV104 will come along first and do exactly that.

 

[StefanM]

A little bit to read...

http://images.nvidia.com/content/volta-architecture/pdf/Volta-Architecture-Whitepaper-v1.0.pdf

 

[eidairaman1]

The HBM2 bandwidth is double that of Vega so, 4 stacks instead of 2? That's a lot of CUDA cores but it is reflected in the relatively low clockspeed. More TFLOP per watt compared to Vega too. The advantages can be explained in part by TSMC 12nm versus GloFo 14nm. Still gives me a sad. I hope AMD is investing those Zen earnings into knocking Navi out of the park.

Navi needs to be a return to form like how the 9700 Pro, 5870 were.

 

[cdawall]

Navi needs to be a return to form like how the 9700 Pro, 5870 were.

Oh I get it you mean you want them to exist! Still waiting for vega. These Volta cards aren't consumer level so judging them for anything other than deep learning is silly