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

We're not getting that until 2018 for consumers

How would you guys develop Cuda 9 support for Volta if the consumer variants don't come out till next year possibly (or late this year....)

Would the Uni purchase a system or would your project have to provide the cash to purchase?

or like Pascal, wait until the consumer variants are out and program accordingly?

'Cause I was thinking one of these would work wonders (wish I had that kinda cash to splurge myself - maybe a gofundme to get started.

https://www.nvidia.com/en-us/data-center/dgx-station/#source=pr

Kinda makes my Titan X Pascal a bit redundant now.

Even the HBM2 memory will be in short supply until some time well into 2018. It is estimated that AMD will be able to deliver less than 16,000 VEGA cards this year due to that. Maybe a few high-end Voltas will appear this year, but not in a price range for most GPUGrid crunchers.

https://www.nvidia.com/en-us/titan/titan-v/

Volta Titan (GV100) now up for sale...

6GPC * (80 SMV) @64 FP32 cores
5120 CUDA
12288 MB HBM2
21.1 billion 12nm transistors

Could be 2x faster than a GTX1080ti at FP32 for here.

Possible >6Tera FP64 which is equal to GTX980ti (GM200) FP32 or 4 OG titans FP64.

Nvidia is asking 3 times cost of the OG titan.

An OG Titan ACEMD 8.0CUDA runtime on par with GTX 1060 / GTX 980.

Will the Volta Geforce GTX series release within 8 months...

I still think that there's no point in buying TITAN series cards for crunching GPUGrid, as this series is overpriced so much that you can buy two gaming (GTX series) cards instead, and have 1.9 times crunching power for the same money.

A Linux GTX 1070 system is nearly as fast as WDDM GTX 1080ti system due scaling issues so a Titan V will certainly want Linux for ACEMD to perform fully.

Considering Tesla GV100 cost is 10k usd some might look at 3k Volta as a compute bargain.

Volta feature set only adds to already impressive Pascal.(Volta FP32 and INT cores are now separate from each other.)

CPU these days are stale rehashes of Haswell. Very few programs take advantage of power hungry AVX512 on Skylake.

Whatever Nvidia decide to do with next gen gaming series they will leave Pascal (Maxwell) in the dust at any program. As Maxwell did to Kelper.

Every generation NVidia lowered their per core or since Maxwell (32 core block warp) energy usage allowing more cores to be added.

Could be 2x faster than a GTX1080ti at FP32 for here.

What makes you say that? It's 13.8 TFlops for a stock Titan V against 11.3 TFlops for a stock 1080Ti. The Titan should have a higher OC headroom due to its lower base clock, but is limited to +20% power.

MrS

---

Scanning for our furry friends since Jan 2002

I still think that there's no point in buying TITAN series cards for crunching GPUGrid, as this series is overpriced so much that you can buy two gaming (GTX series) cards instead, and have 1.9 times crunching power for the same money.

I agree, especially for FP32 projects. Some of the older Titans used to have better FP64 ratios but now they are priced so high its not worth it.

We have already received a Volta GPU from Nvidia.

We have already received a Volta GPU from Nvidia.

Nice! I'm glad Nvidia care so much about science. Is it the V100 or a gaming variant with more FP32 meant for future release?

If it was the latter he wouldn't be allowed to tell us. Besides, the current rumors say there is not going to be any "gamer Volta". They're going to be called Ampere and released in Q2 2018. Very probably it's just a name change, though.

MrS

---

Scanning for our furry friends since Jan 2002

it's a Tesla V100

it's a Tesla V100

FYI: A new Volta v9.1.85 CUDA toolkit is available.

Previous CUDA Toolkit v9.0.176 released September 2017.

388.59 (windows) driver has v9.1.104 CUDA.

Does the new CUDA toolkit help in any way for GPUGRID crunching?

Does the new CUDA toolkit help in any way for GPUGRID crunching?

This is a good question for admins..
There a chance (or not) CUDA9.1 is faster. GPUGRID project has seen previous (ACEMD app) performance increase with CUDA upgrades.

388.59 (windows) driver has v9.1.104 CUDA.

388.71 officially supports v9.1 CUDA in release notes.
The .59 driver does not - a reason why Volta Titan CUDA review benchmarks were inconsistent or not working.