bit-tech.net

Nvidia announces Volta-based Tesla V100 card

Nvidia announces Volta-based Tesla V100 card

Nvidia's new Volta-based V100 GPU, first seen in the Tesla V100 accelerator, is claimed to offer a five-fold performance boost over current-generation Pascal parts.

Nvidia has officially launched the full-fat version of its Volta graphics processing architecture, following its appearance in an automotive-focused system-on-chip (SoC) design, but if you're eager to get your hands on it for gaming you're going to have to wait: its first outing is, unsurprisingly, in a high-end accelerator board for the data centre.

The latest processor to come out of Nvidia, the Volta-based V100 isn't your average general-purpose graphics processing unit (GPGPU). As well as 5,120 CUDA processing cores, the chip includes 640 Tensor Cores designed to accelerate deep-learning and other artificial intelligence workloads - the equivalent, the company claims, of spreading such workloads across 100 CPU cores per GPU. Combined with a next-generation NVLink connection boasting twice the throughput of its predecessor and High Bandwidth Memory 2 (HBM2) memory with a 50 percent throughput boost, Nvidia claims each board can offer 120 trillion floating point operations per second (teraFLOPS) of compute performance when the workload can be accelerated through the Tensor Cores.

'Artificial intelligence is driving the greatest technology advances in human history,' claimed Jensen Huang, Nvidia founder and chief executive, during a keynote speech unveiling the architecture at the company's annual GPU Technology Conference (GTC). 'It will automate intelligence and spur a wave of social progress unmatched since the industrial revolution. Deep learning, a groundbreaking AI approach that creates computer software that learns, has insatiable demand for processing power. Thousands of Nvidia engineers spent over three years crafting Volta to help meet this need, enabling the industry to realise AI's life-changing potential.'

While the Tesla V100 card itself is likely of interest only to those building deep-learning systems, the keynote also revealed some key features of the Volta architecture beyond just core count and a 21-billion transistor total. According to Nvidia's internal testing, Volta offers a five-fold improvement in performance over current-generation Pascal in raw compute without the Tensor Cores' help and 15-fold over the two-year-old Maxwell architecture - four times greater, the company crows, than Moore's Law would predict.

More information on the Tesla V100 and Volta architecture can be found on the company's official website.

8 Comments

Discuss in the forums Reply
edzieba 11th May 2017, 13:31 Quote
Quote:
Nvidia claims each board can offer 120 trillion floating point operations per second (teraFLOPS) of compute performance.
It;s important to mention this is if those FLOPs can be packed into 4x4 matrix-FMA operations, which is what the new Tensor Core components of the SMs perform. If you cannot pack your operations thus, you are limited to a 'mere' 15TFLOP FP (8.5TFLOP DP and 30TFLOP half-precision and). Which is still around 50% more than GP100.
Wakka 11th May 2017, 13:34 Quote
It's not like Nvidia can't afford it, especially given what these are going to sell for, but I wonder what the yields are on a chip that bloody mahusive!

Also, to quote someone on another tech site:
Quote:
Nvidia shows us Volta die and specs, AMD shows us Vega logo...

The more time that goes by without AMD jumping up and down shouting "look what Vega can do!", the more I get worried... I mean, they were quite happy to release performance slides for Ryzen showing how they've caught up to Intel, why are they not doing the same this time?
Vault-Tec 11th May 2017, 13:46 Quote
Nvidia shows us Volta die and specs, AMD shows us Vega logo...

And says it's really good or something :D

Edit. Or was it "The performance is nice".
Anfield 11th May 2017, 13:52 Quote
Quote:
Originally Posted by Wakka

Also, to quote someone on another tech site:
Quote:

Nvidia shows us Volta die and specs, AMD shows us Vega logo...

Lol, I still feel duped having watched that whole live stream where they unveiled the logo:(

Don't get me wrong, I still kind of want a Vega card or two, mostly because the two best ultrawide screens on the market use freesync rather than g-sync plus some competition is always good to slow down inflation of prices.
But at this rate (June 2015 Fury X launch, early 2017 logo of its successor unveiled) we will be on the Volta successor before Vega is in stock at stores.
Gareth Halfacree 11th May 2017, 13:57 Quote
Quote:
Originally Posted by edzieba
It;s important to mention this is if those FLOPs can be packed into 4x4 matrix-FMA operations, which is what the new Tensor Core components of the SMs perform. If you cannot pack your operations thus, you are limited to a 'mere' 15TFLOP FP (8.5TFLOP DP and 30TFLOP half-precision and). Which is still around 50% more than GP100.
Yeah, I thought I'd made that clear with the bit on the compute performance improvements over Pascal and Maxwell in a later paragraph, but I've gone back and made it more obvious - thanks!
rollo 11th May 2017, 15:00 Quote
If Volta launches before Vega AMD would be laughed out of the door.
edzieba 11th May 2017, 17:21 Quote
Quote:
Originally Posted by Vault-Tec
Nvidia shows us Volta die and specs, AMD shows us Vega logo...

And says it's really good or something :D

Edit. Or was it "The performance is nice".
AMD may come to regret certain jibes.
Ending Credits 12th May 2017, 14:52 Quote
Nvidia are so far ahead of AMD when it comes to AI/deep learning that it isn't even funny. Maybe the ultra high end stuff for training deep nets is always going to have limited scope, but when you have developers doing all their work using CUDA, Will they all bother to get the same systems working/optimised with other cards?
Log in

You are not logged in, please login with your forum account below. If you don't already have an account please register to start contributing.



Discuss in the forums