NVIDIA boss explains difference between Volta and Turing

Posted on Friday, August 17 2018 @ 10:38 CEST by Thomas De Maesschalck
NVDA logo
In response to an analyst's question, NVIDIA CEO Jen-Hsun Huang provided an interesting reply about the architectural differences between Pascal, Volta, and Turing. Basically, he explained that Volta and Turing have different target markets.

Volta is intended for large-scale training, with up to eight GPUs that can be connected, with the fastest HBM2, and other features specifically for datacenters. Turing on the other hand is designed with three applications in mind: Pro Visualization, video gaming, and image generation that uses the Tensor Core.

NVIDIA is also keen to point out that Turing will be able to attack the movie rendering industry in a large way. Up until now, that market segment was dominated by CPU rendering farms, but the higher memory capacity of the latest Quadro cards finally makes it possible to switch this to GPU-based rendering.
Jensen Huang:
As a result, comparing Volta and Turing, entering, CUDA is compatible, that’s one of the benefits of CUDA. CUDA, all of the applications that take advantage of CUDA are written on top of cuDNN, which is our network platform to TensorRT that takes advantage -- that takes the output of the frameworks and optimize it for runtime. All of those tools and libraries run on top of Volta and run on top of Turing and run on top of Pascal. What Turing adds over Pascal is the same Tensor Core that is inside Volta. Of course, Volta is designed for large scale training. Eight GPUs could be connected together. They have the fastest HBM2 memories. And it’s designed for datacenter applications, has 64-bit double-precision, ECC, high-resilience computing, and all of the software and system software capability and tools that make Volta the perfect high-performance computing accelerator.

In the case of Turing, it’s really designed for three major applications. The first application is to open up Pro Visualization, which is a really large market that has historically used render farms. And we’re really unable to use GPUs until we now have -- we now have the ability to do full path trace, global illumination with very, very large data sets. So, that’s one market that’s brand new as a result of Turing. The second market is to reinvent computer graphics, real time computer graphics for video games and other real time visualization applications. When you see the images created by Turing, you’re going to have a really hard time wanting to see the images of the past. It just looks amazing.

And then the third, Turing has a really supercharged Tensor Core. And this Tensor Core is used for image generation. It’s also used for high throughput, deep learning inferencing for data centers. And so, these applications for Turing would suggest that there are multiple SKUs of Turing, which is one of the reasons why we have such a great engineering team, we could scale one architecture across a whole lot of platforms at one time.

And so, I hope that answers your question that the Tensor Core inference capability of Turing is going to be off the charts.
Zooming in on the last sentence of his quote, Huang also reiterated the statement that Turing offers 10x higher inferencing performance than Pascal.


About the Author

Thomas De Maesschalck

Thomas has been messing with computer since early childhood and firmly believes the Internet is the best thing since sliced bread. Enjoys playing with new tech, is fascinated by science, and passionate about financial markets. When not behind a computer, he can be found with running shoes on or lifting heavy weights in the weight room.



Loading Comments