Tornado Alley (PROG)

[F6]

NVIDIA Goes Beyond The GPU For AI With Volta

NVIDIA Tesla V100 GPU/Tensor Processor
May 10, 2017
At the company’s annual GPU Technology Conference (GTC), NVIDIA took the giant step toward an accelerator processor architecture customized for artificial intelligence (AI). NVIDIA pioneered the development of artificial neural networks through deep learning with the company’s GPUs and CUDA software platform. While the market is now flush with other solutions, NVIDIA accounts for the clear majority of deep learning networks in use today. The company is engaged with over 40,000 companies and over a half a million developers for neural network applications. However, while the GPU has proven very effective in parallel processing since the addition of shader cores, TIRIAS Research has maintained that even NVIDIA would need to eventually migrate to architectures dedicated to AI while preserving its tools and ecosystem to advance platforms further.
At GTC, NVIDIA announced that step forward with the next generation GPU architecture called Volta. While still referred to as a GPU, Volta is much more. In addition to enhancing the GPU architecture, NVIDIA added 640 new tensor cores capable of processing 4x4x4 matrix multiplies. This provides a specialized math core that works in conjunction with the standard GPU CUDA cores to add additional processing for deep learning environments. It also accelerates the process of inferring a value based on a trained model, making it useful as an inference engine. Essentially, NVIDIA put accelerator cores in an accelerator. This is similar the route Google took in developing its proprietary Tensor Processing Unit (TPU). With the tensor core capabilities incorporated into the NVIDIA SDK libraries and runtimes like cuDNN and TensorRT, developers will be able to take advantage the increase in performance from the tensor cores in their AI frameworks without rewriting their applications.
Volta also includes over 5,000 GPU CUDA cores, 300 GB of system communications bandwidth through six high-speed NV Link interconnects, and 16 GB of the second generation high-bandwidth memory (HBM2) on TSMC’s new 12FFN manufacturing process technology. In all, the new Volta architecture fits in the same power envelope and form factor as the previous Pascal generation GPU with 1.5x the memory performance, 2x the NVLink performance, and 7.5 teraflops of FP64 processing (15 teraflops at FP32) on the GPU (CUDA) cores and a total of 120 teraflops of processing performance with the tensor cores. The first product using the Volta architecture is the Tesla V100. No matter how you look at it, the Volta architecture and Tesla V100 set a new level of computing performance on a chip that will benefit AI. Now imagine what eight of these are capable of in NVIDIA’s DGX-1 platform with the industry’s most mature AI software environment and tools. The Volta architecture and Tesla will also accelerate many high-performance computing (HPC) applications and is already slated for use in the next US supercomputer, the Summit Supercomputer, which is slated to have over 200 petaflops of performance. But, the performance of Volta is only half of the story.
The significance of Volta is that this marks a transition of the most pervasive deep learning engine from a GPU or general processing engine to a more specialized engine for AI. This is a trend that is likely to accelerate and will lead to more innovation in the semiconductor industry. AI is creating a new vector in computing performance that will eventually lead us away from the binary computing solutions we have today to more specialized and revolutionary architectures in the future. While Volta is aimed at the high-end Tesla platform now, TIRIAS Research expects this technology will quickly trickle down to other platforms, such as NVIDIA’s Jetson platform for inference engines that can be incorporated into almost any application from robots to cars. TIRIAS Research projects that within 15 years every new electronics device and system will use some form of AI.
[...]
https://www.forbes.com/sites/tiriasresearch/2017/05/10/nvidia-goes-beyond-the-gpu-for-ai-with-volta/

---

in addition to (linked in) the post to which this is a reply and preceding and (any future other) following, see also (linked in):

http://investorshub.advfn.com/boards/read_msg.aspx?message_id=129743184 and preceding and following,
http://investorshub.advfn.com/boards/read_msg.aspx?message_id=129944214 and preceding (and any future following)

http://investorshub.advfn.com/boards/read_msg.aspx?message_id=130721755 and preceding and following,
http://investorshub.advfn.com/boards/read_msg.aspx?message_id=131201693 and preceding and following