The new GPU unveiled by Nvidia today promises to boost data center performance speeds by as much as 50 times. The company is touting the Tesla P100 as “the most advanced data accelerator ever."
Nvidia's new GPU is the first to be based on its Pascal architecture. The company said that the architecture will enable a new class of servers capable of delivering the performance equivalent of hundreds of CPU server nodes.
The Tesla GPU will help power an upgraded version of Europe's fastest supercomputer, the Piz Daint system at the Swiss National Supercomputing Center (CSCS) in Lugano, Switzerland. The upgrade is expected to more than double Piz Daint's speed.
Jumping to 16nm Fabrication
Nvidia said the Tesla P100 is designed to address inefficiencies in modern data centers. While today’s data centers are able to process large numbers of transactional workloads such as Web services, they are often not up to the task of providing artificial intelligence and scientific applications, the company said.
The company is hoping that its new Pascal architecture will change all that by making server nodes faster and more efficient. Among Pascal’s features is an increase in neural network training performance of more than 12 times compared to previous generations of Nvidia GPUs based on its Maxwell architecture.
The accelerators also feature the NVLink high-speed GPU interconnect that scales applications across multiple GPUs, accelerating bandwidth up to five times more than current solutions, according to Nvidia. Up to eight Tesla P100 GPUs can be interconnected with NVLink to maximize application performance in a single node, the company said.
The new architecture is also based on a 16-nanometer fabrication process, a significant leap over Nvidia's current 28-nanometer technology. The change should mean a major upgrade in energy efficiency for data center workloads.
New AI Algorithms
The Pascal architecture also unifies processor and data into a single package to deliver increased efficiency in computing. Nvidia said it is also taking a new approach to memory design, which it calls Chip on Wafer on Substrate. The new design will provide a three-fold boost in memory bandwidth performance compared to the company’s existing Maxwell architecture.
Finally, the new GPUs feature new artificial intelligence algorithms to deliver peak performance for deep learning. The company said data centers can expect more than 21 teraflops of half-precision performance, 5.3 teraflops double-precision performance, and 10.6 teraflops single-precision performance thanks to the GPU boost.
"Our greatest scientific and technical challenges -- finding cures for cancer, understanding climate change, building intelligent machines -- require a near-infinite amount of computing performance," said Jen-Hsun Huang, CEO and co-founder, Nvidia. "We designed the Pascal GPU architecture from the ground up with innovation at every level. It represents a massive leap forward in computing performance and efficiency, and will help some of the smartest minds drive tomorrow's advances."
The Pascal-based Nvidia Tesla P100 GPU accelerator will be generally available in June as part of the new Nvidia DGX-1 deep learning system. It is also expected to be available beginning in early 2017 from server manufacturers.
Image Credit: Tesla P100 GPU photo via Nvidia.