Major graphics card manufacturer Nvidia recently unveiled what it claims is the most advanced data center accelerator to date, the Tesla P100 graphics processing unit. The new GPU is based on the company's new Pascal GP100 architecture which was designed to combine both cutting edge performance and efficiency in order to deliver top of the line computing power.
Nvidia co-founder and chief executive officer Jen-Hsun Huang said in a statement, "Our greatest scientific and technical challenges - finding cures for cancer, understanding climate change, building intelligent machines - require near-infinite amount of computing performance. We designed the Pascal GPU architecture from the ground up with innovation at every level."
The Tesla P100 GPU is considered to be the largest chip in the market manufactured using the FinFET design. Mass production of the chip started on April 5.
The Tesla P100 GPU is built using the 16nm node fabrication technology. The chip contains 15 billion transistors, 16GB of HBM2 memory and supports NVLink interconnect. According to The Register, including the memories, the Tesla P100 GPU contains 150 billion transistors in total.
Nvidia claims that the Tesla P100 can perform 5.3 teraflops calculations per second using 64-bit floating point numbers. Using 32-bit floating point numbers, the chip can perform 10.6teraflops. When using 16-bit floating point numbers it can perform 21.2 teraflops. The chip also has 4MB of L2 cache and a block of 14MB of shared memory.
Just a day after the Tesla P100 GPU was unveiled, Nvidia said that the chip will be used for an upgrade to a supercomputer called Piz Daint, according to PC World. Around 4,500 units of the Tesla P100 GPU will be installed to the supercomputer which is used at the Swiss National Supercomputing Center in Switzerland. As the seventh fastest computer in the world, the Piz Daint has a maximum performance of 7.8 petaflops.
Aside from powering supercomputers, Nvidia also said that it has already made a desktop style supercomputer with the Tesla P100 at its heart. The computer is called the DGX-1 and can deliver 180 teraflops of performance. The DGX-1 will be released in June and companies providing cloud computing services will start offering online services based on the hardware before the end of the year.