Nvidia recently unveiled the first Pascal architecture powered Tesla cards at the GPU Technology Conference in San Jose, California.
With the P100, the Tesla HPC lineup gets its first significant upgrade since the Kepler cards of 2012. The new GPUs are a consistent performance boost for the Tesla family due to the synergy between the Pascal architecture and the smaller 16-nanometer manufacturing process.
Jen-Hsun Huang, CEO at Nvidia, noted during his keynote that the P100-equipped servers will begin shipping in the first part of next year.
Nvidia announced that partnerships with server system manufacturers such as Hewlett Packard Enterprise, Dell, IBM and Cray are in place, so the servers of the OEMs should come equipped with the Tesla P100 as soon as next year.
However, Nvidia aims to send its high-speed graphic cards to companies that design hyperscale in-house servers and even manufacturing shops.
The GPU will play nice with deep-learning machines, where algorithms work by correlating and classifying huge chunks of data. A myriad of technical areas could benefit from the speed boost, from drones that need to identify objects to self-driving cars and robots. By fitting the gadgets and devices with the P100, the learning time and performance could see a massive improvement.
Nvidia already supplies GPUs for supercomputers: two of the world's best computing machines already run on graphics cards from the manufacturer, according to a report from Top500.
The new P100 brings to the table unprecedented performance that is north of 21.2 teraflops, which is enabled by the device's 15 billion transistors. The OEM used the 16-nanometer FinFET process to reach the peak performance. In addition, the chips are stacked on top each other, which makes room for a lot of features.
The High-Bandwidth Memory 2 (HBM2) packs a bandwidth of 256 Gbps, making it twice as fast as the previous version. Thanks to the novel NVLink interface, data transfers of up to 160 Gbps are enabled. In comparison, PCI-Express cards offer only 20 percent of the speed.
Some tech experts wonder if the servers will play nice with NVLink. IBM claims that its Power architecture is NVLink ready, but Intel's servers require PCI-Express connections to hook up graphics cards to motherboards.
Check out the details regarding Pascal's significant abilities on Nvidia's official developer page.
Previous rumors regarding the ultra-fast graphics card surfaced during this year's CES event, when Nvidia touted the future Tesla GPU.
