The PCIe Tesla P100 GPU for servers from Nvidia is almost ready, and the powerful hardware should land in stores during this year's final quarter.
The OEM says that the latest hardware in its Tesla Accelerated Computing Platform is a radical step up in both performance and power efficiency. The Tesla P100 allows "super nodes," which are the equivalent of over 32 commodity CPU-based nodes. Nvidia estimates that the super-nodes will yield an economy of 70 percent due to reduced capital and operational expenses.
The VP of accelerated computing at Nvidia, Ian Buck, notes that there is a blooming demand for "HPC and AI supercomputing," and providing accelerated computing is the only way to fill it. The Tesla P100 was first announced in April.
"Deploying CPU-only systems to meet this demand would require large numbers of commodity compute nodes," he explains.
This would lead to an explosion in costs but with subpar performance increments.
Buck says that using more Tesla P100-powered nodes will direct the investment toward computing instead of a bulky infrastructure that is expensive to maintain.
The Tesla P100 comes with PCIe support and plays nice with a slew of GPU-accelerated servers.
Double-precision performance for the Tesla P100 shows 4.7 teraflops, while single-precision performance goes as high as 9.3 teraflops. When the Nvidia GPU Boost technology is put to use in half-precision performance, the server roars at 18.7 teraflops.
You can order the Tesla P100 in two models: one comes with a memory bandwidth of 720 GB/s and 16 GB of internal memory. The second one offers 540 GB/s while holstering 12 GB of internal memory.
In May, Nvidia caused quite some hype with the release of its GeForce GTX 1080 GPU. Benchmarking showed that the GPU ranked higher than previous GeForce top models GTX TITAN X and GTX 980 Ti by no less than triple efficiency. The GeForce GTX 1080 does pack 9 teraflops of oomph and 8 GB of GDDR5X memory, after all, so the results were not that surprising.
For the first time, Nvidia embedded its proprietary Pascal architecture into the GeForce GTX 1080. The OEM affirms that the Pascal was designed to power up computers that learn, see and simulate the world.
The first step in harnessing more power from its GPUs was to switch to a 16nm FinFET chip construction. By having a tinier chip design, the GPU uses fewer watts and heats up less. This directly translates into higher clock speed of the GPU, leading to a substantial increase in power.