First Nvidia Tesla P100 Bechmarks Show Giant Leaps In Everything!
At GTC 2016, Nvidia unveiled their most advanced compute system, the Tesla P100. Powering this latest compute monster is the Pascal GP100, which is the company’s fastest GPU yet.
Touted as the largest FinFET chip on the market, Nvidia Pascal GP100 GPU features 15.3 billion transistors on a 610mm2 16nm die, 16GB of HBM2 memory, and the NVLink interconnect. The Tesla P100 features a slightly cut down version of GP100 GPU, delivering 5.3 TFLOPS using 64-bit floating-point numbers, 10.6 TFLOPS using 32-bit, and 21.2 TFLOPS using 16-bit.
Further, the GPU has 4MB of L2 cache, and a block of 14MB of shared memory. For clock speeds, Pascal GP100 runs at 1328 MHz core and 1480 MHz boost clock – which is no doubt an incredible leap for that small a chip. Nvidia claims the GPU delivers 65 percent high speed, around 2 times the transistor density increase and 70 percent less power than its 28HPM tech.
Besides revealing the above details at its Keynote, Nvidia also showed the first Tesla P100 benchmarks compared to its predecessors. As it turns out, the new Tesla P100 models are always positioned well ahead of the previous gen chips, be it a K40 or M40.
At the same time, the performance delivered by Pascal based P100 offers max scalability up to 8x the current K80 GPU.
The chip maker also illustrated the cross sectional view of the entire interposer of the P100 solution, including the GPU, memory and others.
As shown in the image above, there are basically four stacked memory chips lying beneath a much larger Spacer, featuring 16GB of stacked HBM2 VRAM for a total of 720GB/s of bandwidth. This innovative approach to memory design is termed as Chip on Wafer on Substrate (CoWoS) which provides a 3x boost in memory bandwidth performance, compared to the Maxwell architecture.
The Tesla P100 GPU is aimed at hyperscale data center workloads crunching deep-learning AI and HPC apps. Servers with the chips are set to hit the shelves in Q1 2017, while Nvidia’s DGX-1 supercomputer, which also uses the GPU, is due out in June.