During its annual GPU Technology Conference, Nvidia has unveiled its new Tesla P100 GPU high-performance computing (HPC) accelerator based on Pascal GP100 GPU.

The GP100 on the Tesla P100 HPC accelerator is a 610mm² GPU built on TSMC's 16nm FinFET manufacturing process and featuring 15.3 billion transistors. It also comes with 16GB of the 2nd generation High Bandiwidth Memory (HBM2) and offers 5.3 TFLOPS of FP64, 10.6 TFLOPS of FP32 and 21.2 TFLOPS of FP16 compute performance.

The GPU on the Tesla P100 HPC accelerator comes with 56 streaming multiprocessor (SM) units, which adds up to a total of 3584 CUDA cores. Bear in mind that the fully-enabled GP100 packs 60 streaming multiprocessors, which is 3840 CUDA cores.

The GPU is made on TSMC's Chip-On-Wafer-On-Substrate technology, which allows them to put the GPU and HBM2 DRAM on the same interposer. The 16GB of HMB2 is paired up with a 4096-bit memory interface and with memory clock of 1.4Gbps, it means that GP100 on the Tesla P100 has a rather impressive memory bandwidth of 720GB/s.

Nvidia was also keen to note that Tesla P100 supports NVLink, featuring 4 NVLink controllers. This allows the GPUs to connect to each other or the CPU and offer much higher bandwidth compared to the PCI-Express 3.0 interface.

According to Nvidia, the Tesla P100 HPC accelerator is currently in volume production and should be available in OEM HPC servers in the first quarter of next year.









Source: Nvidia.com.