NVIDIA company announced the first graphics processor architecture Pascal - GP100, and the first product based on it - Accelerator computing Tesla P100. Features of the new chip is truly impressive, as the performance level , but in the presented video card used a few stripped-down version of the processor.
Thus, the graphics processor NVIDIA GP100 promoted on 16-nm FinFET process technology, has an area of 610 mm2 and consists of 15.3 billion transistors. Chip 60 includes streaming multiprocessors (Streaming Multiprocessor - SM), each of which has 64 CUDA kernel. The novelty has 3840 CUDA cores. The number of texture units is 240. Next to the GPU are four memory chips HBM2 total volume of up to 32 GB, which are connected by a 4096-bit interface.
As for the video Tesla P100, designed for high-performance computing (HPC), it is built on the GP100 graphics processor that uses only 56 SM (3584 CUDA cores) and 224 texture units. In this case, the chip is equipped with 16 GB of memory HBM2 which operates at a frequency of 1400 MHz and has a bandwidth of 720 GB / s. The GPU has a high frequency potential - in the presented video card it runs at a frequency of 1328/1480 MHz. The TDP is 300 watts, which means that the cards with full GP100 will be hotter.
Performance Tesla P100 is 21.2 teraflops of computing half-precision (FP16), 10,6 teraflops of single-precision computations (FP16) and 5.3 teraflops of computing double-precision (FP64). According to the head of NVIDIA Jen-Hsun Huang Tesla P100 are already in the stage of mass production and "will soon be sent to the first customers." The first new products will receive cloud services, and OEM-customers will have to wait until the first quarter of 2017. The cost is unknown.
Related Products :
|