Nvidia new calculation card Tesla P40/P4 configuration how

at the GTC 2016 technical conference held in Beijing, NVIDIA CEO Huang Renxun personally released a new generation of high-performance computing cards Tesla P40, Tesla P4, specifically for artificial intelligence, depth learning, neural network deduction.

They will replace the Tesla M40/M4 of Maxwell's architecture, which have evolved to the latest Pascal architecture and, of course, 16nm processes.

Tesla P40 is based on the high-end core GP100 and opens all 3,840 stream processors (previous Titan x only 3,584), core frequency 1303-1531MHZ,FP32 single precision floating-point computing capability 12tflops,int8 (eight-bit data-specific inference instruction) 47TOPS, the response speed is 45 times times faster than the CPU.

The card is paired with 384-bit 24GB GDDR5 memory, equivalent frequency 7.2Gbps, whole card power consumption 250W, passive heat dissipation.

Incidentally, GP100 big core of the Tesla P100 also only opened 3,584 flow processor, complete is 3,840, with video memory is gddr5x.

Tesla P4 Core is the GTX 1080/1070 GP104, all 2,560 stream processors, the core frequency only 810-1063mhz, single precision performance 5.5tflops,int8 22TOPS, with 256-bit 6Gbps 8GB GDDR5 video memory, power consumption 50-75w, half high passive heat dissipation.

Compared to predecessors, their computational performance has been increased by 70%, 1.5 times times, INT8 characteristics are the first to join, the power consumption is maintained unchanged.