Professional accelerators Nvidia Tesla P4 and P40 based on the Pascal GPU
Nvidia has presented professional accelerators Tesla P4 and P40. Both models are based on GPU generation of Pascal, but much different.
Some time ago, Nvidia divided the accelerators Tesla, which primarily focuses now on systems deep learning into two categories: in fact, maps for learning and inference. The new adapters are just to the second category.
Let’s start with the technical characteristics. Tesla model P4 is focused on blade servers and is based on GP104 GPU with 2560 cores CUDA. That is analog desktop GeForce GTX 1080. Work frequency is much lower 810-1063 MHz for the core and 6000 MHz for 8 GB of GDDR5 memory. As a result, the performance (FP32) stands at 5.5 TFLOPS. But Nvidia decided that for these cards is also important to the performance of the instructions of logical inference on the basis of operations eight-bit (INT8), so that new adapters Tesla has received specialized support high-speed operations INT8. So, for Tesla P4 performance stated at 22 TOPS (trillion operations per second). Interestingly, TDP is 50 or 75 watts depending on the execution, and the card itself uses a compact, low profile circuit Board.
Accelerator Tesla P40 founded on the GPU GP102 with 3840 cores CUDA. In this case the core operates at frequencies 1303-1531 MHz and 24 GB of GDDR5 memory at the frequency of 7200 MHz. The production capacity reaches 12 TFLOPS (FP32) or 47 TOPS (INT8). Power consumption is in the range of 250 watts. Both models feature passive.
Along with Nvidia cards and introduced a new specialized software. For example, the library TensorRT created for optimization of deep learning models for the development of production.
DeepStream SDK allows cards with Pascal GPU simultaneously decode and analyze up to 93 streams video in HD resolution in real-time.