Until now Pascal architecture computing among NVIDIA accelerators offered only series products Tesla P100 , based on the graphic processor and is equipped with a memory GP100 type HBM2. NVIDIA announced the imminent start of deliveries of accelerators Tesla P40 (GP102) and Tesla P4 (GP104).
It is noteworthy that Tesla P40 and Tesla P4 give preference to GDDR5 memory, refusing not only on HBM2, but also on GDDR5X. According to rough estimates , if Tesla P40 is limited to frequencies around 1600/7200 MHz with TDP level not more than 250 W, in the case of the Tesla P4 desire to limit power consumption 50-75 W range forces to adhere to the frequency of about 1100/6000 MHz. At the same time the volume of GDDR5 memory was reduced from 24 GB to 8 GB, respectively, while the number of active stream processors - was cut to 2560 units from 3840 respectively.
A more productive Tesla P40 will be available in October, and more affordable Tesla P4 will have to wait until November. Traditionally, these accelerators will be used in server solutions. NVIDIA recommends to use these solutions in systems of artificial intelligence, including products from the object recognition function on multiple streams of video simultaneously.
If Tesla P40 delivers performance of 12 teraflops, Tesla P4 limited at speed of 5.5 teraflops.
Related Products :
|