Nvidia knows very well how far its Pascal architecture can go, thanks to this today we have items on the market such as the new GeForce GTX 1080 or the recent Tesla range update which, as the company announces, now seeks to position itself as one of the most powerful options for artificial intelligence environments, neural networks and even deep learning environments.
Thanks to this intention, today we can talk about the new Tesla P4 y Tesla P40, which have been created as a generational replacement of those known as the Tesla M4 and Tesla M40 presented just a year ago. According to the company, these new cards exceed up to four times the performance of the previous ones, being ideal even to be used as a single server due to their power and especially their capabilities.
One Nvidia Tesla P4 can replace 13 CPU-based servers
Going into a little more detail, we find the Nvidia Tesla P4, a card that stands out for its low consumption, located between 50 and 75 watts, as well as for having a much more compact size than that of its sister the Tesla P40 . As far as power is concerned, this card is capable of offering 5,5 teraflops, 22 trillion operations per second INT8, 2560 CUDA Cores and 8 GB of GDDR5 memory with a bandwidth of 192 GB / s.
As you can see, we are talking about more than interesting data for a card of this type since, due to consumption and power, they do it, as announced by Nvidia, up to 40 times more efficient than a GPU. In addition to this, the server that is equipped with one of these cards can replace up to 13 CPU-based servers, without a doubt a more than interesting test of what the Nvidia Tesla P4 can offer.
The Nvidia Tesla P40 stands as the top of the range
Second we find the version Tesla P40, an option that would become the top of the range thanks to a computing power of 12 teraflops, capacity to perform up to 47 billion operations per second INT8 thanks to its 24GB of GDDR5 memory with a bandwidth of 346 GB / s and 3840 CUDA cores.
If we put all this into practice, as advertised by Nvidia, a server equipped with eight Tesla P40 cards would be able to replace up to 140 CPU-based servers which, as you can see, would represent a saving in both costs and dimensions quite high. In addition to this, its capacity would be ideal for use in applications related to autonomous driving, robotics and, above all, artificial intelligence systems.
Further information: AnandTech