According to engadget, in the early hours of Beijing time, Intel held the 2019 Artificial Intelligence Summit in San Francisco, officially launched two Nervana neural network processors for training and reasoning, and unveiled a new generation of Movidius Myriad vision processing unit.
It is reported that the two Neurona neural network processors include the training-oriented NNP-T1000 and the reasoning-oriented NNP-I1000. This is also the chip that Intel developed after acquiring the two machine learning companies of Nervana Systems and Movidius Ltd. in 2016. At the meeting, Intel announced that the two chips officially began commercial delivery.
Among them, NNP-T uses TSMC's 16nm process technology, with 27 billion transistors and a total area of 680 square millimeters. Intel said in the August announcement that the NNP-T1000 can handle up to 119 trillion operations per second.
In addition, based on the 10nm Ice Lake processor architecture, the NNP-I1000 has an evaluation efficiency of 4.8 TOPs/W on ResNet50, which Intel calls the most energy efficient chip in its class.
At the same time, Intel also said that companies such as Facebook and Baidu can't wait to deploy these two Nervana chips. According to Misha Smelyanskiy, director of artificial intelligence at Facebook, these chips are key to accelerating the work of artificial intelligence, such as processing 6 billion translations a day.
It is worth mentioning that for the next generation of Movidius vision processing unit, Intel is more confident to compare it with NVIDIA products.
It is said that the visual processing unit code-named Keem Bay can provide 4 times more Tops reasoning than NVIDIA Xavier. When fully utilized, the chip can help customers get 50% extra performance. And this new generation of vision processing unit will be available in the first half of 2020, and you need to wait patiently.