According to Engadget, deep learning applications such as computer vision and speech recognition typically require matrix calculations on large arrays, which are not a long term for common chips such as Intel Core or Xeon. Intel expects to fill the AI gap with the upcoming NNP , And invited to engage in deep learning and AI exhibition Facebook involved in chip design.
In addition to Facebook's strong social media applications, Intel plans to roll out its AI chips to include healthcare, automotive and meteorology.
Nervana NNP is a purpose-built integrated circuit (ASIC) that enables training and execution of deep learning algorithms with extreme computational efficiency. Intel discards common caches on the CPU and uses special software chips for specific algorithms Memory management, hoping to take this chip computing density and performance to a new level.
Nervana NNP can also support a large number of bi-directional data transmission through high-speed interconnection inside and outside the chip, and if you connect multiple NNP chips, you can form a huge virtual chip to cope with the scale of the deep learning module.
It is worth mentioning that Nervana NNP uses a less accurate Flexpoint format.Naveen Rao, vice president of Intel AI, said neural networks have a fairly high level of tolerance for data-based noise and can even assist The neural network brings together new solutions, and the lower accuracy of Flexpoint helps to increase system parallelism, reduce latency and increase bandwidth.
Before Intel invested in AI development, NVIDIA took the first step through the parallel computing power of the GPU to capture the market, but the GPU was good at algorithmic training rather than execution.On the other hand, Intel's biggest rival, Qualcomm, AI program chip development.
Intel's NNP chip also targets training and execution of AI and will release new versions in succession, in addition to a new neuromorphic chip called Loihi, as well as the Myriad X R & D.
Just as Intel struggled to catch up, NVIDIA also launched the V100 for AI apps and took Clement Farabet as vice president of AI Architecture, hoping to enhance the ability of its chips to execute deep learning programs.
In the meantime, Google has created a Tensor Processing Unit (TPU) chip for its data center applications and IBM has released a neuromorphic chip called True North.