The growth of machine learning is staggering, and FPGAs and ASIC chips are expected to become the new mainstay

At the beginning of 2016, machine learning was still considered a scientific experiment, but now it has been widely used in data mining, computer vision, natural language processing, biometrics, search engine, medical diagnosis, credit card testing Fraud, securities market analysis, voice and handwriting recognition, strategic games and robots, etc. In just one year, machine learning has grown faster than expected.

Deloitte Global's latest forecast indicates that large and medium-sized companies will place a greater premium on the use of machine learning in the industry by 2018. Compared with 2017, the number of projects deployed and implemented using machine learning will double, and will again turn around in 2020 Times

At present, more and more types are beginning to enrich the new term "AI chip", including GPUs, CPUs, FPGAs, ASICs, TPUs, optical flow chips, etc. According to Deloitte's forecast, in 2018, GPUs and CPUs will still be the domain of machine learning Of the mainstream chip GPU market demand is about 500,000 or so, the demand for FPGA in the machine learning tasks more than 200,000, while the demand for ASIC chips in the 100,000 or so.

Notably, Deloitte said it expects FPGAs and ASIC chips to accelerate machine learning in more than 25% of data centers by the end of 2018. It is clear that FPGAs and ASICs are expected to emerge in machine learning.

In fact, some of the early adopters of FPGAs and ASIC chip accelerators mainly used them for machine learning inference tasks, but soon after, FPGA and ASIC chips will also be able to work on module training Play

In 2016, the global FPGA chip sales have exceeded 4 billion US dollars.And in early 2017 report "Can FPGAs Beat GPUs in Accelerating Next-Generation Deep Neural Networks?", The researchers said that in some cases, FPGA Speed ​​and computing power may be even stronger than the GPU.

At present, such as Amazon (AWS) and Microsoft (Azure Azure Cloud Services), have introduced FPGA technology; the domestic Alibaba also announced cooperation with Intel (Intel), using Xeon-FPGA platform to accelerate cloud applications; Intel recently Constantly emphasize that the data center can adjust the cloud platform through FPGA to enhance the efficiency of machine learning, audio and video data encryption and other work.

In addition, although the ASIC is a single-chip implementation of the task, but the current number of ASIC chip maker in 2017, the entire industry's total revenue of about 15 billion US dollars.It is reported that Google and other manufacturers began to use ASIC in machine learning, Chips based on TensorFlow machine learning software have also been released.

Deloitte believes that the combination of the CPU and the GPU has contributed greatly to the growth of machine learning, and if the future of various FPGA and ASIC solutions can also exert enough influence on processing speed, efficiency and cost reduction, the machine Learn the application will once again be explosive progress.