At the beginning of 2016, machine learning was still regarded as a scientific experiment, but now it has been widely used in data exploration, computer vision, natural language processing, biometric recognition, search engine, medical diagnosis, detection of credit card fraud, securities market analysis, voice and handwriting recognition, strategic games and robots. In this short period of one year, the growth rate of machine learning exceeded the expectations of the outside world.
The latest forecast report of Deloitte global points out that in 2018, large and medium-sized enterprises will pay more attention to the application of machine learning in the industry. Compared with 2017, the number of projects deployed and implemented with machine learning will double, and it will double again in 2020.
At present, more and more types begin to enrich the new term “Ai chip”, including GPU, CPU, FPGA, ASIC, TPU, optical flow chip, etc. According to Deloitte’s prediction, GPU and CPU will still be the mainstream chips in the field of machine learning in 2018. The market demand of GPU is about 500000, the demand for FPGA in machine learning task is more than 200000, and the demand for ASIC chip is about 100000.
It is worth noting that Deloitte said that it is expected that FPGA and ASIC chips will be used to accelerate machine learning in more than 25% of data centers by the end of 2018. It can be seen that FPGA and ASIC are expected to rise in the field of machine learning.
In fact, some users who started using FPGA and ASIC chips earlier mainly use them in the inference task of machine learning, but soon, FPGA and ASIC chips will also play a role in module training.
In 2016, the global sales of FPGA chips exceeded US $4 billion. At the beginning of 2017, the report can FPGAs beat GPUs in accelerating next generation deep neural networks In, researchers say that in some cases, the speed and computing power of FPGA may be stronger than GPU.
At present, for example, Amazon’s AWS and Microsoft’s azure cloud services have introduced FPGA technology; Domestic Alibaba also announced cooperation with Intel (Intel) to accelerate cloud applications using Xeon FPGA platform; Recently, Intel has continuously stressed that the data center can adjust the cloud platform through FPGA to improve the execution efficiency of machine learning, audio-visual data encryption and so on.
In addition, although ASIC is a chip that only performs a single task, there are many manufacturers of ASIC chips at present. In 2017, the total revenue of the whole industry was about $15 billion. It is reported that Google and other manufacturers have begun to apply ASIC to machine learning, and chips based on tensorflow machine learning software have also come out.
Deloitte believes that the combination of CPU and GPU has greatly promoted the development of machine learning. If various FPGA and ASIC solutions can also exert enough influence in improving processing speed, efficiency and reducing cost in the future, machine learning applications will make explosive progress again.