ARM announces a new AI processor "ARM Machine Learning", allowing machine learning processing by terminals without cloud

ARM of semiconductor design is a processor for AI processing "ARM Machine Learning"And the second generation"ARM Object DetectionWe announced. In the future, the process of increasingly increasing machine learning processing on the terminal side rather than the cloud is likely to accelerate at a stroke.

Arm's Project Trillium offers the industry's most scalable, vers - Arm
https://www.arm.com/news/2018/02/arm-project-trillium-offers-the-industrys-most-scalable-versatile-ml-compute-platform

ARM unveils two new AI chip designs to ride the machine learning wave - The Verge
https://www.theverge.com//2018/2/13/17007174/ai-chip-designs-arm-machine-learning-object-detection

Technique for processing AI processing on the terminal side already "DynamiQARM developed two kinds of AI chips, "ARM Machine Learning (ARM ML)" and the second generation "ARM Object Detection (ARM OD)".

ARM OD is a processor optimized to detect human faces and objects. As a technical design, ARM OD, which is the second generation, has real-time detection at full HD · 60 fps and DSP performance up to 80 times higher than conventional. The 2nd generation ARM OD is expected to be used for IoT products such as drone and smart security camera with collision safety function, and will be provided to processor manufacturer by the end of February 2018.

And, the newly announced "ARM ML" is a dedicated processor that accelerates processing of general AI applications such as automatic translation and face recognition, and when using it on a mobile terminal, 4.6 trillion times or more Processing can be executed, and it is said that it has high power efficiency of 2 to 4 times as compared with the conventional one. More versatile ARM ML is expected to be used on mobile terminals such as smartphones, and release within 2018 is scheduled.

Both of the processors do not perform the machine learning process used in the AI ​​technology on the terminal side but the cloud side. At present, machine learning processing that requires high processing capability is mainly performed using the cloud, but because it uses cloud transmission and reception data, it is disadvantageous in response speed and data transfer amount. There is also the risk of intercepting (stealing) when sending and receiving data. On the other hand, if it is a dedicated AI chip specialized for machine learning processing, machine learning processing on the terminal side is possible, which is superior in response and security.

According to Gem Davies, vice president of machine learning technology at ARM, the new chip is a new technology not based on the existing CPU and GPU architecture. We are already using not only as SoC for mobile terminals but also IoT terminal, and it is revealed that consultation with each manufacturer is proceeding.


Apple is equipped with iPhone X / 8 SoC "A11 Bionic"Huawei said"Kirin 970"We adopted a custom chip that can do machine learning processing exclusively, but adoption of AI chip is limited to expensive high-end models. However, ARM ML plans to adopt AI chips for entry level terminals. Mr. Davis has already talked about using AI chips with Chinese manufacturers and cheap entry terminals, and he seems to believe that the AI ​​chip will quickly become a standard specification even at low cost terminals.

Both ARM ML and ARM OD are expected to be used on relatively small terminals, but ARM is planning to expand the scope of AI chips. "According to Google, if all users use the voice search feature only every three minutes every day, Google has to double the number of servers Today's chip design focuses on mobile terminals However ARM's AI chip designer can be extended for servers as well, "Davis says.

According to ARM, development of semiconductors capable of dedicated machine learning processing that can utilize AI technology is a trend of computing, which will greatly change the trend of semiconductor development for mobile that aimed at power saving and high performance is.