Recently, a research group at the Massachusetts Institute of Technology (MIT) has announced a low-power chip for automatic speech recognition that can reduce power consumption by up to 99 percent.
Since the beginning of 2017, the enthusiasm of human-computer interaction has been increasing. In addition to speakers, television and other smart home products, many people will also look to the mobile device: Apple's smart headset Air Pods just a listing on the consumer's attention; LG released last week's smart watch became the first Google's Android Wear 2.0 operating system using a smart watch, built-in intelligent voice assistant Google Assistant; has not yet revealed the real Samsung mobile phone Galaxy S8 also because of the mysterious voice assistant Bixby provoke rumors constantly.
For voice AI on mobile devices, power consumption is always a big problem. Want these voice assistants can listen to the call at any time, you need to always keep the state in the background, real-time monitoring of the surrounding sound. This will inevitably affect the battery life. According to MIT researchers, the existing mobile phone in the voice recognition function of the power consumption of 1 watt.
The current chip needs to keep all the neural networks running, to detect all the sounds and noise. The MIT's new chip uses a "voice activity detection" (voice activity detection) circuit, in the detection of human voice will activate the more complex voice recognition circuit. Therefore, this chip can reduce the power consumption of 90% to 99%, power consumption is only 0.2 to 10 milliwatts.
This technique means that it is possible to use speech recognition and AI assistants on simple small electronic devices. Whether it is mobile phones, watches, glasses or headphones, this chip for their intelligent road and swept away an obstacle.
Chip - another battlefield of artificial intelligence
Artificial intelligence not only means the algorithm, but also the physical hardware behind it. Whether it is voice interaction or identification of images, you need to do a lot of computing to deal with large and complex information. For those who use artificial intelligence, not only value the results (such as recognition accuracy), but also concerned about the speed of operation - if the slow response will greatly reduce the use of experience. Therefore, although the chip low-key, but the cornerstone of indispensable.
In addition, although the current cloud computing is the mainstream of deep learning and artificial intelligence trends, but in some occasions - such as mobile devices - and inevitably in the local processing of data to achieve some real-time features, which requires fast, low energy consumption Of the chip to provide support.
Traditional chip giant transformation AI
NVIDIA is an early chip maker in the field of artificial intelligence. The company, which is known for its graphics processors (GPU), has helped the game become clearer and smoother, and its powerful GPU is well positioned to cope with the computational effort required by the deep learning of neural networks. See the future of artificial intelligence, NVIDIA will be positioned as a "artificial intelligence chip manufacturers", focus on the development of specifically for the artificial intelligence of the chip.
After the acquisition of chip makers Altera, Movidius and Nervana, Intel also announced in 2016 to enter the field of artificial intelligence chips. This year, they will release the first depth of learning for the chip Knights Mill, and the introduction of its acquisition of the company Nervana Systems developed machine-trained chips.
Internet giants admission
Internet companies are also aware of the importance of the chip. Continue to invest in software research at the same time, have also tried to open up the field of artificial intelligence chip.
In 2015, Amazon acquired the Israeli chip maker Annapurna Labs, a year later sold ARM-based chips for wireless routers, streaming media devices, home appliances and data storage devices. This chip is suitable for low-power devices, main notebook computers and wireless routers.
In 2016, Google announced at the I / O conference the development of a special depth of learning tasks for the chip Tensor Processing Unit (TPU). TPU is tailored for Google's TensorFlow open source deep learning framework, which provides better performance and priority for machine learning. Today, TPU has been to enhance the quality of Google Maps, Google search and other services contribute a lot of power, Alpha Go and Li Shishi's game also has TPU participation. Although Google is not directly selling artificial intelligence chip, but because a large number of enterprises using Google's cloud computing services, but also indirectly swallowed the British Weida, Intel and other old chip manufacturers market. mg100q2ys42
On the other side, Microsoft put the future on the programmable chip. Field Programmable Gate Array (FPGA) allows hardware to be programmed, so that developers can "customize" their own chip, such as both to meet the needs of image processing, but also to meet Scientific computing needs, compared to the development of two dedicated chips, the cost is greatly reduced. Microsoft has applied FPGAs to Bing search and Azure cloud computing platforms, and the development team said that FPGAs will improve the machine learning speed by 100 times.
Apple's W1 chip is more famous. In 2013, Apple acquired the Bluetooth startup company Passif Semiconductor, began to develop wireless technology, and eventually in 2016 launched the first wireless chip W1 chip. Apple has not yet revealed the technical details of the W1 chip, but we see, equipped with W1 chip wireless headset is not only stable connection is also low power consumption, can work up to five hours, the use of Siri is also relatively smooth.
In addition, Baidu last year released the open source benchmark test program DeepBench, used to measure the depth of the chip processing speed. This program is designed to accurately measure the performance of the chip to help hardware vendors and users communicate with each other. They also work with NVIDIA to develop intelligent vehicle control platforms that make driving safer. bsm200gb60dlc
The market is starting
The powerful chip is a solid foundation behind artificial intelligence. With the artificial intelligence into the phone, speakers, watches, headphones, the original chip and face new problems. Thus, the calculation of fast, low energy consumption, small size, low price of artificial intelligence chip has become an urgent desire for the hardware. We believe that when such a chip is born, artificial intelligence will move forward a big step forward.