MIT Technology Review (MIT Technology Review) reported that venture capitalists in 2017 artificial intelligence chip venture capital investment of 113 million US dollars, more than three times 2015 full year, of which British semiconductor startup Graphcore just from Silicon Valley venture Sequoia Capital raised 50 million US dollars to develop new artificial intelligence chip Mythic, Wave Computing, Cerebras, China Deep Kam Technology (Cambex), Cambricon Technology (Cambricon), Cambrian Technology recently Get a $ 100 million investment from the Chinese government.
Why startups have a chance to report that many businesses are now investing heavily in hardware that runs deep learning systems that highlight the limitations of existing chips such as NVIDIA's graphics chips that render pixels in parallel with thousands of microcomputers running through After some adjustments, it has been possible to run deep learning algorithms, but also involves a lot of parallel computing, the biggest drawback is too power.
Carnegie Mellon also requires researchers to cut chip usage time as it puts pressure on the school's power system and they are looking for alternative sources of energy to ease the problem. Startups plan to produce more efficient chips, but they're real Think of a custom chip that makes artificial intelligence applications.
Now the new generation of chips combines multiple processing functions into one step, whereas the graphics chip lets multiple steps produce the same result.The former functions are often tied together to optimize specific use cases, such as training algorithms that help detect potential obstacles from driving .
Graphcore expects to ship in the first quarter of next year, claiming that its chips perform 10 to 100 times faster than other products. "Cambrian customer Huawei also said that for deep learning applications such as training algorithms and identification images, the Cambrian chip Six times faster than graphics chips, and many in the industry who focus on advanced applications such as robotics or computer vision, the chips provided by startups do contribute to the research in this area.
Of course, the chip giant will not be sitting dead, Intel is about to announce a series of new processors by a previously acquired machine learning startup Nervana Systems design, NVIDIA also step by step to upgrade their own chips.Another challenge is the newly created main design highly specialized hardware , Take several years to go public, and it is hard to predict when the industry will change.
But if startups design chips that are too broad, they could sacrifice performance levels, compete with big companies, and some may eventually be bought and sold. It has not always been easier for shrimp to fight big sharks.