When we started this niche market, the first question is: what is our artificial intelligence chip? The best way is to think about what the artificial intelligence software needs: a good processing speed, as well as high processing speed required for high power. However, the operation method of the processor is also very important. This section from the Technology Review MIT cited text explains why we can not just use the high-end Intel processor chip to achieve artificial intelligence:
A top Intel processor contains too much punch used to run huge financial spreadsheet or business software, and in depth study of the chip optimization of some specific types of problems (such as understanding the voice commands or image recognition) divided into numerous small block (bite-size chunks. Because GPUs has thousands of small processor cores on the same chip, so they can handle thousands of small pieces at the same time. Distribution of an Intel processor to will be in this kind of work is a huge waste of resources, because each of such a processor contains dozens of originally designed to run complex algorithm kernel, and the depth of learning chip as long as to deal with all those small tasks, not to need to think so much more complex operations.
GPU (Processing Units Graphical) is the application of artificial intelligence (such as image recognition and computer vision) will be used in the chip. The actual method you can frame a chip can be optimized for specific artificial intelligence tasks, such as image recognition, speech recognition, or any type of large data analysis. In the case of deep learning, you use artificial neural networks to generate simulated neurons to stimulate the brain behavior. The following is to create a chip and hardware solutions and commitment to optimize the artificial intelligence tasks of the 5 companies.
KnuEdge is actually not a start-up company, which was created by the NASA's former head of the company, has been operating in a stealth mode for 10 years. KnuEdge recently came out of the hidden model and let the world know that they get $100 million from an anonymous investor to invest in developing a new neuron chip ":
KUNPATH provides LambaFabric based chip technology, LambaFabric will be on the market with the GPUs, CPUs and FPGAs completely different architectures for the calculation of the neural network. LambdaFabric essence is in the high requirements of the computing environment to expand to 512000 equipment designed and frame to frame delay time is only 400 ns, low-power 256 processors. KNUPATH technology, based on biological principles, will redefine the chip level / system level calculations in the data center and consumer devices market.
Compared to other similar chips, this chip technology should provide 2 times to 6 times the performance advantage, and the company has been through the sale of their prototype system to get the income. Under the "KnuEdge umbrella structure, KnuEdge by three separate companies, KnuPath provide their chips. KnuVerse provide through the verification of military grade speech recognition and verification technology, Knurld.io is to allow developers to simply fusion voice verification to their patent products of public cloud service API (public cloud API service). KnuEdge claims that now only need to speak a few words on the microphone can be done to verify the computer, network, mobile applications and networking equipment. Don't remember how many great things will be in the future.
Founded in 2014, is located in the San Diego startup nervana systems has grown from 20 different investment institutions received $24.4 million funds, and one is very respected Wheelock Jie venture capital firm Draper Fisher Jurvetson, Dfj). Nervana Engine The (will come out in 2017) is a dedicated to the depth of learning and optimization of ASIC chips. The implementation of this program is due to a new memory technology called Bandwidth Memory High, while having a high capacity and high speed, to provide 32GB on-chip storage and 8TB memory access speed per second. The company currently offers a artificial intelligence service "in the cloud", they claim that this is the fastest in the world and has been used in financial services, health care providers, and government agencies, their new chip will ensure nervana cloud platform within the next few years remains the fastest speed.
Founded in 2015 by the Chinese start-up companies Robotics Horizon (horizon robot) has been obtained from the Sequoia and the legendary venture capitalist Milner Yuri and other investors did not disclose the amount of seed funds. They are working on the establishment of a one-stop artificial intelligence solutions, the definition of things smart, so that life is more convenient, more interesting and more secure.
Krtkl, founded in 2015, is dedicated to creating "a tiny wireless computer used to create something completely different."". Technology will be infatuated with Snickerdoodle, a dual core ARM processor, FPGA, WiFi, Bluetooth, starting at $65, "to the minimum, the most difficult, the most affordable Fu can robots, unmanned platform for human-computer and computer vision." This product is in fact through the public to raise more than 160 thousand dollars of funds. The latest information is that they have received a primary version of the Snickerdoodle, and will soon be shipped.
Eyeriss is in fact not a start-up, but because it was developed by MIT and received a lot of media coverage, so we can not exclude it from this list. Eyeriss is a highly efficient depth convolution neural network (CNN) accelerator hardware, and the architecture is as follows:
The main harvest here is: the hardware described above is more efficient, 10 times faster and consumes 10 times less than the existing graphics processor GPUs. The U. S. Department of Defense Advanced Research Project Agency (DARPA) provides part of the funds, by vivienne Sze led the MIT team in this year's conference discloses chip is the most advanced neural network for the first time in the custom chip presentation.
Original text:5 Building Startups Artificial Intelligence Chips
Compile: Zheng Yao
Commissioning editor: Zhou Jianding
CCAI 2016 China Conference on Artificial IntelligenceWill be held in Beijing on August 27-28, chairman of the AAAI headed more than 30 global technology leader and pioneer industry create a platform for domestic artificial intelligence frontier, 6 major theme report, human-computer interaction, machine learning, pattern recognition, industrial real four thematic forums, 1000 + high quality honored guests.Forty percent off discount tickets.