Prosperous prospects, IT companies cross-border semiconductor industry
The global semiconductor industry is undergoing a long rising cycle. South Korean semiconductor companies have submitted bright reports. Samsung Electronics reported preliminary operating data for the first quarter, showing a quarterly operating profit of $14.7 billion, a year-on-year increase of 57.58. %. Samsung’s semiconductor division also surpassed smartphones many years later to become the group’s profit winner. Preliminary data released by LG Electronics recently showed that its profit growth in the first quarter of 2018 was as high as 20%.
Analysts believe that the red fire in the semiconductor industry will not end in the short term. On the contrary, new applications such as the Internet of Things, smart cars, and cloud computing are leading the new round of semiconductor business prosperity.
In the face of new prospects, IT companies have become a trend to enter the semiconductor industry cross-border. The most prominent is the chip design business. In the field of AI chips, IBM, the first to show the AI charm to the public, has created its own neuron chip. Developed its own deep learning chip. Already dozens of AI chips have been released. The product abbreviation has already exceeded the scope of CPU. From APU, it has been ranked to UPU. On this node, Nabiss has released AI chip across the border. What is the business logic?
AI chip, an important industry development direction
Anting answered the question from the reporter of the science and technology daily newspaper. The reason he gave is that 'artificial intelligence chip is a very important industry development direction'.
He explained that artificial intelligence covers a wide range of technologies and fields, and there is a great deal of cross-convergence between them. At this stage, it can be roughly assumed that the core technology of artificial intelligence is based on machine learning, and machine learning generally includes deep learning and nerves. Neuromorphic Two technical directions. The deep learning system is a hybrid platform built by software and hardware. Most of the software is operated according to the principle of multi-layer neural network. The hardware mainly uses CPU, GPU and other relatively 'traditional' Processor. The neuromorphic computing system is a hardware platform that builds computing nodes according to the neural network structure. The representative product is a neuron chip.
Anting Chen emphasized that deep learning technology is a great innovation in software technology, but the hardware still belongs to the von Neumann architecture, and it is difficult to overcome its shortcomings in terms of parallel processing speed and energy consumption. Neuron chips have unprecedented ultra-low energy. Consumption and processing efficiency. This chip is based on hardware, uses a massively parallel processing architecture, and the input data is simultaneously poured into the memory of each neuron, thus enabling high-speed processing of information at a lower frequency while maintaining low power consumption. State. More importantly, neuron chips work through pattern learning and recognition. All neurons can complete an identification within the same clock cycle. Complex application scenarios can be addressed by increasing the number of neurons. Neuron chips Can handle various types of data, including non-specific data generated by various sensors.
Neuron chips, training learning is done almost instantaneously
NM500 chip package size is only 4.6 × 4.5 mm, frequency 36 MHz, peak power consumption is less than 200 MW. NM500 contains 576 identical neurons, each neuron includes a logic device and a 256-bit memory. Si also showed the corresponding evaluation board, providing stacking capabilities, integrated FPGA, motion sensor, USB port and SD card slot, support Windows and Arduino development environment. Science Daily reporter saw in the live demo, one by the evaluation board The experimental device constructed with a lightweight robotic arm can identify chips of different denominations based on the pattern of the chip surface and classify the chips. It is impressive that training and learning new chip patterns are almost instantaneously completed. The development environment is simple and convenient, the whole process is very easy.
Antinho said that the NM500 was productized after obtaining QV's design license, and Napas applied a large number of proprietary technologies during the productization process. Napes is optimistic about the future prospects of neuron chips. In the networking age, the number of interconnected devices is 10 billion or 100 billion. The data generated by these devices needs to be processed to generate value. The application scenario of neuron chips is full of enormous imagination space.