Google enlarged the recruitment of self-developed A chip.
Google has made another big move, releasing the latest generation of artificial intelligence chip TPU v4, the best thing is the integration of optical interconnection switches.
First of all, what chip is Google's TPU? Its full name is TensorProcessor Unt, which is a special computer processor developed by Google. Brain-like neurons are used to speed up AI applications.
TPU chips accelerate the training and inference process of large-scale data through parallel processing, and can process large amounts of image, sound, language and other types of data.
The working principle of TPU chip is based on deep learning technology, through mathematical operations to build neural networks, so as to achieve machine learning and artificial intelligence. TPU chips are widely used in Google's cloud computing and data centers, such as speech recognition, image recognition, natural language processing and machine translation. Now TPU has evolved to the fourth generation, and its most important feature is the use of optoelectronic chip technology. Traditional chips use electrical switches to control the on-off or disconnection of circuits, thereby representing one and zero. However, this method has relatively high power consumption and low stability.
So Google replaced electricity with light, using a small mirror hundreds of times smaller to reflect light to an area, using dark and bright to represent 0 and 1, not only fast but also basically no loss, stability is also much higher.
Based on this principle, Google developed a technology called configurable optical interconnect switch (reconfigrbl pftclswith), and used it in the fourth generation of TPU chips. And based on it, interconnect scaling and performance tuning are carried out, for example, the interconnect topology of the chip can be adjusted according to different model data streams, so as to achieve the optimal performance improvement can be more than twice as high.
Another importance of optical interconnect switches is high reliability. Simply put, if some chips have errors or failures, the interconnection can be dynamically adjusted, so that the failed part does not affect the overall function, and there may be a small performance loss. The money-rich and wayward Google is not yet looking at the existing commercial solutions.
So it developed a new optical path switch chip and realized the world's first data center-level configurable optical interconnect. Using MEMS-based mirror array-based technology, optical switch chips can achieve low latency, low switching delay, and low power consumption, and the cost of optical path systems is also very low. It accounts for less than 5% of the overall TPU v4 supercomputer cost.
Recently, the research of optoelectronic chips is very hot, and it is also a very good entrepreneurial direction. Interested audiences can follow, want to see more chip technology commercialization prospects, remember to follow me, I am chip engineer Lao Shi, don't forget to give me a little red heart. Old stone talking core.
