MIT has developed the architecture of a new photonic processor. It will increase the efficiency of optical neural networks by a thousand times compared to similar devices.
The chip will reduce the amount of electricity consumed by the data center. Let's talk about how it's set up.
A photo - β unsplash
Why a new architecture is needed
Optical neural networks are faster than traditional solutions using electronic components. Light isolation of signal paths, and laser streams are able to pass through each other without mutual influence. Thus, all signal paths can operate simultaneously, which ensures a high data transfer rate.
But there is a problem - the larger the size of the neural network, the more energy it consumes. To solve this problem, special accelerator chips (AI accelerators) are being developed that optimize data transfer. However, they do not scale as well as we would like.
The solution to the problem of energy efficiency and scaling of optical chips was taken up at MIT and a new photonic accelerator architecture that reduces device power consumption by a thousand times and works with tens of millions of neurons. The developers say that in the future the technology will find application in data centers that interact with complex intelligent systems and machine learning algorithms, as well as analyze big data.
What is she
The new chip is based on an optoelectronic circuit. The transmitted data is still optically encoded, but balanced homodyne detection is used for matrix multiplication (). This is a technique that allows you to form an electrical signal based on two optical ones.
To transmit light pulses with information about input and output neurons, one signal path is used. Data on the weights of neurons, on the contrary, come through separate channels. All of them "diverge" through the grid nodes of homodyne photodetectors, which calculate the output value for each neuron (determine the signal level). This information is then sent to the modulator, which converts the electrical signal back into an optical one. Further, it is sent to the next layer of the neural network and the process is repeated.
In their scientific work, engineers from MIT the following scheme for one layer:
Image: / CC BY
The new architecture of the AI ββaccelerator requires only one input and one output channel for each neuron. As a result, the number of photodetectors is equated to the number of neurons, not their weights.
Such an approach allows saving space on the chip, increasing the number of useful signal paths, and optimizing power consumption. Now engineers from MIT are creating a prototype that will test the capabilities of the new architecture in practice.
Who else is developing photonic chips
Similar technology developments Lightelligence is a small startup based in Boston. Employees of the company say that their AI accelerator will allow them to solve machine learning problems hundreds of times faster than classic devices. Last year, the team completed the prototype of their device and prepared to conduct tests.
Works in the field of photonic chips and Cisco. At the beginning of the year, the company announced startup Luxtera, which designs photonic chips for data centers. In particular, the company produces hardware interfaces that allow fiber optics to be connected directly to servers. This approach increases network bandwidth and speeds up data transfer. Luxtera devices use special lasers to encode information and germanium photodetectors to decrypt it.
A photo - β unsplash
Other large IT companies, such as Intel, are also involved in optical technologies. Back in 2016, they began producing their own optical chips that optimize data transfer between data centers. Recently, representatives of the organization that they plan to implement these technologies outside of data centers - in lidars for unmanned vehicles.
With the result that
So far, photonic technologies cannot be called a universal solution. Their implementation requires high costs for the technical re-equipment of data centers. But developments like those being developed at MIT and other organizations will make optical chips cheaper and likely to "promote" them in the mass market of data center equipment.
We are we help companies develop their IT infrastructure and provide private and hybrid cloud services. Here's what we write about in the corporate blog:
Source: habr.com