Scientists at China’s Tsinghua University have developed a fully analog photoelectronic ACCEL chip that promises to revolutionize high-speed computer vision applications. The chip, which combines electronic and optical technologies, is capable of demonstrating unprecedented power efficiency and highest computing speed for computer vision tasks. In this area, the new chip is radically superior to modern GPUs.
Traditional processors have limited processing speeds and consume enormous amounts of power when solving computer vision problems such as image recognition for autonomous driving, robotics and medical diagnostics. These tasks require high-resolution image processing, accurate classification, and extremely low latency.
The ACCEL chip takes advantage of the emerging field of photonic computing, which uses light to process information. By integrating diffractive optical analog computing (OAC) and electronic analog computing (EAC) on a single chip, ACCEL achieves remarkable energy efficiency and computing speed.
The OAC method uses the manipulation of light waves through diffraction to encode and process information. Using interference patterns created by light, calculations are performed in an analog fashion, processing the data continuously rather than in discrete digital steps. The EAC method uses electronic components to manipulate continuous physical quantities. Instead of working with digital signals in the form of zeros and ones, EAC uses constantly changing analog signals.
Both methods provide advantages for certain types of computers and make it easier to develop high-speed vision problems.
ACCEL image processing does not require an ADC to convert the image, but directly uses light-induced photocurrents for calculations, resulting in significantly reduced latency. ACCEL achieves system power efficiency of 74.8 peta-ops per watt, more than three orders of magnitude higher than current GPUs. Computing speeds reach 4.6 peta operations per second, with more than 99% of calculations performed optically.
By integrating optoelectronic computing and adaptive learning, ACCEL achieves competitive object classification accuracy across a variety of tasks. The new chip achieved accuracies of 85.5%, 82.0%, and 92.6% for Fashion-MNIST, ImageNet 3-class classification, and time-lapse video recognition tasks, respectively. In particular, ACCEL has high reliability even in low light conditions, making it suitable for portable devices, autonomous driving and industrial applications.
The new chip’s extremely low power consumption significantly reduces heat dissipation, paving the way for further improvements and miniaturization. Unlike traditional optoelectronic digital computing systems, ACCEL flexibly combines diffractive optical computing and electronic analog computing, and its architecture achieves scalability, nonlinearity and high adaptability.
In a study published in the journal Nature, researchers found: “Developing a computer system based on a completely new principle is a huge task. More important, however, is the successful translation of this next-generation computing architecture into real-world applications that meet the critical needs of society.”
In a review of the study published in the journal Nature’s Research Briefing, experts expressed their belief “ACCEL could ensure that these architectures play a role in our daily lives much sooner than expected.”
Everything new is undoubtedly the forgotten old. The very first analogue computing device is the slide rule, which is well known to the older generation.
Another well-known example of analog computing devices is the MH-7 desktop analog computer developed in 1955. She successfully solved ordinary differential equations up to the 6th order. No less successfully, with the help of such machines, mathematical models of physical processes were created, which were used to solve automated process control problems.
In an analog computer (AVM), the instantaneous value of the original variable quantity is linked to the instantaneous value of another quantity, which often differs from the original physical nature and scaling factor. Each elementary mathematical operation usually corresponds to a physical law that establishes mathematical relationships between physical quantities at the output and input (e.g. Ohm’s law).
Features of the representation of initial quantities and the construction of algorithms determine the high speed of AVM operation and ease of programming, but limit the scope and accuracy of the obtained result. AVM is characterized by low universality (algorithmic limitation) – when solving problems of a different class, it is necessary to rebuild the structure of the machine and the number of crucial elements.
And now we are witnessing how, in a world of seemingly victorious digital technologies, analog data processing, which has reached a new level of development, is being used again.