First light processors operate AI efficiently

Contents

The increasing complexity of artificial intelligence is challenging conventional electronic computers and consuming more and more energy. Photonics, which uses light instead of electricity for data processing, offers a promising alternative.

Two independent research teams have presented photonic processors that are designed to increase computing power and reduce energy consumption. Previously, only simple demonstrations such as benchmarks were possible on photonic processors, but advanced AI models did not run on them. The work has been published in the journal Nature.

Photonic computing

With conventional processors and GPU accelerators, data movements cost a lot of energy. This is not only expensive due to the electricity costs for operation, but also complex to cool – all energy must be dissipated in the form of waste heat, which itself costs more energy.

Photonic or optical computing generates less waste heat; efficiency potentially increases. At the same time, other challenges arise. Instead of the binary values 0 and 1, photonic computers use a continuous spectrum of possible values. They therefore require a fundamental rethink. Intel and Nvidia, among others, are therefore pursuing a mixed approach: CPUs and GPUs calculate internally as usual, but exchange data optically with each other.

"Photonic computers have been under development for decades, but these demonstrations could mean that we can finally harness the potential of light to build more powerful and energy-efficient computer systems," writes researcher Anthony Rizzo from Dartmouth College in an accompanying commentary in Nature.

Lightmatter

The Lightmatter team led by Nicholas Harris from California has developed a photonics processor that implements AI models efficiently and with high accuracy. "For the first time in computing history, we have demonstrated a non-transistor-based technology that can perform complex, real-world tasks with similar precision to existing electronic systems," explains Nicholas Harris, CEO of Lightmatter.

The authors show that their photonic processor can run advanced AI models such as BERT and ResNet with an accuracy comparable to electronic processors. It can also run a range of applications: For example, it creates Shakespeare-like texts, classifies movie reviews and plays "Pac-Man". How well the processor copes with the tasks assigned to it depends on the type: for image classifications, it achieves 90 percent of the accuracy of a conventional processor, but only just under 30 percent for questions in natural language.

Mario Chemnitz, junior professor for intelligent photonic systems at Friedrich Schiller University Jena, who was not involved in the study, says: "The amount of electronic and optical components that functionally interact on a single computing card that can be plugged directly into a conventional computer architecture is unique and marks a significant technological step for the field of photonic computing."

He does the math: The new photonics processor is three times faster and ten times more energy efficient than an A100 GPU accelerator from Nvidia's penultimate Ampere generation, and its performance is comparable to its successor H100 from the previous Hopper generation.

However, he notes that the start-up does not fully disclose its system and the publication therefore does not contain the necessary depth of information to conclusively evaluate the technology. He also sees a fundamental problem in the still inadequate accuracy and the discrepancy between analog (i.e. optical) hardware in the midst of digital infrastructure.

Lightelligence

Independently of this, the Lightelligence team led by Bo Peng from Singapore has developed a photonic accelerator called PACE. This is designed to efficiently solve complex combinatorial optimization problems. In tests, PACE showed a drastically reduced latency compared to conventional GPUs from Nvidia. Low latency is particularly relevant for processing data in real time.

PACE consists of more than 16,000 photonic components, enables high-speed computing at up to 1 GHz and shows up to 500 times lower minimum latency compared to small circuits or single photonic components.

Both teams point out that their systems are scalable, but need further optimization to take advantage of optical computing.

(mma)