Lightmatter’s New Passage Chip Reduces AI Energy Use Photonics Technology

From powering artificial intelligence (AI) algorithms to optimizing tasks on our smartphones, the computer chip is the silent engine that drives everything in the modern world.

Nowhere is this more evident than in the thriving AI chips market, expected to reach $383.7 billion by 2032, as per Allied Market Research. Yes, that’s a staggering number!

Yet, the market is expected to witness major growth, exacerbating the need for quicker and more energy-efficient AI computing hardware. After all, AI has already reached formerly unimaginable levels amid its exponential rise.

For decades, the computing industry has adopted shrinking transistors, adhering to Moore's Law (the tendency for computer chips to get smaller and more powerful over time.) Now, potential miniaturization of technology has reached a stage where researchers are looking at “light" to solve the problems of speed and energy that we face with current computers.

Among many in this pursuit of light-based computing is Lightmatter, a Silicon Valley startup that is making waves with an innovative chip. So, what did Lightmatter announce that has the potential to completely change AI computing in the years to come?  

What Did Lightmatter Announce?

Valued at an astounding $4.4 billion after raising $850 million in venture capital,
Lightmatter has revealed that it has developed a new computer chip that uses light to speed up AI operations and reduce electricity consumption.

This signifies a substantial departure from conventional electronics. While the idea of using light (or photonics) to transfer data between computers has been gaining popularity for a while, Lightmatter has made real-life advancements toward it with its photonic interconnect technology, Passage.

While it promises speedier connections, essential for sophisticated AI software, Lightmatter says its advanced idea can also be utilized to carry out computations. The company's innovation directly addresses the shortcomings of traditional computing chips, as described in a report published in the renowned scientific journal Nature.

The advanced chip from Lightmatter alters precisely calibrated light beams to get around traditional limitations. The chip directs these light beams toward one another rather than through transistors, and an integrated package of chips developed by their partner, GlobalFoundries, measures the outcomes.

Remarkably, this processor achieves accuracies comparable to conventional 32-bit floating-point digital systems right out of the box. The processor integrates six chips within a single package, utilizing high-speed interconnects between vertically aligned photonic tensor cores and control dies. 

What Went Into This Breakthrough?

Acquiring computational precision was one of the primary challenges that previous attempts at photonic computing had to overcome.

Consider a situation in which just a small value is recorded as zero because it disappears in the noise.

Well, Lightmatter has ingeniously found a way around this issue by using a technique that breaks down both very large and very small numbers into manageable groups before they are processed by photonic circuits. This clever strategy ensures that even the most minute values are preserved throughout the computation.

According to Nick Harris, Lightmatter's CEO, the outcome is a computing chip capable of tackling current AI challenges with a level of precision not seen before. While he anticipates it might take another decade for this technology to become mainstream, the implications are enormous.

Harris stated, "What we're doing is looking at the future of where processors can go. We fundamentally care about computers, and this is one of the alternative paths. There's trillions of dollars of economic value that's behind the idea that computers will keep getting better.”

This announcement by Lightmatter paints a clear picture of the computing industry’s inflection point. The relentless demands of AI workloads are pushing the boundaries of traditional scaling, perhaps even breaching Moore's Law, Dennard scaling, and memory scaling, all of which have effectively stalled on a per-silicon-area basis.

The current approach of simply doubling the silicon area or the number of memory chips to enhance performance comes with an exponential increase in cost. This trend is already evident in the prohibitively expensive consumer GPUs, requiring new approaches.

Lightmatter's unveiling of their photonic processor addresses three critical areas for the future of computing: memory, interconnect, and compute. While developing a scalable, DRAM-like memory solution remains a significant challenge, their groundbreaking Passage photonic interconnect technology has already garnered attention for its potential to revolutionize data movement. 

So, What Does It Mean For The Future?

With its revolutionary photonic processor, Lightmatter is tackling the core of computation itself. High bandwidth, low latency, and the possibility of increased energy efficiency through light-based computation are some of the inherent benefits that photonics offers for these activities. Despite its potential, earlier photonic accelerators frequently lacked the accuracy needed for real-world AI applications.

Lightmatter's innovation also stands out by achieving the necessary computational precision to run advanced AI models like ResNet, BERT, and DeepMind's Atari deep reinforcement learning algorithm – without requiring any modifications.

Lightmatter's breakthrough marks a turning point in computer history. The ability of non-transistor-based technology to do intricate, real-world tasks with precision and efficiency, compared to that of conventional digital electronics, has been achieved for the first time.

As the AI industry struggles with the constraints of traditional silicon scaling, Lightmatter provides a viable alternative. While electronic computing will surely be the foundation, photonics has begun to emerge as a promising alternative.

Can Lightmatter bring in a new era of computing?

Let us know in the comments below!