Microsoft uses "light" to run AI and makes it onto Nature. With 100 times the energy efficiency, it disrupts GPUs. A Chinese chief researcher takes the lead.
In the past few decades, major companies have been secretly competing in the chip field: chip price hikes, GPU shortages, and anxiety about AI computing power...
While everyone was focusing on the iterative upgrades of chips, Microsoft was quietly doing something else: redefining computing with light.
They spent four years using a mobile phone camera, Micro LEDs, and lenses to piece together an analog optical computer (AOC).
Now, this experiment has been published in Nature, bringing an imagination of the future that is enough to disrupt GPUs.
The Arrival of Photons: The Secret of Fixed-Point Search
For decades, the story of computing power has almost been written on silicon wafers: the acceleration of Moore's Law, the stacking of GPUs, and the anxiety about energy consumption.
But in Cambridge, UK, a small team at Microsoft Research took a completely different path - letting light do the calculations.
They pieced together an analog optical computer (AOC) using materials that are not rare at all: Micro LEDs, optical lenses, and a camera sensor from a mobile phone.
It looks more like a "lab-assembled machine," but it opens up another possibility for computing power.
A detailed image of the analog optical computer in the Microsoft Research lab in Cambridge, UK. It is made using commercially available components, such as micro-LED lights and sensors from smartphone cameras.
Actually, the concept of optical computing was proposed as early as the 1960s, but at that time, limited by technology, it remained at the theoretical level.
Now, the Microsoft team has really made it a reality.
The real secret of the AOC lies not in these parts, but in its operating mode - fixed-point search.
It puts optics and analog electronic circuits into a loop: the optical part performs matrix-vector multiplication, and the electronic part handles nonlinearity, addition and subtraction, and annealing operations.
Each cycle takes only about 20 nanoseconds. The signal iterates continuously in the loop until it converges to a stable "fixed point."
And this fixed point is the answer to the problem.
The internal structure of Microsoft's analog optical computer: the upper left is the overall schematic, and the lower right is the link of alternating photon and electron calculations.
This method solves two long-standing problems that have plagued optical computing:
Firstly, it avoids the high-cost digital-to-analog conversion in the hybrid architecture, significantly reducing energy consumption.
Secondly, it naturally has the advantage of being resistant to noise.
During the iteration process, the fixed point is like a magnet, firmly attracting the answer and not easily deviating.
That's why the AOC can handle both optimization problems and AI inference on the same platform.
Four years ago, this was just an adventurous attempt in the lab.
Now, it has been published in Nature, for the first time making optical computing no longer a paper concept but truly entering the public eye.
Microsoft CEO Satya Nadella reposted the AOC research on X, calling it "a new way to solve complex real-world problems more efficiently" and emphasizing that the result has been published in Nature.
From Banks to Hospitals: The First Real-World Application of AOC
What the Microsoft team most wants the public to see is not a display of skills, but that this technology can really be used in the real world.
So the Microsoft team chose two of the most representative scenarios - finance and healthcare - for verification.
In the financial field, they cooperated with Barclays Bank to transfer the "delivery versus payment" settlement problem that clearinghouses face every day to the AOC.
Traditional clearinghouses need to find the most efficient settlement method among hundreds of thousands of transactions. Here, the team first built a scaled-down version:
46 transactions and 37 participants, transformed into an optimization problem with 41 variables.
The results showed that the AOC found the optimal solution in only 7 iterations.
How transactions between multiple financial institutions can get the optimal solution through the AOC.
Shrirang Khedekar, a senior engineer at Barclays who also participated in the paper, commented:
"We believe there is great potential to explore. We also have other optimization problems in the financial industry, and we believe the AOC technology has the potential to play a role in solving these problems."
Hitesh Ballani guiding research on future AI infrastructure in the Microsoft Research lab in Cambridge, UK.
The healthcare field also showed breakthroughs.
The team rewrote MRI compressed sensing imaging into an optimization problem that the AOC can run. They first tested a 32×32 Shepp–Logan phantom brain slice image on the hardware and successfully restored the original image using 64 variables.
Furthermore, they used digital twin (AOC-DT) to reconstruct a real brain MRI dataset containing 200,000 variables.
MRI image reconstruction: the restoration process of the Shepp–Logan phantom and the reconstruction of large-scale brain MRI using AOC-DT.
Michael Hansen, the director of biomedical signal processing at Microsoft Health Futures, said bluntly:
"To be transparent, we can't use it clinically now. This is just a small-scale experiment, but it gives the feeling that - if it really reaches full scale, the consequences will be unimaginable."
He also envisioned that in the future, the original MRI data could be directly streamed to the AOC on Azure, and then the results could be sent back to the hospital in real-time.
That would mean that the scanning time could be shortened from 30 minutes to 5 minutes, not only greatly improving efficiency but also reducing the suffering of patients.
"We must find a way to obtain the original data and stream it to where the computer is located."
From finance to healthcare, the signals released by these two cases are very clear:
The AOC is no longer a conceptual attempt in the lab but is truly moving towards the transformation of the real world.
A New Path for AI: Possibilities Beyond GPUs
Actually, the breakthrough that most excited the research team is not in the financial or healthcare fields, but in artificial intelligence.
A lunch conversation in the lab led to a turning point.
Researcher Jannes Gladrow realized that the "fixed-point search" mechanism of the AOC is naturally suitable for equilibrium models that require repeated iterations and ultimately converge to an equilibrium state (such as deep equilibrium networks DEQ and modern Hopfield networks).
Three equivalent representations of the Deep Equilibrium Network (DEQ, an equilibrium model).
On GPUs, these models consume a huge amount of computing power, while on the AOC, they are almost "born for photons."
So the team tried to map some simple AI tasks to the AOC. The results soon emerged:
In the MNIST and Fashion-MNIST classification tasks, the results of the AOC and the digital twin (AOC-DT) were almost 99% aligned.
In nonlinear regression tasks (such as fitting Gaussian curves and sine curves), the AOC also performed stably, and the curves almost coincided with the simulation results.
Through time multiplexing technology, the researchers also expanded the hardware to an equivalent scale of 4096 weights, proving that it can not only handle "small toys" but also has the potential for further expansion.
The experimental results of the AOC in MNIST classification and nonlinear regression (Gaussian curves, sine curves).
These experiments show a new path beyond GPUs.
Microsoft researchers believe that in the future, the most laborious part of large language model inference - state tracking - may be handed over to the AOC.
Imagine if the complex inference process no longer depends on energy-consuming GPUs but is completed by an optical computer, the required energy consumption may be reduced by two orders of magnitude.
In an era anxious about computing power and energy consumption, such results undoubtedly ignite the imagination of the industry.
A Long-Distance Race and Vision: Another Track for Computing Power
The Microsoft research team is well aware that the current AOC is just a prototype, and there is still a long way to go before it can be commercially available.
It can currently handle a weight scale in the hundreds, but the researchers have already drawn up an expansion roadmap:
In the future, through modular expansion, each module can support about 4 million weights.
By splicing dozens to thousands of modules, the overall scale can be pushed to 0.1 - 2 billion weights.
What's even more impressive is the energy efficiency comparison.
The team estimates that the mature version of the AOC is expected to reach 500 TOPS/W (about 2fJ/operation), while the current most advanced GPUs (such as the NVIDIA H100) are only about 4.5 TOPS/W.
This means that the energy efficiency gap is as high as two orders of magnitude.
As project researcher Jannes Gladrow said:
"The most important feature brought by the AOC is that we estimate its energy efficiency can be increased by about a hundred times. This is almost unheard of in the hardware field."
In other words, in future large model inference tasks, if GPUs are "gas guzzlers," the AOC may become an "electric vehicle."
It can not only run but also operate continuously with extremely low energy consumption.
A Team of Stars: The People Behind the Optical Computer
Behind this machine that thinks with light is not the solitary struggle of a genius, but the collective wisdom of a group of interdisciplinary researchers.
Francesca Parmigiani is the chief research manager at Microsoft's Cambridge Research Institute.
She led the team to turn a concept that has circulated in the academic circle for half a century into real hardware and insisted on opening up the "digital twin" so that more researchers can participate in the experiment.
She often says that the AOC is not a general-purpose computer but an "optical accelerator" that can run new possibilities in key scenarios.
Jannes Gladrow is a machine learning expert on the team.
