AI News
All stories
Reported

Optical computing firms aim for 10- to 80-fold energy gains over GPUs in AI inference

IEEE Spectrum reports that optical computing companies aim for 10 to 80 times better energy efficiency than today's GPUs, a goal not yet shown at scale.

Chips & Infrastructure · · 4 min read

Optical computing companies aim for 10 to 80 times better energy efficiency than today's GPUs, according to IEEE Spectrum. The energy advantage has not been conclusively demonstrated, and the goal has not yet been shown at scale. Sector valuations now total US $2.22 billion, up from $0.4 billion in 2022. The article links much of that interest to AI inference, whose power demands are pushing data-center construction.

The case for optics rests on a scaling argument dating to a 1978 proposal by Joseph W. Goodman and colleagues. Light-emitting diodes encode a vector into light intensity. A cylindrical lens copies that vector across a grid, and a fixed attenuator with one pixel per matrix element multiplies each copy by a matrix value. A second lens sums the results row by row, and detectors read out the answer. Multiplying an N-by-N matrix by a vector of length N takes N squared multiplications in electronics, so energy scales with N squared. In optics, the multiplications are passive once the matrix is fixed. The main costs are encoding the vector and reading out the answer, which scale only with N.

A 2022 Cornell demonstration led by Peter McMahon used an optical neural network that consumed less than one photon of energy, about 2.5 x 10-19 joules, per multiplication. A single multiply in a standard computer can take between 10-14 and 10-12 joules, which suggests optics could reduce energy use by a factor of 100,000. McMahon cautions against that extrapolation because the calculation left out the cost of encoding data into light. His example: if an electronic multiply costs 1 picojoule and encoding one vector element costs 1,000 picojoules, the electronic approach wins for any vector shorter than 1,000 elements. He gives a rough crossover point of somewhere between N equals 1,000 and N equals 10,000 for various neural networks.

The article also applies Amdahl's law, which limits the gain when only part of a workload can be accelerated. Matrix-vector multiplication is only part of inference. Other nonlinear operations still have to run on the data, so the share of the workload optics can handle is below 100 percent and varies by approach. Lumai, an Oxford, England company, uses a programmable spatial light modulator in a free-space design. It claims its next-generation Iris Aura server will perform prefill using roughly one-tenth of the energy a GPU would use. Lumai made its earlier Iris Nova server available for evaluation in April 2026 and calls it a proof of concept. The electronics still handle the nonlinear functions.

Neurophos, based in Austin, Texas, uses metamaterial modulators in a chip with a 1,000 by 1,000 array on a square with 5-mm sides. Its CEO, Patrick Bowen, claims a target of 50 times the speed and 35 times the energy efficiency of Nvidia's Blackwell B200. The company has a road map to 250 times the speed and roughly 90 times the energy efficiency in later generations. It is still choosing a production site ahead of rack deliveries to customers in 2028.

Other companies take an on-chip route. Q.ANT, based in Stuttgart, Germany, and Lightelligence, based in Shanghai, trade the three-dimensional layout of free-space optics for chips that can be made in semiconductor fabs. In Q.ANT's design, vector elements are sent as sequential light pulses through waveguides and interferometers. That approach is slower than Goodman's all-at-once method, though multiple wavelengths can offset some of the speed loss. Q.ANT's CEO, Michael Förtsch, identifies the converters as the main obstacle. He says that without a fast, low-energy converter, "we don't see realistic chances that" a general optical vector-matrix multiplication machine makes sense. A Q.ANT press release says its second-generation processor showed up to a 100x performance increase over the first generation in an evaluation at the Leibniz Supercomputing Centre. The company's internal benchmarking indicates up to 30 times higher energy efficiency than conventional processors.

Some companies are moving away from matrix multiplication. LightSolver, based in Israel, runs differential equations on a laser-style cavity with a gain medium and a programmable modulator. Its CEO, Ruti Ben-Shlomi, says differential equations can be solved much faster and with higher energy efficiency on its device. The team showed in emulation that the device can solve such equations 40 to 80,000 times as fast as a GPU, depending on the workflow. The article says the company is still working on explicit energy benchmarks. Optalysys, based in Leeds, England, places passive optical components in the networking path, where switch cards today use power to combine results. Its founder, Nick New, argues this approach suits the decode phase of inference, which is limited by moving data in and out of memory.

Yale's Logan Wright is less optimistic that optics will replace GPUs soon. He argues that algorithm development is more likely to move in a direction that favors GPUs, because model developers want their models to run well on them. He points out that investment by Nvidia and TSMC in co-packaged silicon photonics, aimed at circuit switching and interconnects rather than computing, may still open the door to more optical processing later.

Why it matters

Buyers planning inference capacity should treat optical efficiency claims as unproven. The article notes that energy gains depend on system size, encoding costs, and how much of the workload optics can actually handle.

Original source

Further reporting

Where we saw it

1 source

IEEE Spectrum

LightSolverLightelligenceLumaiNeurophosNvidiaOptalysysQ.ANT

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works