Just out todayAugmented reality: The AR trainees were faster on the model they trained on, and no better anywhere elseVirtual reality: They matched the exercise dose exactly, and the VR advantage fell below the line that mattersVirtual reality: Both groups used the same headset. One watched a video first and picked up 10.8 more blocks

Artificial intelligence/China

A 1.5 billion parameter language model with no matrix multiplication, running on 13.8 watts

A spike-based 1.5 billion parameter language model with the matrix multiplication removed ran at 161.8 tokens per second on 13.849 watts.

What the paper found

A neuromorphic spike-based large language model framework converts transformer behaviour into neural dynamics such as spike trains, combined with quantization and sparsification, eliminating matrix multiplication completely. A custom MatMul-free hardware core on a VCK190 FPGA ran a 1.5-billion-parameter model at 13.849 W and 161.8 tokens per second, improving energy efficiency 19.8 times, memory usage 21.3 times and throughput 2.2 times against an A800 GPU.

Grade 13 reading level6 min read

What happened

Transformers run on matrix multiplication. It is the dominant operation, it is why GPUs are the hardware of choice, and it is where most of the energy goes.

Several groups have proposed removing it rather than accelerating it – replacing the multiplications with additions, or with the event-driven arithmetic of spiking neural networks. Those proposals are nearly always evaluated in simulation, which can show the maths works and cannot tell you what it costs in watts.

A team at the Institute of Automation of the Chinese Academy of Sciences, the School of Artificial Intelligence at the University of Chinese Academy of Sciences, and the Brain-Inspired Models Group at the Beijing Academy of Artificial Intelligence built the hardware.

A 1.5 billion parameter model, 161.8 tokens per second, 13.849 watts.

The test

The framework transforms large language models into neuromorphic spike-based models by converting their behaviours into neural dynamics such as spike trains, through mathematical modelling complemented by quantization and sparsification.

Because a spike is a discrete event rather than a continuous value, the arithmetic reduces to additions and comparisons. Applying a hardware-algorithm co-design paradigm, matrix multiplication is eliminated entirely rather than approximated.

A custom MatMul-free hardware core was designed on a VCK190 field-programmable gate array to validate the 1.5-billion-parameter model, with dynamic power and inference throughput measured and compared against an A800 GPU.

The numbers

Measure Result
Model size 1.5 billion parameters
Hardware custom MatMul-free core on VCK190 FPGA
Dynamic power consumption 13.849 W
Inference throughput 161.8 tokens per second
Energy efficiency vs A800 GPU 19.8x improvement
Memory usage vs A800 GPU 21.3x improvement
Throughput vs A800 GPU 2.2x improvement
Matrix multiplication eliminated completely
Data public benchmark datasets on Hugging Face

The memory figure is the one worth sitting with. A 21.3 times reduction in memory usage is arguably more consequential than the energy saving, because memory capacity is what determines whether a model fits on a device at all.

The result

The implementation achieved 13.849 watts dynamic power and 161.8 tokens per second on a 1.5-billion-parameter model, improving energy efficiency by 19.8 times, memory usage by 21.3 times and inference throughput by 2.2 times compared with an A800 GPU.

The authors present the framework as addressing two problems at once: the energy consumption of large language models, and their interpretability, by conceptualising the model as neural populations that computational neuroscience tools can analyse.

How this compares to what came before

MatMul-free language models have been published before, and they are an active research line. The usual pattern is a software implementation showing that accuracy survives the transformation, with projected hardware benefits.

Projections are the weak point. A model that eliminates multiplications still has to run on hardware that was designed for them, and on a GPU a multiplication-free model does not go faster. The benefit only appears when the hardware is built to match.

That is what this paper supplies. Building the core on a VCK190 and reporting measured watts and tokens per second converts a projection into a measurement.

Spiking neural network hardware has a longer history and a consistent problem: it works well on small vision tasks and has not scaled to language. 1.5 billion parameters is a meaningful step in that direction.

What it means

The memory result reframes the significance. A 21.3 times reduction means a model that would need substantial GPU memory fits in a fraction of it, and memory capacity is the binding constraint on running language models at the edge far more often than compute is.

The throughput result is more modest. 2.2 times against an A800 is real and is not the kind of multiple that reorganises an industry. 161.8 tokens per second is usable for interactive generation and nowhere near what a datacentre needs.

The claim that matters most is the one the abstract handles least: whether accuracy survives. The paper says high performance is maintained. Spike conversion, quantization and sparsification each cost something, and three of them together usually cost more than any one alone. Without benchmark-by-benchmark numbers against the original model, the cost is unquantified.

The interpretability argument is interesting and speculative. Treating a language model as a neural population, analysable with computational neuroscience tools, is a genuinely different angle on a hard problem, and this paper proposes it rather than demonstrating it.

Business ideas from this paper

Edge language model inference. 13.849 watts and a 21.3 times memory reduction puts a 1.5-billion-parameter model within reach of devices that cannot host a GPU. Industrial equipment, vehicles and handsets are the obvious targets.

FPGA deployment services. Building a custom core on a VCK190 is specialist work. As MatMul-free models mature, the implementation gap between a published architecture and a working board is a services business.

Accuracy-preservation tooling. Converting a trained model to spikes without losing accuracy is the hard part and the least documented. Tooling that measures and minimises that loss would be used by everyone working in this area.

Interpretability through neural population analysis. If modelling LLMs as neural populations yields real analytical traction, there is a research-to-product path in AI interpretability tooling built on neuroscience methods.

What could be wrong with it

1.5 billion parameters is small, and nothing here demonstrates that the approach scales to the model sizes that matter commercially.

The comparison is against an A800 GPU running a general-purpose software stack, which is not a matched inference accelerator. Some of the 19.8 times efficiency gain is the specialisation, not the architecture.

Accuracy is reported as maintained rather than quantified against the original model’s benchmarks.

One FPGA implementation, one model, one research consortium, no independent replication.

The interpretability claim is a proposal rather than a result.

How sure can you be?

We score this 8 out of 10.

It is a peer-reviewed paper in National Science Review, with a hardware implementation measured rather than simulated, three named institutions at the Chinese Academy of Sciences, the University of Chinese Academy of Sciences and the Beijing Academy of Artificial Intelligence, and the benchmark datasets used are publicly available on Hugging Face.

It loses points because there is no independent replication and the evidence is a hardware demonstration rather than a controlled comparison.

The measurements are the strength. The missing accuracy comparison is the weakness, and it is the number we would most want before drawing conclusions.

Do this today

If you are planning edge AI deployments, the memory figure matters more than the power figure. 21.3 times less memory determines what fits; watts determine how long it runs.

If you evaluate accelerator claims, check what the baseline was. An A800 running a general software stack is a different comparison from a purpose-built inference chip, and the gap between those is often most of the claimed advantage.

If you work on model compression, the combination of spike conversion, quantization and sparsification is aggressive. Ask for the accuracy numbers before assuming performance was maintained – three lossy transformations applied together rarely come free.

Source: Neuromorphic spike-based large language model, National Science Review, February 2026. PMID 41696420 · doi.org ().

Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.

What to remember
  • The framework converts LLM behaviour into spike trains through mathematical modelling, quantization and sparsification.
  • Matrix multiplication is eliminated completely rather than accelerated.
  • Measured on a VCK190 FPGA: 13.849 W dynamic power, 161.8 tokens per second, 1.5 billion parameters.
  • 19.8x energy efficiency, 21.3x memory usage and 2.2x throughput against an A800 GPU.
Where this really is

A custom MatMul-free hardware core was designed on a VCK190 FPGA and used to run a 1.5-billion-parameter model, with measured power and throughput compared against an A800 GPU.

Questions people ask

What is a spiking neural network?

A network where information is carried by discrete events in time rather than continuous values. Because a spike is either present or absent, the arithmetic reduces to additions and comparisons instead of multiplications.

Why eliminate matrix multiplication?

It is the dominant operation in transformers and the reason GPUs are needed. Removing it rather than accelerating it changes what hardware the model can run on.

Is 1.5 billion parameters large?

Small by current standards - frontier models are hundreds of times larger. It is large enough to be a real language model rather than a toy, and it fits on one FPGA.

Does interpretability really improve?

The authors argue that modelling an LLM as neural populations lets computational neuroscience tools analyse its information encoding. That is a plausible research direction and the paper does not quantify an interpretability gain.

About the author

Mark Alex

Mark Alex is the founder and Managing Director of Real Biz Digital, a technology company operating out of Nairobi since 2018. He works in agentic AI and the Model Context Protocol, AI governance, enterprise software architecture and cybersecurity. He holds an MSc in Mechatronical Engineering from Obuda University in Budapest and a BSc in IT, Forensic Technology and Cybercrime, from USIU-Africa in Nairobi, and has published IEEE conference research on an AI-powered digital twin for greenhouse systems. He is the author of seven books. Between 2020 and 2024 he mentored more than 200 university students and interns in Nairobi. He writes every Just Out Tech article from the original research paper.