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/Hong Kong SAR, China

116 times faster and only half the energy: read both numbers

An analogue resistive memory system generated images 116.5 times faster than digital hardware - and cut energy by only 52 percent.

What the paper found

A resistive memory in-memory computing system acting as a time-continuous neural differential equation solver was validated on fabricated 180 nm macros. It accelerated unconditional generation 69.0 times and conditional generation 116.5 times against advanced digital hardware, while reducing energy by 31.5 and 52.0 percent respectively, at generative quality equivalent to the software baseline.

Grade 13 reading level6 min read

What happened

Image generation models work by removing noise in small steps. Start with static, denoise a little, repeat fifty or a hundred times, end with a picture. Each of those steps is a complete pass through a neural network, which is why generating one image costs what it costs.

The steps are an artefact. The underlying mathematics describes a continuous trajectory, and it is chopped into discrete pieces because that is how digital computers work.

A team at the Department of Electrical and Computer Engineering and the Center for Advanced Semiconductors and Integrated Circuits at the University of Hong Kong, with the School of Microelectronics at Southern University of Science and Technology in Shenzhen, built a system that does not chop it. Resistive memory integrates storage and computation and acts as a time-continuous neural differential equation solver.

Conditional generation ran 116.5 times faster. Energy fell by 52 percent.

The test

The argument the paper makes about conventional hardware has two parts. Physically separated storage and processing units cause data-transfer bottlenecks – the familiar memory wall. And discrete operations disrupt naturally continuous generation dynamics – which is the less familiar claim and the one this architecture addresses.

Resistive memory solves both at once by putting computation where the weights are and by evolving continuously rather than stepping.

Validation was experimental, on fabricated 180 nanometre resistive memory in-memory computing macros. Both unconditional and conditional generation were measured for speed and energy against advanced digital hardware, with generative quality compared to the software baseline.

The numbers

Measure Unconditional generation Conditional generation
Speedup vs advanced digital hardware 69.0x 116.5x
Energy reduction 31.5% 52.0%
Generative quality equivalent to the software baseline
Process node 180 nm resistive memory macros
Validation experimental, on fabricated macros
Data main text, Supplementary Information, processed datasets on GitHub

The result

The system accelerated unconditional generation by a factor of 69.0 and conditional generation by 116.5, compared with advanced digital hardware, while maintaining generative quality equivalent to the software baseline.

Energy consumption fell by 31.5 percent and 52.0 percent respectively.

The authors frame this as expanding the horizon for hardware solutions in edge computing for generative AI.

How this compares to what came before

Analogue in-memory computing for neural networks is a mature research area, and the standard demonstration accelerates matrix multiplication for inference in a classifier. Those results report energy per multiply-accumulate and typically claim large efficiency gains.

What is different here is matching the hardware’s natural behaviour to the algorithm’s natural form. Resistive memory evolves continuously; diffusion models are continuous processes that we discretise for digital convenience. Removing the discretisation is a better fit than simply accelerating each discrete step.

Against the usual claims for in-memory computing, this result is notable for an unusual reason: the energy saving is modest. Most analogue computing papers lead with efficiency multiples in the tens or hundreds. Reporting 31.5 and 52.0 percent, next to speedups of 69 and 116.5, is a more careful accounting than the field’s norm.

What it means

Read both numbers together and the picture is informative.

A 116.5 times speedup with a 52 percent energy reduction means the system finishes much sooner while still doing comparable total work. That is valuable – latency matters for interactive generation – and it is not the order-of-magnitude efficiency revolution that analogue computing is often promised to deliver.

The 180 nanometre node cuts the other way. That process is roughly a quarter-century old. Getting these figures against advanced digital hardware from such an old node suggests the architectural advantage is real and that a modern-node implementation would do better. It also means the comparison is not like for like, and the exact multiples should not be quoted as if it were.

For edge generative AI, the latency result is the useful one. Running a diffusion model on a device rather than a data centre is bounded by how long a user will wait, and a hundredfold reduction in that time changes what is feasible.

Business ideas from this paper

On-device generative models. The latency barrier is what keeps image generation in the cloud. A hundredfold speedup at the edge changes the privacy and cost calculation for any application that currently sends user content to a server.

Analogue accelerators for continuous algorithms. The general principle – match the hardware’s physics to the algorithm’s mathematics – applies to other continuous processes: physical simulation, optimal control, differential equation solvers.

Honest efficiency benchmarking. This paper separates speed from energy. Most do not. A benchmarking standard that forces both to be reported, on a stated process node, would be useful to every buyer in this space.

Process node migration. Taking a validated analogue architecture from 180 nm to a modern node is specialised engineering work that device research groups are not set up to do.

What could be wrong with it

The comparison crosses process generations. 180 nanometre macros against advanced digital hardware is architecture against architecture, and the speed and energy multiples are not silicon-to-silicon.

Generative quality is reported as equivalent to the software baseline without the standard image quality metrics in the abstract. For generative models, that is the claim most in need of numbers.

The energy reductions of 31.5 and 52.0 percent are far smaller than the speed multiples, which is reported clearly and is worth noticing when the headline is quoted.

Analogue computing has inherent precision limits. How those interact with generative quality over long trajectories is the question, and it is not resolved here.

One research consortium, no independent replication.

How sure can you be?

We score this 8 out of 10.

It is a peer-reviewed paper in Nature Communications, experimentally validated on fabricated hardware rather than simulated, with three named institutions at the University of Hong Kong and Southern University of Science and Technology, and processed datasets available on GitHub.

It loses points because there is no independent replication and the evidence is a laboratory hardware demonstration.

What raises our confidence is the reporting of two different multiples for speed and energy. A paper that wanted to oversell would have led with 116.5 and left the energy figure in a table.

Do this today

If you build anything that generates images on a device, track analogue in-memory computing. The latency barrier is the thing keeping generation in the cloud and this is the clearest attack on it we have seen.

If you evaluate accelerator claims, insist on the process node. A multiple quoted against hardware two decades newer or older is not a measurement of the architecture.

If you work on generative models, the continuous-trajectory framing is worth understanding independently of the hardware. The discretisation into steps is a convenience, and anything that exploits the underlying continuity has a structural advantage.

Source: Resistive memory-based neural differential equation solver for score-based diffusion model, Nature Communications, 2026. PMID 42120416 · 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
  • Resistive memory integrates storage and computation, acting as a time-continuous differential equation solver.
  • 69.0 times faster unconditional generation, 116.5 times faster conditional, against advanced digital hardware.
  • Energy fell 31.5 and 52.0 percent - a much smaller multiple than the speed figures.
  • Validated experimentally on 180 nm macros, which is an old process node.
Where this really is

The system was validated with 180 nm resistive memory in-memory computing macros, with measured speed and energy against digital hardware baselines.

Questions people ask

What is a score-based diffusion model?

A generative model that starts from noise and removes it step by step to produce an image. Each step is a separate pass through the network, which is why generation is slow and expensive.

What does treating it as a differential equation change?

The denoising trajectory is continuous in theory and chopped into discrete steps only because digital hardware works that way. Analogue memory can evolve continuously, so the steps collapse.

Why is the energy saving so much smaller than the speedup?

Because finishing faster is not the same as doing less work. The speedup comes partly from parallelism and continuous evolution; the energy still has to be spent on the computation and on the peripheral circuitry around the analogue core.

Is 180 nm a disadvantage?

It is an old node by current standards. The comparison against advanced digital hardware is therefore architecture versus architecture, not silicon versus silicon, and a modern-node implementation would presumably do better still.

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.