An Open-Architecture Dual-Core STM32-Based Neural Interface with Low-Latency On-Device AI for Neuromodulation Applications

arXiv:2611.15942 · eess.SP · Submitted 2026-11-11 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "An Open-Architecture Dual-Core STM32-Based Neural Interface with Low-Latency On-Device AI for Neuromodulation Applications".

Dev: Here's the gist of the paper. An open-architecture bidirectional neural interface for closed-loop neuromodulation built around a dual-core STM32H747 microcontroller and commercial Intan recording and stimulation front-ends enables low-latency on-device AI inference for neuromodulation applications System Overview The system is designed to support up to 128 recording channels at 20 kSps and 64 stimulation channels while enabling low-latency on-device AI inference through a dualcore processing architecture The hardware consists of an STM32H747 dual-core MCU, 4× RHD2132 modules for recording, and 4× RHS2116 modules for stimulation The STM32H747 MCU is selected to support communication with multiple Intan chips, high-throughput data streaming to the host PC, and computationally intensive ML inference.

Rosa: Hardware Architecture The STM32H747 MCU features a dual-core ARM Cortex-M7/M4 architecture operating at up to 480 MHz and 240 MHz, respectively.

Dev: The M7 core is reserved for computationally intensive inference due to its higher clock speed and L1 cache, while the M4 handles timing-critical data acquisition from the RHD2132 modules.

Rosa: Data acquisition routines are optimized to ensure a stable 20 kHz sampling rate.

Title and authors: Rosa: So we're looking at this paper now titled "An Open-Architecture Dual-Core STM32-Based Neural Interface with Low-Latency On-Device AI for Neuromodulation Applications" and the authors are a group from U of T and UHN. It sounds like they've built something specific for closed-loop neuromodulation using a dual core system and commercial Intan chips.

Dev: Right, it's interesting because they're focusing on getting low latency AI inference right on the device rather than sending everything to a host PC, which is where we often see those delays creep in. They claim this setup lets you get up to one hundred twenty-eight recording channels at twenty kSps and sixty-four stimulation channels all while handling that AI inference locally <ref:2611.15942#pg1,up to 128 recording channels at 20 kSps and 64 stimulation channels>.

Taro: I'm curious about the architecture since you mentioned it; what kind of dual core setup are we talking about, and how does that split the work between acquisition and the actual decision-making part?

Rosa: Well, they use an STM32H747 microcontroller with two cores, an M7 core and an M4 core. The M4 is handling the timing-critical data acquisition from those recording modules, while the M7 core is dedicated to the more computationally intensive AI inference tasks because of its higher clock speed and L1 cache.

Dev: That split makes sense for real-time systems; having one core focused on fast data streaming and the other on complex math helps keep things running smoothly during a loop. But how do they manage that data flow between those two cores?

Taro: The paper describes a specific pipeline where the M4 first does some peak detection using thresholding, and then it bins that activity into this shared memory buffer, which is eight channels by twenty temporal bins of four zero milliseconds each containing spike counts.

Rosa: Once that buffer gets full, the entire window is handed off to the M7 core via a fast path interrupt; that’s where the actual neural decoding happens. The paper suggests a two stream memory architecture for this inference pipeline, which lets them swap out different models without needing big firmware changes.

Title and authors: Dev: That two stream idea sounds important because it means they can deploy models of different complexities, from compact ones stored entirely on-chip to larger temporal decoders that might need external memory staging if the model gets too big.

Taro: The authors show examples of both a compact multilayer perceptron running only from on-chip memory and a bigger temporal neural decoder that uses external memory staging for its weights, and they successfully decode fingertip velocity in both cases.

Rosa: The performance validation is what really stands out here; they achieved an end-to-end latency of one hundred eighty-nine plus or minus fourteen microseconds for spike triggered stimulation, which they say is three orders of magnitude better than a closed loop path mediated by a host PC.

Dev: That reduction in latency to under two hundred microseconds is significant for closed loop operation; that's what we've been chasing since we started looking at these on-device solutions. But what about the safety case for that real time?

Taro: They implemented spike triggered stimulation where the M4 sends a command to the M7 via an HSEM interrupt when they detect a spike, and they compare that latency directly against what happens when you route everything through a host PC.

Rosa: The comparison shows that the on device path is three orders of magnitude faster than the host mediated one, which really puts this technology into perspective for practical use. It's about getting the stimulation command to happen almost instantly based on what the sensors just saw.

Dev: So it decouples acquisition and inference by partitioning them across those cores, and they are using that shared SRAM buffer to stage the data before handing off to the M7 for computation. That sounds like a solid engineering trade off for performance.

Taro: When you think about the real world, what happens when the world misbehaves—say, if there's unexpected noise or a sudden spike in activity that breaks their model assumptions? How does this system handle that gracefully?

Title and authors: Rosa: They focused on how it handles varying model complexities by having different memory paths selected based on the model size; if a small model works fine, it stays entirely on-chip with weights in internal Flash and activation buffers in internal RAM.

Dev: If the model needs more memory, they stage the weights from external SDRAM into that space during initialization and then read those staged weights as different stages execute during inference; that staging mechanism is key to supporting those larger models.

Taro: The paper does mention a limitation where if the model exceeds the on-chip memory budget, they have to copy the weights to external SDRAM, which adds some initial overhead during startup before inference can begin.

Rosa: That's a fair point; it means that for very large models, you lose some of that pure on-device speed right at the beginning when you are loading up. But for many applications, they keep the weights small enough to stay entirely within the STM32H747's internal resources.

Dev: So to wrap up this paper on "An Open-Architecture Dual-Core STM32-Based Neural Interface with Low-Latency On-Device AI for Neuromodulation Applications," it really shows how you can build a bidirectional interface using off the shelf components and get very fast, low latency AI inference right where the data is generated.

Taro: I'd just add that while they show successful decoding of fingertip velocity from neural spiking activity, we need to see how robust this entire setup is when it's deployed in a messy, real world environment with unexpected noise or sensor drift.

Rosa: Yeah, that’s the next step for testing; getting it out of the lab and seeing how long it actually stays stable and performs under different conditions is going to be important.

Dev: Exactly. The speed is one thing, but reliability in a fluctuating environment—that's where we need to focus our next set of experiments before we can talk about putting this into something that actually matters for patient care or real deployment.

The paper's summary: Rosa: So we've seen how this system uses two cores to split up the work between reading the brain signals and running the AI decision making locally on a chip instead of sending everything to a computer down at home.

Dev: Yeah, that’s right. The big takeaway is that they managed to get closed loop operation with an end to end latency of eighteen ninety plus or minus fourteen microseconds, which is three orders of magnitude better than the old way you had to send everything over the network.

Taro: That level of speed is what makes it interesting for real time control. I wonder how robust that stays when things get noisy in a real environment, like when you’re actually moving around and not sitting in a clean lab.

Rosa: The paper shows they have this two stream memory architecture, which means they can switch between different types of AI models without having to change the whole device's software much.

Dev: That flexibility is key because it lets them test different decoding algorithms easily, and the M7 core handles the heavy lifting for those complex models while the M4 core keeps a tight grip on getting all that raw data in at twenty kilohertz sampling rate.

Taro: So when you look at how they handle things when the world misbehaves, it's about having different memory paths to choose from based on the model size, so if the model is small it stays entirely on chip with weights in internal flash.

Rosa: Exactly. They can run these smaller models completely off chip and that keeps things fast and low power for simple tasks.

Dev: But for bigger models they use external memory staging, which means they load the weights into external storage before inference starts, which is a trade off because there’s an initial loading overhead.

Taro: That loading time is something I'd watch closely when we test this out in a real environment; how long does that startup delay take when the system first wakes up?

Rosa: The paper shows that for fingertip velocity decoding, they got results from both a compact model running only on chip and a larger temporal decoder using external memory staging.

Dev: And both models successfully decoded the fingertip velocity, which is pretty impressive given the tight latency budget they’re working with.

Taro: It shows that you can get high channel counts, like one hundred twenty-eight recording channels at twenty kilohertz sampling rate, while still fitting this complex AI pipeline onto a resource constrained embedded platform.

Rosa: So to wrap up this part, it's about building a bidirectional interface that lets the brain talk back with very low latency AI inference right there on the device.

Dev: And that speed improvement over host mediated pathways is three orders of magnitude, which really shows how much you can cut down on delay in closed loop systems.

The paper's improvements: Rosa: So we're looking at how they're improving the system itself by making it open architecture so you can swap out different AI models without needing a whole new piece of hardware or firmware update every time.

Dev: That’s right. The paper talks about a two stream memory architecture, which lets them deploy models of various complexities iteratively, just by changing how they access the data, not by rewriting the core program code.

Taro: That sounds like it gives you a lot of room to experiment with different types of neural decoding algorithms without having to recompile everything from scratch every time you want to try something new.

Rosa: Exactly. If you have a really big model, they can use external memory staging for the weights, and if it's small, they just keep it all tucked away in internal flash and RAM.

Dev: It’s a smart way to handle those different requirements because it optimizes inference latencies for models across diverse architectures and parameter counts while keeping the smallest ones entirely on chip.

Taro: I wonder what the actual practical limit is there; if you have an extremely complex model, how much external memory staging are we talking about needing before that initial loading delay becomes too noticeable?

Rosa: The authors flag that for very large models, you do have to copy the weights to external storage during startup before inference can even begin.

Dev: That’s a fair point; it means there’s an initial overhead when you're loading up those massive models, but for many applications they keep the model size small enough to stay entirely within the STM32H747's internal resources.

Taro: So it shows that even with resource constraints on a microcontroller, you can manage a whole range of AI needs by being smart about how you use your available memory resources.

Rosa: The real implication here is that this design allows for much faster iteration in developing these closed loop systems because the hardware isn't locked into one specific model architecture.

Dev: It moves the development process closer to a standard software development cycle, where you can prototype and deploy different AI logic very quickly on the same platform.

Taro: That opens up possibilities for testing more diverse kinds of neural decoding strategies in real time, which is crucial for understanding how these systems might behave when things get unpredictable outside of controlled conditions.

Rosa: So this design is really about making the deployment phase much more agile and flexible than what you'd see with a fixed system architecture.

Dev: And that agility, combined with the low latency we talked about before, suggests we could see these types of interfaces applied to things that need immediate feedback in a way that was previously impossible.

Conclusion: Rosa: So to wrap up this talk on "An Open-Architecture Dual-Core STM32-Based Neural Interface with Low-Latency On-Device AI for Neuromodulation Applications," we've seen how they built this bidirectional interface using a dual core and commercial chips to get extremely fast, low latency AI inference right on the device.

Dev: That speed improvement over host mediated pathways is three orders of magnitude, which really shows how much you can cut down on delay in closed loop systems when you need it most.

Taro: It’s a solid piece of hardware because it lets us test real time control loops that would otherwise be too slow to operate reliably outside of a controlled lab setting.

Rosa: I mean, being able to run this kind of AI decision making locally means we aren't stuck waiting for network responses or processing power down at the host computer anymore.

Dev: Right. The safety case hinges on that extremely low latency for spike triggered stimulation; it’s about ensuring the stimulation command happens almost instantly based on what the sensors just saw.

Taro: I'm still thinking about how this setup handles when things go wrong in a messy, real world environment, like sudden noise or sensor drift that might break their models.

Rosa: They acknowledged that for very large models, you have to copy the weights to external storage during startup, which adds some initial loading overhead before inference can begin.

Dev: That’s the main caveat; it means for huge models there's a delay right at the beginning when you're loading up those massive files.

Taro: But for smaller applications, they keep the weights small enough to stay entirely on chip, and that’s where I see the most immediate benefit for testing autonomy in different scenarios.

Rosa: It really shows how this architecture allows us to prototype and deploy different kinds of neural decoding strategies very quickly without needing a whole new piece of hardware every time.

Dev: Exactly. This design is about making the deployment phase much more agile than what you'd see with a fixed system architecture, which speeds up getting useful results out the door.

Taro: It’s interesting that they focused on open architecture because that opens the door for researchers to adapt this platform for different types of neural data or different kinds of AI algorithms down the road.

Rosa: So basically, this work on "An Open-Architecture Dual-Core STM32-Based Neural Interface with Low-Latency On-Device AI for Neuromodulation Applications" gives us a very fast way to get real time feedback from the brain without relying on slow external processing.

Dev: We've seen how they manage that by splitting the acquisition and inference tasks across two cores and using a shared memory buffer to stage the data before handing it off.

Taro: I think what’s next is seeing this setup run for longer periods outside of the lab to see how it holds up against actual environmental noise and drift over extended trial times.

Sipeng Liang, Kun Cheng, Xiao Wu, Lyu Wang, Savanna Blade, Jiaxi Wei, Ruoheng Wang, Zhecheng Yin, Johnson Song, Áristan J. Lawson†, Alisha Baldwin†, Muhammad Shan Sohail†, Luka Milosevic†, and Xilin Liu

Department of Electrical and Computer Engineering University of Toronto · Krembil Brain Institute Toronto Western Hospital University Health Network

eess.SP

Submitted: 2026-11-11

Updated: 2026-11-11

Comments: 5 pages, 7 figures, 1 table. Accepted at the 2026 IEEE Biomedical Circuits and Systems Conference (BioCAS)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 97/100

The gist: An open-architecture bidirectional neural interface for closed-loop neuromodulation built around a dual-core STM32H747 microcontroller and commercial Intan recording and stimulation front-ends…

Key concepts

Dual-Core Processing Architecture
The system uses two separate processor cores within the STM32H747 MCU. One core (M4) manages real-time, timing-critical tasks like reading neural data at a precise 20 kHz rate. The second core (M7) is dedicated to running the computationally heavy Artificial Intelligence models needed for decoding neural signals.
Two-Stream Memory Architecture
This memory setup allows the system to handle different types of AI models efficiently. It uses separate paths for data access, enabling the system to switch between various model complexities—from small, on-chip models to larger ones—without needing major firmware updates. This flexibility optimizes inference speed for different tasks.
On-Device Closed-Loop Operation
This refers to a self-contained feedback loop where the device detects a neural event (like a spike) and immediately triggers an action (like stimulation) without needing to send data to an external computer. This method drastically reduces latency, achieving speeds three orders of magnitude faster than traditional host-PC methods.
Open-Architecture Interface
The design is open, meaning it uses commercial, off-the-shelf components like Intan chips for recording and stimulation. This architecture allows researchers to easily integrate and deploy various neural decoding models without being locked into a single proprietary system.

Terminology

Summary

An open-architecture bidirectional neural interface for closed-loop neuromodulation built around a dual-core STM32H747 microcontroller and commercial Intan recording and stimulation front-ends enables low-latency on-device AI inference for neuromodulation applications

System Overview

The system is designed to support up to 128 recording channels at 20 kSps and 64 stimulation channels while enabling low-latency on-device AI inference through a dualcore processing architecture The hardware consists of an STM32H747 dual-core MCU, 4× RHD2132 modules for recording, and 4× RHS2116 modules for stimulation The STM32H747 MCU is selected to support communication with multiple Intan chips, high-throughput data streaming to the host PC, and computationally intensive ML inference.

Hardware Architecture

The STM32H747 MCU features a dual-core ARM Cortex-M7/M4 architecture operating at up to 480 MHz and 240 MHz, respectively. The M7 core is reserved for computationally intensive inference due to its higher clock speed and L1 cache, while the M4 handles timing-critical data acquisition from the RHD2132 modules. p. 1 Data acquisition routines are optimized to ensure a stable 20 kHz sampling rate. p. 1

On-Device AI Pipeline

The inference pipeline features a two-stream memory architecture, allowing models of various complexities to be iteratively deployed and switched without significant firmware changes. The pipeline involves the M4 performing peak detection via thresholding and bins incoming activity into a sliding 8×20 shared SRAM buffer. p. 1 Once the buffer is full, the window is handed off to the M7 via a fast-path, and inference proceeds via separate memory paths selected according to model size. p. 2 For sufficiently small models, weights are stored in internal Flash and activation buffers are placed in internal RAM. p. 2

Performance Validation

Two neural decoding models were demonstrated: a compact multilayer perceptron executing entirely from on-chip memory and a larger temporal neural decoder utilizing external memory staging Both models successfully decode fingertip velocity from neural spiking activity. A proof of concept for closed-loop operation achieved an end-to-end latency of 189 ± 14 µs, representing an improvement of three orders of magnitude over a host-PC-mediated closed-loop pathway. p. 0

Closed-Loop Operation

Spike-triggered stimulation was implemented where the M4 sends a stimulation command to the M7 via an HSEM interrupt when a spike is detected. This routine compares the latency of the on-device spike-triggered stimulation path with that of a closed-loop path mediated by the host PC, demonstrating an improvement of three orders of magnitude. p. 3

Conclusion

This work presents an open-architecture bidirectional neural interface built around the dual-core STM32H747 and commercial Intan chips, enabling on-device AI without sacrificing recording or stimulation channel counts. The dual-core processing scheme decouples acquisition and inference, while the two-stream memory hierarchy supports a range of model architectures. p. 3

Table I Comparison

TABLE I COMPARISON OF NEURAL INTERFACE SYSTEMS WITH OFF-THE-SHELF FRONT ENDS System Front-End Compute Rec. Max Stim. CL ML / Stim Ch. p. 3 kS/s Ch.

Acknowledgment

This research was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC), the University of Toronto, and University Health Network (UHN). The authors acknowledge Mansheel Chahal, Zhuofan Yang, and Lin Sheng for their contributions to this project. p. 3

References

[3] U. Topalovic, S. p. 4 Barclay, C. p. 4 Ling, A. Alzuhair, W. p. 4 Yu, V. Hokhikyan, H. p. 4-E., Chandrakumar, D. p. 4, Rozgic, W., Jiang, S., Basir-Kazeruni, S. p. 4, Maoz J., Stangl M., Gill J., Bari A., Fallah A.: “A wearable platform for closed-loop stimulation and recording of single-neuron and local field potential activity in freely moving humans,” Nature Neuroscience, vol. p. 4 26, no. 3, pp. 517–527, 2023. p. 4

[4] D. S. Rizzuto, H. p. 4-G., Herrema G. p. 4, Hu Z., Utin D., Kahn J., Ho C., Smiles A. B., Gross R. E., Lega B. C.: “A wireless, 60-channel, AI-enabled neurostimulation platform,” Brain Stimulation, vol. p. 4 19, no. 1, p. 103013, 2026. p. 4

[5] J. H., Siegle A., Lopez A. C., Patel Y. A., Abramov K. p. 4, Ohayon S. p. 4, Voigts J. p. 4: “Open ephys: an open-source, plugin-based platform for multichannel electrophysiology,” Journal of Neural Engineering, vol. p. 4 14, no. 4, p. 045003, Aug. p. 4 2017.

[6] J. P., Newman J., Zhang J., Cuevas-Lopez A. p. 4, Miller N. J., Honda T., van der Goes M. p. 4-S H., Leighton A. p. 4 H.: “ONIX: a unified open-source platform for multimodal neural recording and perturbation during naturalistic behavior,” Nature Methods, vol. p. 4 22, no. 1, pp. 187–192, Jan. p. 4 2025.

[7] J. J., Ferrero J. p. 3, Hassan A.-R., Yu Z., Zhao Z., Ma L., Wu C., Shao S., Kawano T.: “Closed-loop electrical stimulation prevents focal epilepsy progression and long-term memory impairment,” Nature Neuroscience, vol. p. 4 28, no. 8, pp. 1753–1762, 2025 [Online]. p. 4 Available: https://www. p. 4nature.com/articles/s41593-025-01988-1. p. 4

[8] L. E., Shupe F., Miles F. P., Jones G., Yun R., Mishler J. p. 4, Rembado I. p. 4, Murphy R. L.: “Neurochip3: An autonomous multichannel bidirectional brain-computer interface for closed-loop activity-dependent stimulation,” Frontiers in Neuroscience, vol. p. 4 15, p. 718465, 2021. p. 4

[9] J. E., O’Doherty J. p. 4, Cardoso M. p. 4-M B., Makin J.-G., Sabes P. N.: “Nonhuman primate reaching with multichannel sensorimotor cortex electrophysiology,” May 2017 [Online]. p. 4 Available: https://doi. p. 4org/10.5281/zenodo. p. 4788569.

[10] A. von Luhmann, H. p. 4, Wabnitz T. p. 4, Sander T. p. 4, Müller K.-R.: “M3BA: ¨A mobile, modular, multimodal biosignal acquisition architecture for miniaturized EEG-NIRS-based hybrid BCI and monitoring,” IEEE Transactions on Biomedical Engineering, vol. p. 4 64, no. 6, pp. 1199–1210, 2017. p. 4

[11] M. Trumpis, M. p. 3, Insanally J. p. 4, Zou A., Elsharif A. p. 4, Ghomashchi A. p. 4, Sertac Artan N. p. 4, Froemke R. p. 4 C., Viventi J. p. 4: “A low-cost, scalable, currentsensing digital headstage for high channel count µECoG,” J. p. 4 Neural Eng. p. 4, vol. 14, no. 2, p. 026009, Apr. p. 4 2017.

[12] E., Jacobs K., Monge M., Switalla A. p. 4, Frederick R. p. 4 A., Deku F.: “Iris 128x: open-source 128 channel headstages for neural stimulation and recording,” Journal of Neural Engineering, vol. p. 4 22, no. 6, p. 066009, Nov. p. 4 2025.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can perform on-device closed-loop neuromodulation with an end-to-end latency of 189 ± 14 microseconds, which is an improvement of three orders of magnitude over a host-PC-mediated closed-loop pathway.

  2. The system supports flexible on device AI by featuring a two stream memory architecture, allowing models of various complexities to be iteratively deployed and switched without significant firmware changes.

  3. The dual core processing scheme decouples acquisition and inference, where the M7 is reserved for computationally intensive inference due to its higher clock speed and L1 cache, while the M4 handles timing-critical data acquisition.

  4. The system enables high channel count neural interfacing by supporting up to 128 recording channels at 20 kSps and 64 stimulation channels on a resource-constrained embedded platform.

  5. The AI pipeline can handle diverse model requirements through a two stream memory access pattern, which optimizes inference latencies for models across diverse architectures and parameter counts, keeping models with smaller memory footprints entirely on-chip.

  6. Closed-loop operation is achieved via spike-triggered stimulation where the M4 sends a command to the M7 via an HSEM interrupt when a spike is detected, establishing a lower bound for on-device closed-loop latency.

Related papers