Competition 2024
Competition: Hardware Implementation

Arrhythmia Analysis Accelerator : A-Cube

We propose the A-Cube design methodology to create medical decision support on the edge. The design and implementation of an atrial fibrillation detector hardware core was selected as a proof-of-concept study. To facilitate the required atrial fibrillation functionality, we adopted an established AI model, based on Long Short-Term Memory (LSTM) technology for hardware implementation. The adaptation was done by varying design parameters such as data window and the number of LSTM units. We found that a data window of 40 beats and 20 LSTM units are sufficient to achieve a classification accuracy of 99.02%. We are confident that the A-Cube methodology can be used to implement this model in hardware. Doing so, will create a low power and low latency atrial fibrillation monitoring solution which has the potential to extend the observation duration while being convenient for patients.

 

Project Milestones

  1. Milestone #1

    Target Date

    Implement and test the Soclab encryption example in the ZCU104 FPGA board.

  2. Milestone #2

    Target Date

    Implement an atrial fibrillation detection model in the ZCU104 FPGA board.

  3. Milestone #3

    Target Date

    Analyse the model performance based on different signal length and algorithm complexity.

  4. Milestone #4

    Target Date

    Analyse the model performance based on different quantisation levels.

  5. Milestone #5

    Target Date

    Select a suitable implementation candidate to server as atrial fibrillation detection core.

  6. Milestone #6

    Target Date

    Simulate (RTL) the selected atrial fibrillation detection core.

  7. Milestone #7

    Target Date

    Implement an AHB bus interface for the selected atrial fibrillation detection core.

  8. Milestone #8

    Target Date

    Integrate Soclab and the selected atrial fibrillation detection core.

  9. Milestone #9

    Target Date

    Simulate the integration results.

  10. Milestone #10

    Target Date

    Implement the integration results in the ZCU104 FPGA board.

  11. Milestone #11

    Target Date

    Test the hardware on the ZCU104 FPGA board.

  12. Milestone #12

    Target Date

  13. Milestone #13

    Target Date

Comments

Hi,

Great to see the project live. It is very easy to add the initial milestones and you can come back as each milestone progresses and make updates. 

IITH are a good example of milestones in their project https://soclabs.org/project/real-time-edge-ai-soc-high-speed-low-complexity-reconfigurable-scalable-architecture-deep

You can simply select milestones from the generic design flow stages or create any unique milestones.

add milestones via edit of the Project

The HLS4ML environment can use various backend tool chains for High Level Synthesis (HLS). These tools support the generation of variations of hardware implementation for an algorithm written in a high level language portable language. The benefit of High Level Synthesis is that it bridges hardware and software design domains. 

It was good to get some details of your design environment. As I understand things you are using the Xilinx® Vivado® High-Level Synthesis (HLS) tool chain as your backend. HLS4ML uses streams to pass data and these are mapped into implementation buffers. As with other abstractions it provides an accelerated design environment but then requires effort to optimise the hardware implementation. This can be done by developing optimization directives that direct the tool chain to refine the automated generation of the implementation.

Alternatively some design teams move to use the next layer of tooling down in the design abstraction environment once they have a conceptual design. The data model for your project is relatively easy to understand. It was interesting discussing the use of the  Xilinx® HLS streaming interface.  We also discussed some of the challenges of moving beyond the FPGA fabric support into the full ASIC implementation flow and the subtle differences in performance of logic and memory between the two hardware instantiations. 

It will good to hear how things are progressing. 
 

If you are still planning to use HLS4ML to generate a model you might be interested in one of last year's projects. I have added a few comments on what has been happening on Aba’s project by and the design challenges.

Enhancing HLS4ML: Accelerating DNNs on FPGA and ASIC for Scientific Computing | Review

Is my current overview of their project environment and

Enhancing HLS4ML: Accelerating DNNs on FPGA and ASIC for Scientific Computing | Bridge

Is an overview of the bridge issues from nanosoc to their custom accelerator

 

The workflow proposed by Aba and the team in their project is developing but still has some issues being resolved to map the model to the nanoSoC reference design when dealing with the specifics of physical design for a specific technology node and ensuring there are no DRC violations. The section 'Workflow of Our system' is helpful in giving an overview. You should be able to get some understanding of the next milestones, I think these are 6,7,8.  

One significant difference I am aware of is that the volume of data Aba and team are looking to move through the system is such they plan to use DMA engines to offload the data movement across the bus. In your case you have said the data volume is low and so you don't plan to use DMA. I am assuming the firmware on the M0 will handle the transfers?

In this interesting Siemens EDA presentation on Catapult + HLS4ML for Inference at the Edge it is clearly shown why generating the right intermediate code is important to allow efficient ASIC implementation.

Example of generated code that does not allow loop unroll for efficient hardware implementation

David Burnette in this presentation illustrates a hardware design pattern, Sliding Window, and how generating the appropriate code from the HLS4ML environment allows the design pattern to be used to make an efficient implementation.

Example of generated code that does allow loop unroll for efficient hardware implementation

Looking at this diagram from the ML Developers Guide for Cortex-M Processors and Ethos-U NPU your application is in the vibration detection class I suspect.  As you are planning on using DMA then the M0 processor will handle the data movement through your custom accelerator.  

Diagram from Arm showing the various M class processors and the ML applications that are bet suited

                Copyright © 1995-2024 Arm Limited (or its affiliates). All rights reserved

I recently updated the Interest on Firmware.  I suspect you will need to develop a device driver for your custom accelerator to handle the interactions and interrupts.

Perhaps a Milestone 14 could be added? Now that Europractice have set out their schedule for 2025 fabrication shuttles then the 16th April mini@sic shuttle looks a good candidate for a nanoSoC version 2 tape out. 

Europractice mini@asic shuttle schedule as of November 2024

 

Is there a Milestone for the implementation of the Hyperbolic tangent and sigmoid functions ? When will this be complete?

Hi,

It was great to hear today how you worked through the analysis to come to the view that the data window of 40 beats and 20 LSTM units are sufficient to achieve a classification accuracy of 99.02%. It was also interesting to explore how the data movement across the bus and into your accelerator will form based on this requirement.

Different people have taken similar but slightly varying approaches to how the data transactions flow into their working registers within their accelerators. There are a few projects that can act as examples:

SHA-2 Accelerator Engine | SoC Labs

Given the volume of data flowing I think you concluded the Master for the bus was likely the M class CPU and not a DMA controller.

All sounds very promising and it will be great to understand more on the design and expected milestones.

Hi,

It was good to open the discussion on what will be your frame work for creating test benches. There are a variety of approaches people are adopting. The Verification pathway parallels the Design pathway through the Design Flow  stages. There are even standarisation efforts in this area: 

Universal Verification Methodology | SoC Labs

It often comes down to experience of various tools and languages that are easily adopted, or the ability to re-use prior verification assets that reduce the overall project timeline rather than extend it with new skills acquisition for unknown techniques. 

There are some  verification assets in SoC Labs and some projects to continue to develop these.

System Verification of NanoSoC | SoC Labs

I think you had some specific focused test bench requirements you wanted to work on for the data flow across the bus. I guess this is part of the usual ebb and flow of design, either top down for the SoC systems architecture or bottom up from foundation IP blocks.

Looking forward to seeing how these develop on your project.

John. 

We propose the A-Cube design methodology to create medical decision support on the edge. The design and implementation of an atrial fibrillation detector hardware core was selected as a proof-of-concept study. To facilitate the required atrial fibrillation functionality, we adopted an established AI model, based on Long Short-Term Memory (LSTM) technology for hardware implementation. The adaptation was done by varying design parameters such as data window and the number of LSTM units. We found that a data window of 40 beats and 20 LSTM units are sufficient to achieve a classification accuracy of 99.02%. We are confident that the A-Cube methodology can be used to implement this model in hardware. Doing so, will create a low power and low latency atrial fibrillation monitoring solution which has the potential to extend the observation duration while being convenient for patients.

Hi,

It would be good to get an update on your milestones at the end of the month.

I think you have been working on the interface Atrial Fibrillation Detection Accelerator core to main Nanosoc. As I understand things, you were considering the need for a DMA for data transfer. If you do not use a DMA the processor will have to initiate all transfer of data between your sensor input and your custom accelerator. 

Let us know where you have got to?

John.

Hi Rami,

We've just updated the way you can configure nanosoc, to hopefully make it more easy for people to make changes to the overall configuration without having to change multiple files. If you pull the latest changes you should see a nanosoc.config in the nanosoc_tech directory

If you leave the options blank, they will not be included and if you write a yes (or really any text after) then it will be included. There is also a section to edit the paths to arm IP

You should now not have to make any edits to flists (apart from to include your accelerator) or to any of the makefiles

Let me know how it goes

Daniel

It was good to hear you confirm the Architectural design pattern for data movement within your proposed SoC. It was interesting talking through the various options:

  1. Using the processor to Poll status registers for results. Using the processor to periodically inspect status registers associated with the processing of the external data for both data handling and obtaining results can consume unnecessary processor cycles and may not allow the processor to enter low power modes. Polling might be needed during initiation sequences before the main operating cycles of the SoC are established.
  2. Using an Interrupt driven approach where by your custom hardware raises an interrupt request signal on the processor when it requires service. The processor then accesses the status registers only when required and avoids the overhead of period polling. The processor is still active in managing all the data transfer in the system. 
  3. Using a separate Direct Memory Access (DMA) controller as an Initiator (Master) for data movement transactions. Interrupts on the processor are raised only when additional processing needs to be undertaken. 

Given the low data bandwidth requirements for you specific project you felt the DMA approach was not needed. A simple interrupt architectural design pattern with a mapping of registers within your custom hardware into the address space of the processor and accessible from the main system bus was I recall your proposed design. This will be a combination of data, status and result registers.

It will be good to see how this develops.

Atrial Fibrillation (AF) is a common heart rhythm disorder which increases the stroke risk fivefold . It is estimated that AF prevalence in the UK is around 1% of the general population and increases to 10% within the older population. AF is diagnosed by analysing the electrocardiogram (ECG) of the human heart. Figure 1 shows an ECG signal before and during an AF episode.   This analysis can be done manually, which is labour intensive or with artificial intelligence (AI) models.  Currently the AI models are executed in cloud servers and the data must travel from the point of measurement to that server over a network. Communicating the data requires network connectivity and energy expenditure associated with accessing the network. Furthermore, communication channels pose data safety and security issues. Moreover, a setup procedure is needed for transferring patient data to the cloud server. These systemic issues limit the observation duration of measurement equipment, such as patch sensors for heart rate capturing. The reduced observation duration will reduce the number of observable AF episodes which has a negative impact on AF diagnosis.  

You are able to build nanoSoC with the DMA removed. The two nanoSoC designs taped out recently had either the DMA 230 or the DMA 350. Aba and the team are developing an accelerator core with micro DMA engines within their accelerator design. This allows them to generate similar firmware for their FPGA implementation as their ASIC implementation. They have been varying the nanoSoC project to build without the  DMA 230 or DMA 350 IP but containing their addressing within the accelerator expansion region. 

Add new comment

To post a comment on this article, please log in to your account. New users can create an account.

Project Creator
rami hariri

researcher student at anglia ruskin university
Research area: AI

Related Articles
2024 IEEE 37th International System-on-Chip Conference

Submitted on

Actions

Log-in to Join the Team