Jan 8, 2026

Invited Vijay Janapa Reddi Petrut Bogdan jyik@g.harvard.edu steve.furber@manchester.ac.uk Eran Briman Pete Bernard raashid.ansari@silabs.com Ed Doran Rosina Haberl Chetan Singh Thakur EAIF Team Calendar Gideon Intrater klaus.knobloch@infineon.com priya.panda@yale.edu friedemann.zenke@fmi.ch dhireesha.kudithipudi@utsa.edu n.beladel@student.tudelft.nl Venkat Rayudu jtapson@brainchip.com Frank Ghenassia Jason Eshraghian andre.vanschaik@manchester.ac.uk

Attachments EDGE AI Neuromorphic WG EDGE AI Neuromorphic WG kick off meeting – 2025/09/29 08:56 CST – Notes by Gemini 

Summary

Petrut Bogdan, Rosina Haberl, Benedetto Leto, and Jason Yik discussed extending NeuroBench to act as a front-end for hardware benchmarking across different system backends, aiming to automate the process for users to get energy and latency metrics from actual boards, with SSense boards as the first pilot hardware. Jason Yik and Benedetto Leto confirmed intentions to engage with original NeuroBench paper contributors like Intel and SpinCloud, while Jonathan Tapson offered accessible hardware and proposed publishing a reference power measurement design to aid standardization. Petrut Bogdan, Rosina Haberl, and Jonathan Tapson planned a fully online, live-streamed mini-conference day focusing on neuromorphics from an AGI perspective, targeting the week of February 16th, and suggested speakers including Guido de professor and Greg Cohen.

Details

Notes Length: Standard

• Start of Meeting and Well Wishes Petrut Bogdan and Rosina Haberl exchanged New Year greetings, with Rosina mentioning they were ill and could not use their camera. Petrut Bogdan noted that Rosina’s voice sounded rough. The discussion briefly covered the logistics of other participants like Gideon from Vivid, who was likely at CES. Petrut Bogdan also described travel difficulties related to snow and high winds in the Netherlands, including helping a delivery van that was stuck in the snow.
• Introduction of NeuroBench Update Petrut Bogdan introduced the idea of discussing the February conference day and welcomed additional guests, Benedetto Leto and Jason Yik, who were working on NeuroBench. Jason Yik, representing the NeuroBench project, explained that they are interested in extending NeuroBench as a front-end for benchmarking across different system backends. Benedetto Leto, a main developer, added that they aim to develop the harness further to be closer to hardware benchmarking, and introduced Gianvito Urgese as their supervisor.
• Goal of NeuroBench Extension for Hardware Benchmarking Petrut Bogdan sought clarification on the development goal for NeuroBench, asking if it involved native communication with neuromorphic hardware. Jason Yik confirmed the goal is to take trained models and datasets from the user interface and automate the process of pushing them to different hardware backends, adding that this would provide metrics like energy and latency from actual boards, alongside existing metrics like sparsity and synaptic operation count. The NeuroBench extension assumes that the target boards have the necessary instrumentation for measuring latency and power.
• Initial Hardware and Vendor Engagement for NeuroBench Jason Yik and Benedetto Leto identified SSense boards as the first pilot hardware for integration with NeuroBench due to their clean interface. Petrut Bogdan inquired about engaging with original NeuroBench paper contributors, such as Intel and SpinCloud, to avoid duplicating effort and gather requirements upfront. Jason Yik confirmed that they intend to reach out to them, noting the timing aligns with their return from the holidays and the start of this initiative. Jonathan Tapson confirmed that their hardware is accessible for the project and they would be happy to help integrate the KA from the NeuroBench front end.
• Discussion on Power Measurement and Standardization Petrut Bogdan asked if NeuroBench could support off-the-shelf lab equipment for performance analysis due to varying on-board instrumentation. Jonathan Tapson responded that while their evaluation boards have hooks for power measurement, they were unsure about the benefits of integrating beyond that. Jonathan Tapson suggested that system providers could expose an API for NeuroBench to request metrics, especially if the hardware lacks on-board instrumentation. Jason Yik agreed that from the user’s perspective, the process should be a “push button” experience for collecting power stats, and documentation on measurement sources would suffice, rather than requiring standardized external hardware. Jonathan Tapson proposed publishing a reference power measurement design to allow NeuroBench or customers to duplicate their power measurement systems.
• InoTra Development Board Status Jason Yik asked Petrut Bogdan for an update on the availability of an InoTra development board for users. Petrut Bogdan stated that they are working on ramping up production and determining the version of the evaluation kit that will include instrumentation, expecting it to be available within the next couple of months. They noted that Talamu, being PyTorch-based, should allow for relatively easy integration with NeuroBench, especially if the model was trained for an InoTra chip.
• Neuromorphic Conference Day Planning Petrut Bogdan initiated the discussion on organizing a mini-conference day, fully online and live-streamed, focusing on neuromorphics from an AGI perspective, with a target date around the week of February 16th. Petrut Bogdan suggested the conference day be held around the 18th, following a dry run at the start of the week. The goal of the event is to demonstrate that neuromorphics are ready for deployment across small, medium, and large scales, covering existing applications, systems, training methods, and benchmarking infrastructure.
• Conference Content and Speaker Suggestions Petrut Bogdan outlined a draft timetable and topics, including use cases, training methods, and physical AI, suggesting Guido de professor at Delft as a potential speaker for their work on low-latency, low-power control problems. Jonathan Tapson supported the program and was enthusiastic about hearing Guido speak. Petrut Bogdan clarified that the initial conference is invite-based to establish a strong case, which Rosina Haberl agreed with as a good starting point before potentially moving to an abstract call for future recurring events. Rosina Haberl requested a theme or title for the event to assist with promotion and thumbnail creation. Jonathan Tapson suggested inviting Greg Cohen from Western Sydney University to speak about their work with neuromorphic systems on the space station. Other speaker suggestions included Andre from the working group, Miguel Mastela from Neonova, and Massimiliano Versace from Analog Devices.

Suggested next steps

• Jonathan Tapson will publish a reference power measurement design to help customers or NeuroBench duplicate the power measurement systems.
• Petrut Bogdan will look into and continue debating the suggestions for conference speakers offline.
• Rosina Haberl will arrange with the marketing team and Pete to create a good thumbnail and start promoting the mini conference day.
• Petrut Bogdan and Rosina Haberl will finalize the date for the mini conference day and propose a title suggestion offline.