Julian Büchel

LinkedIn | Google Scholar

Bio

Julian Büchel is a Pre-Doctoral Researcher in the Analog In-Memory Computing group at IBM Research – Zurich and a PhD student at the Department of Computer Science at ETH Zürich under the supervision of Prof. Martin Vechev. He studied Computer Science and Neural Systems and Computation at ETH Zürich from 2016 to 2022. During his Master's, he worked as a Research Engineer at SynSense, a company focused on digital neuromorphic vision. During his Master's thesis conducted at IBM Research in 2021, he demonstrated that adversarial attacks in the weight domain can be used to enhance the robustness of DNNs (see ICLR paper). After his thesis, he joined IBM Research as a Pre-Doctoral Researcher in 2022. Work co-authored by him has won the 2023 OSS award and an honorable mention in the 2024 Pat Goldberg Memorial Award. His current research interests evolve around the topic of bringing LLMs to Analog In-Memory Computing architectures based on high-density non-volatile memory technology. He primarily focuses on re-training LLMs to improve the robustness against various types of noise such as quantization noise and noise introduced by the analog hardware.

News

• [23 May 2025] v1.0.1 of AIHWKIT-Lightning is released. Check it out here.
• [15 May 2025] We released a new paper called "Analog Foundation Models" where we show that LLMs can be trained to tolerate weight noise introduced by AIMC hardware. We loose accuracy comparable to the 4W8I scenario in quantized LLMs.
• [01 May 2025] Iason Chalas successfully defended his MSc thesis. Glad to announce he will join my team focusing on algorithms for analog AI.
• [08 January 2025] The paper "Efficient scaling of large language models with mixture of experts and 3D analog in-memory computing" got the cover at Nature Computational Science.

Software

• AIHWKIT-Lightning is a scalable and light-weight library for HW-aware training.
• 3D-CiM is a simulator that can be used to assess the system level performance of a high-level 3D AIMC-based accelerator.
• Analog-MoE is a triton kernel for a MoE layer equipped with HW-aware training capabilities such as input/output quantization, tiling, and noise injection.
• Sigma-MoE is a library extending Robert Csordas original Sigma-MoE implementation by making it compatible with torch.compile and huggingface.
• analog-foundation-models is code supporting our "Analog Foundation Models" paper and can be used as a recipe to train your own analog foundation model.

Selected publications

Analog In-Memory Computing:

• "Analog Foundation Models", Büchel et al., under review
• "The inherent adversarial robustness of analog in-memory computing", Lammie et al., Nature Communications 2025
• "AIHWKIT-Lightning: A Scalable HW-Aware Training Toolkit for Analog In-Memory Computing", Büchel et al., NeurIPS 2024 (MLNCP workshop)
• "Kernel Approximation using Analog In-Memory Computing", Büchel et al., Nature Machine Intelligence 2024
• "Efficient Scaling of Large Language Models with Mixture of Experts and 3D Analog In-Memory Computing", Büchel et al., Nature Computational Science 2024 (cover)
• "A 64-core mixed-signal in-memory compute chip based on phase-change memory for deep neural network inference", Le Gallo et al., Nature Electronics 2023 (cover)
• "ML-HW Co-Design of Noise-Robust TinyML Models and Always-On Analog Compute-in-Memory Edge Accelerator", Zhou et al., IEEE Micro 2022
• "Exploiting the State Dependency of Conductance Variations in Memristive Devices for Accurate In-Memory Computing", Vasilopoulos et al., IEEE Transactions on Electron Devices, 2023
• "Gradient Descent-Based Programming of Analog In-Memory Computing Cores", Büchel et al., IEDM Proceedings, 2022
• "Programming Weights to Analog In-Memory Computing Cores by Direct Minimization of the Matrix-Vector Multiplication Error", Büchel et al., IEEE JETCAS (highlighted paper), 2023

Spiking Neural Networks:

• "Supervised Training of Spiking Neural Networks for Robust Deployment on Mixed-Signal Neuromorphic Processors", Büchel et al., Nature Scientific Reports, 2021
• "Adversarial Attacks on Spiking Convolutional Neural Networks for Event-Based Vision", Büchel et al., Frontiers in Neuroscience, 2022
• "Network Insensitivity to Parameter Noise via Adversarial Regularization", Büchel et al., ICLR, 2022

Selected Patents

• US20250005431A1
• US20240419971A1
• US20240420762A1
• US20240160348A1
• US20240127009A1
• US20240086682A1
• US20230305893A1

Reviewing

• Nature Communications
• Nature Computational Science
• JETCAS
• ISCAS
• DAC
• DATE
• ICCAD

Get in touch

jub[at]zurich[dot]ibm[dot]com