Simon Bührer

Simon Jonas Bührer

Engineer working on novel ML Paradigms, co-designed across software and hardware. Instead of compiling a high-level model down onto a chip, the network is learned directly at the lowest level representation the hardware gives you.

I studied electrical engineering and information technology at ETH Zürich, and I build neural networks that are hardware-native. The usual approach designs a model with high-level math and lets a toolchain compile it down to whatever the target chip actually runs. I drop that step and learn the network directly at the lowest level the hardware (FPGA, CPU, GPU, ASIC) offers. Built this way, the model is what the chip prefers, so it needs far less computation for the same accuracy and runs with lower energy, lower latency, and higher throughput. Working at that level, most of the standard deep-learning toolkit no longer holds and has to be rebuilt from scratch.

That rebuild changes how a neural network behaves, and reshapes how we think about it: its architecture, its components, and the way it trains and runs.

Efficient inference, live

A live MNIST demo: a conventional dense neural network, run as an int8 GEMV (matrix-vector multiply), against a lookup-table (LUT) based network of logic gates. Both classify the same handwritten digits, but the LUT network gets there for far fewer boolean operations per sample.

Loading the live demo

Research interests

  • Hardware-native AI for high-throughput, low-latency, low-energy inference, on both existing and newly designed hardware. This moves models out of datacenters and onto edge devices, reaching new users and unlocking new use cases.
  • Moving past the split between training and inference, where a model trains once and is then frozen. I want models that keep learning from new data and new tasks as they run, and that can let go of what no longer matters.
  • Distributed AI: instead of a few state-of-the-art expert models in datacenters, each device runs a local model that learns what matters to its user and exchanges information with other models and the cloud at several levels.

Education

  1. 2024-2026

    ETH Zürich, MSc Electrical Engineering & Information Technology

  2. 2020-2023

    ETH Zürich, BSc Electrical Engineering & Information Technology

  3. 2015-2019

    Kantonsschule Kreuzlingen, Gymnasiale Maturität

Experience

  1. 2025-2026

    Everllence (former MAN Energy Solutions)

    Internship to Werkstudent, Zürich

  2. 2022-2023

    ARIS Project Prometheus

    Fokusprojekt, Dübendorf

  3. 2021-2022

    ETH Zürich

    Teaching Assistant, Engineering Mechanics

  4. 2020

    Bühler Group

    Internship, London, UK

  5. 2019

    Swiss Armed Forces

    Panzerjägersoldat, Aufklärungskompanie, Thun

Languages

German (native), English (professional), French (limited), Russian (elementary)