Grigori Fursin Email & Phone Number

Paris, Fr

Paris, Île-De-France, Fr

https://flex.ai

San Francisco, Us

* created the Collective Mind automation technology (CM, CM4MLOps, CM4MLPerf, CM4ABTF) and donated it to MLCommons to help everyone modularize, benchmark and optimize AI/ML systems across diverse and continuously changing models, datasets, software and hardware from different vendors:* prepared white paper with my vision for CM: https://arxiv.org/abs/2406.16791* gave invited keynote about CM at ACM REP'23: https://doi.org/10.5281/zenodo.8105339

San Francisco, Us

I helped MLCommons bootstrap the development of the CM automation recipes for MLOps, MLPerf and ABTF as a collaborative engineering project to run MLPerf inference benchmarks across diverse models, data sets, software and hardware from different vendors in a unified way: * https://access.cknowledge.org/playground/?action=scripts .* https://github.com/mlcommons/cm4mlops* https://github.com/mlcommons/cm4abtfCM was successfully validated by automating ~90% of all MLPerf inference v4.0 performance and power submissions while finding some top performance and cost-effective software/hardware configurations for AI systems: see our report for more details.

San Francisco, Us

* established the MLCommons Task Force on Automation and Reproducibility after I donated my CK and CM automation technology to MLCommons in 2022 to benefit everyone: https://github.com/mlcommons/ck/blob/master/docs/taskforce.md . * continued developing CM as a community effort to modularize, automate and unify benchmarking and co-design of efficient and cost-effective AI/ML systems: https://github.com/mlcommons/ck * has successfully accomplished the mission in May 2024 - our on-going developments are funded by MLCommons and integrated with several MLCommons Working Groups.

I have established the cTuning foundation to connect academia and industry to automate the development of more efficient and cost-effective computer systems, and helping the community reproduce the state-of-the-art research projects and validate them in the real world: * https://royalsocietypublishing.org/doi/10.1098/rsta.2020.0211* https://learning.acm.org/techtalks/reproducibility* https://www.youtube.com/watch?v=7zpeIVwICa4 .Our goal is to develop open-source tools to help researchers and engineers automate their tedious and repetitive tasks, improve productivity, unleash creativity, accelerate innovation, reduce all R&D costs and make AI accessible to everyone.That is why we have developed the open-source Collective Knowledge technology v1 to empower everyone to quickly validate their ideas in an automated and reproducible way across diverse and rapidly evolving AI/ML models, data, software and hardware from the cloud to mobile and tiny devices.We became a founding member of MLCommons and have donated Collective Knowledge Technology to MLCommons in 2021 to continue open and transparent developments with the new MLCommons taskforce on automation and reproducibility to benefit everyone.We deprecated Collective Knowledge v1 in 2021 and v2 in 2022 to develop a new Collective Mind workflow automation language with MLCommons to unify benchmarking and optimization of AI systems across any AI/ML/SW/HW stack.We continue helping AI, ML and systems conferences set up and automate artifact evaluation and reproducibility initiatives: https://cTuning.org/ae

Seattle, Washington, Us

I have developed the 2nd generation of the open-source Collective Knowledge workflow automation technology (aka Collective Mind with CM4MLOps automation recipes) - it was adopted by @MLCommons to modularize MLPerf inference benchmarks and automate development, optimization and deployment of AI systems across diverse models, data sets, software and hardware:* https://cKnowledge.org* https://github.com/mlcommons/ck* https://github.com/mlcommons/cm4mlops* https://docs.mlcommons.org/inference

News: cKnowledge.io platform was acquired by OctoML.ai and our open-source CK technology was donated to MLCommons to benefit everyone.I designed and implemented a prototype of the Collective Knowledge platform based on my open-source CK technology to automate the tedious and time consuming process of building, testing, benchmarking, comparing, optimizing and deploying emerging ML Systems in the most efficient way in terms of speed, accuracy, energy, size and costs on any tech stack from the cloud to the edge. It helps to automate MLPerf submissions, reduce R&D costs and accelerate the time to market for innovative products by several orders of magnitude: https://cKnowledge.io or https://cKnow.io

Co-founded CodeReef with Nicolas Essayan to prototype a platform that enables benchmarking and optimization of Machine Learning systems.➡️Incepted @StationF as part of Entrepreneur First program.➡️Started developing the 2nd generation of my open-source Collective Knowledge workflow automation framework

London, Gb

Selected for the EF's second cohort in Paris, I learnt how to build deep tech startups and MVPs from scratch while avoiding numerous pitfalls and minimizing all risks. This knowledge and experience helped me to meet many amazing people, start long-term collaboration with Nicolas Essayan, and create the cKnowledge.io platform acquired by OctoML.ai in 2021.

Cambridge,, Gb

- I co-founded a commercial engineering company to validate my Collective Knowledge framework in production and automate the development of efficient, reliable and affordable computing systems from HPC servers to edge devices: https://github.com/ctuning/ck- With the help of my CK framework, our customers automated the tedious benchmarking, optimization and co-design process of emerging AI, ML and quantum workloads while reducing R&D costs and accelerating time to market for DNN-based technology by several orders of magnitude: https://cKnowledge.org/partners .

Seattle, Wa, Us

I successfully prototyped CK workflows to automatically scale of deep learning on AWS using C5 instances with MXNet, TensorFlow and BigDL from the edge to the cloud.

Detroit, Michigan, Us

With my CK technology, GM reduced the time to explore and select Pareto-optimal AI/SW/HW stacks from different vendors for self-driving cars by several orders of magnitude.

Cambridge, England, Gb

I started a new educational initiative with the Raspberry Pi foundation and developed a CK workflow to expose compiler optimization problems to ML researchers while hiding the system complexity: https://cknow.io/c/report/rpi3-crowd-tuning-2017-interactive .

Mountain View, Ca, Us

* developed portable workflows for performance autotuning and software/hardware co-design of some workloads.

Cambridge, Cambs, Gb

- Won a 6-month grant from the EU TETRACOM initiative to develop the open-source Collective Knowledge technology with a permissive license and validate it in production with Arm.- Released all technology at https://cKnowledge.org.- Was honored to receive the European technology transfer award for this successful project.

Le Chesnay Cedex, France, Fr

- Was among the first researchers in the world to demonstrate how to automatically learn optimization heuristics in production compilers, reduce optimization cost by an order of magnitude and get up to 4x speedups and 30% code size reductions on previously unseen workloads in comparison with the best manually tuned heuristics.- Developed a compiler plugin framework and program feature extractor included to the mainline GCC in collaboration with Google and Mozilla to support ML-based testing and optimization in production compilers.- Led the development of the world's first ML-based compiler across 20 researchers and engineers in 5 teams in the €1.5M MILEPOST project with IBM, ARC, U.Edinburgh, CAPS Entreprise and INRIA. This work was commercialized by IBM, Synopsys, Intel and STMicroelectronics.- Co-advised 2 PhD students who successfully graduated from the University of Paris Saclay.- Developed cBench and 2 data sets (MiDataSets and KDataSets) to create more realistic conditions for SysML research used by the leading universities and companies including Arm, Google and Facebook.- Developed the cTuning.org platform considered the first in the world to crowdsource ML-based program autotuning across diverse platforms provided by volunteers similar to SETI@home: https://bit.ly/ctuning.

Le Chesnay Cedex, France, Fr

- Co-authored the successful €1.5M MILEPOST grant proposal with 3 universities and IBM and presented it for the EU commission to build the world’s first self-optimizing compiler based on my PhD research and powered by ML. - Was awarded a tenured research scientist position at INRIA to lead this project.

Hillsboro, Or, Us

- On sabbatical from Inria to co-direct Intel Exascale Lab in France with 20 researchers and engineers reporting directly to the Lab’s CEO.- Proposed, designed and led the development of an internal framework and a web-based platform to automate the software/hardware co-design process for Exascale systems using portable workflows and machine learning.- Was promoted to senior research scientist at INRIA and received the award of scientific excellence based on the results of this sabbatical.

Armonk, New York, Ny, Us

* led the development of a self-optimizing and machine-learning based compiler (MILEPOST GCC) in close collaboration with IBM.* demonstrated that it is possible to use machine learning to accelerate co-design of efficient computer systems and reduce the time to market for innovative products by several orders of magnitude* received the ACM CGO'17 test of time award for this technology which is considered to be the first in the world:- https://www-03.ibm.com/press/us/en/pressrelease/27874.wss- https://cTuning.org/milepost-gcc- https://github.com/ctuning/reproduce-milepost-project* developed the concept of cTuning.org to crowdsource training of the machine learning model inside this compiler across diverse platforms and workloads provided by volunteers:- https://cTuning.org- https://cTuning.org/repo- https://hal.inria.fr/inria-00436029v2

Edinburgh, Gb

- Led 2 work packages out of 4 in the €1M EU MHAOTEU project. This comprised 12 researchers and administrators. - Built an open-source framework and a client-server infrastructure used by 5 universities and EPC (Edinburgh Parallel Compilers) in the €1M EU MHAOTEU project to optimize real world HPC applications. It was commercialized by EPC.- Built an open-source polyhedral compiler and autotuning infrastructure used by the HiPEAC network of excellence.- Our project successfully delivered all deliverables and was highly rated by the EU.