The problem
We're fixing an important problem
Models keep getting better, but the software underneath them hasn't kept up. Few stacks come close to the peak performance each chip is capable of, and much of the world's compute sits underused as a result.
Training and inference live in separate stacks, so models get rewritten on their way to production. Each new accelerator brings another round of hand-tuned kernels, and teams maintain several versions of the same model just to run it well everywhere.
Our mission
We're on a mission
Compute is the raw material of intelligence, and today too much of it goes to waste. Our mission is to make every chip count.
We start from how each chip moves data and computes, and build upward: CubeCL to write kernels for any GPU or CPU, Burn to train and run models, and Metabolic to deploy them. One unified stack, from the kernel to the deployed model.
The future
We're building the future
A future where every accelerator in the data center runs close to its peak, whatever its vendor or generation, and where operators turn that capacity into intelligence, on their own terms.
Training and inference will share the same fleet and the same code, so capacity follows demand instead of sitting idle. Models won't ship frozen: they'll keep learning on the infrastructure that serves them, close to their data. We're building the software that makes this possible.
Open source
Built on open source
We believe contributing to open source is the best way to create new technologies. It accelerates adoption, fosters collaboration, and enables rapid iteration, outpacing any closed-source solution.
Through open development, we create a foundation that benefits the entire AI community and ensures transparency in our technological advances.
The team
Meet the people behind Tracel
Founders
Nathaniel SimardChief Executive Officer
Started Burn, CubeCL and Metabolic, and leads Tracel's direction across the whole stack. Master's in machine learning, Mila.
Louis Fortier-DuboisChief Computing Officer
GPU kernel specialist and main author of CubeK and its tile-based approach. Master's in complexity theory and three years of PhD in ML, Université Laval.
Machine learning team
Guillaume LagrangeMachine learning
Core maintainer of Burn and its community manager. Master's in machine learning, Mila; five years of computer vision in production.
Charles RenaudMachine learning
Brings models to Tracel's stack and optimizes them, from training to inference. Master's in computer vision, Université Laval.
High-performance computing team
Genna WingertCompiler
Works on CubeCL's compiler, shaping how kernels are written in Rust and how they compile for every platform.
Marc-Antoine ManninghamCompiler
Built CubeCL's CPU backend and works on its compiler infrastructure. Master's on FPGA compilation for CubeCL, Polytechnique Montréal.
Samuel BélangerKernels
Makes modern architectures like mixture-of-experts run fast on Tracel's stack. Software engineering, Université Laval.
Thierry Cantin-DemersRuntime
Works on the Burn and CubeCL runtimes and on distributed execution for training and inference. Software engineering, Université Laval.
Application team
Sylvain BennerInfrastructure
Builds Tracel's infrastructure. More than two decades of experience, including Shopify's machine-learning platform; creator of Spacemacs.
Jonathan RichardFull-stack
Builds Tracel Console, with a focus on experiment tracking and model management. Computer science, Université Laval.
Marc-Anthony GirardFull-stack
Builds the foundations of Tracel Console and the app side of Metabolic. Software engineering, Université Laval.
Our partners
Thank you for building with us.
