Customer stories

How teams that run AI at scale work with us.

We work with teams for whom performance on the hardware is not a detail. These are their stories, told as far as each of them has agreed to.

Customer story

Meta

A customized G-Watch for internal AI workloads

Meta runs AI at a scale where the efficiency of a single GPU kernel is a question of fleet capacity. Under a commercial collaboration, MarsCompute provided Meta with a customized version of G-Watch to help its engineers optimize internal AI workloads.

G-Watch gives an engineer, or a coding agent, a timeline of what happens inside a kernel on the binary that really ships, so the time spent waiting can be found and removed rather than guessed at.

Explore G-Watch
Customer
Meta
Product
G-Watch, customized for the engagement
Goal
Optimize internal AI workloads
Engagement
Commercial collaboration

The problem at hyperscale

A few percent is thousands of GPUs.

The kernels behind large models are already heavily tuned. What is left to gain is hidden inside them, and totals do not show it.

Totals hide the cause

Hardware counters say that a kernel is slow. They do not say which stage of it is waiting on which.

A trace shows it

Xtrace records a timeline of every warp, on the compiled binary, and adds 0.9 to 2.8% to the kernel's time.

Agents act on it

With an exact trace, a coding agent tunes a kernel in far fewer iterations than with counters alone.

The numbers on this page are from our published results, not from the work with Meta, whose details are confidential. See the published results.

More teams

Who else we work with.

Collaborator

AI Singapore

AI Singapore develops the SEA-LION family for Southeast Asia's languages. Qwen-SEA-LION v4.5 is live on Token Factory today.

See the model
Customer

4Paradigm

Enterprise AI has to perform on the hardware a business already owns. The full story is in preparation.

Story coming
Your team

Become a design partner

We take on a small number of teams as design partners and work on their kernels, hardware and workloads directly.

Work with us

Want numbers like these on your own kernels?

Talk to the optimization team about your workloads and the hardware they run on.