Meta
Hyperscale AI infrastructure
At hyperscale, a few percent of kernel throughput adds up to thousands of GPUs. That is the problem our optimization work exists to solve.
Company
A GPU systems team, building a model platform.
MarsCompute comes out of long-running research on GPU acceleration. We build the optimization layer between models and silicon, and a model API on top of it.
Where this comes from
GPU systems research since 2007.
MarsCompute is led by Prof. Bingsheng He of the National University of Singapore, with a team working on kernels, compilers and AI systems.
-
2007
Mars begins
A research project on GPU acceleration led by Prof. He, started the year CUDA was born.
-
March 2026
First kernel reports
Automated optimization results published for FlashAttention-2, 3 and 4.
-
September 2026
Xtrace released
The intra-kernel tracer in G-Watch goes open source, with a paper describing it.
-
Today
The model API
One OpenAI-compatible endpoint for open and commercial models, from the same team.
Who we work with
Working with teams that run AI at scale.
Our customers include Meta and 4Paradigm, among others, and we collaborate with AI Singapore.
4Paradigm
Enterprise AI platforms
Enterprise AI has to perform on the hardware a business already owns. Getting that across mixed accelerators is a planning problem, not a hand-tuning one.
AI Singapore
Models for Southeast Asia
AI Singapore develops the SEA-LION family for the region's languages. Qwen-SEA-LION v4.5 is live on our API today.
Work with us
Three ways to start.
We work with teams that run production AI systems, where performance on the hardware is critical.
Technical architecture review
A technical look at your workloads and the hardware they run on.
Partnership
Collaboration with hardware, platform and ecosystem teams.
Design partner program
Work with us directly as a design partner.
Who we build for
Teams for whom performance is infrastructure.
- Hyperscalers operating AI fleets across more than one architecture.
- AI infrastructure and platform teams inside enterprises.
- GPU and accelerator vendors scaling production adoption.
- Model serving teams focused on latency and efficiency.
- Developers who want models behind one well-run API.
A model to call, or a fleet to tune?
Request access to the API, or talk to the team about your own kernels and hardware.