94–98%
We trace the kernel you ship
Xtrace leaves 94 to 98% of a kernel's instructions intact. Existing tracers keep 8 to 48%.
Model-as-a-Service, engineered from the silicon up
The model API from the team that tunes the silicon.
We build the tools that see inside GPU kernels and the system that tunes them, then put models behind one OpenAI-compatible API.
- Kernel-level optimization
- NVIDIA and AMD
- OpenAI-compatible API
Producer
MMA
Softmax A
Softmax B
Epilogue
Tiles finished in this window
3
Matrix-multiply warp stalled
48%
A schematic, not a measurement. Real traces of FlashAttention, cuDNN and more are in G-Watch Open Traces.
- NVIDIA Volta to Blackwell
- AMD CDNA1 to CDNA4
- AMD RDNA1 to RDNA4
- CUDA
- HIP
- CuTe DSL
- Triton
- TileLang
- FlyDSL
- cuDNN
- cuBLAS
- TensorRT-LLM
What we do
Most model platforms resell GPU hours. We engineer what runs on them.
01
Model API
Open and commercial models behind one OpenAI-compatible endpoint, with keys, limits and metering built in.
Explore the platform 02Optimization engine
The Mars Optimization Brain plans and applies kernel-level optimizations across NVIDIA and AMD hardware.
See the technology 03G-Watch
Our open-source toolbox that lets engineers and coding agents see inside a GPU kernel.
Meet G-WatchWhy MarsCompute
Strengths you can measure.
3.9×
Agents do the tuning
A coding agent needed 3.9× fewer iterations to tune a kernel with our trace.
19
NVIDIA and AMD alike
19 GPU architectures covered, from Volta to Blackwell and CDNA to RDNA.
+50.8%
Gains you can check
Our best published gain so far: FlashAttention-2 on an NVIDIA A100.
Model library
Four models. One integration.
Two are live today. Two more are coming on GPUs we operate.
Who we work with
Working with teams that run AI at scale.
Start building on MarsCompute.
Request access and we will set up your workspace. Bring the client you already use.