Technology
Infrastructure that improves itself.
We build AI agents that make the GPU code under every AI model run faster, and that get better at it with every change they keep.
01The bottleneck
AI is only as fast as its GPU kernels.
Kernels are the small programs that do a model's math on the GPU. Every new model and every new GPU needs them tuned again, and today that work is slow, scarce and done by hand.
-
Tools see only totals.
Profilers report figures for a whole kernel, not when each warp waited or why, so engineers and their agents guess.
Our answer: MarsTrace traces every warp
-
Experts do not scale.
A few specialists tune one kernel at a time, while new models and GPUs keep arriving.
Our answer: Agents work in parallel
-
Lessons get lost.
What one fix teaches is rarely written down, so the next kernel starts from zero.
Our answer: A memory every agent reads
02Our answer
Agents that tune kernels, and learn from every change.
We call it recursive self-improvement, one layer down from the model. Instead of a model training a model, agents improve the infrastructure every model runs on, in three loops.
- Turn
- One kernel, made faster and checked.
- Round
- A model's slowest kernels, taken by agents in parallel.
- Memory
- Every kept change, read by the next agent on any model.
Inside one turn
-
Trace
MarsTrace records a timeline of every warp in the binary you ship, closed-source libraries such as cuDNN included.
94–98% of instructions untouched
-
Optimize
Agents name what limits the kernel and try the cheapest fix first: tuning, lower precision, fusion, a rewrite, a patch to the binary.
Seven kinds of change, cheapest first
-
Verify
A change stays only if it is faster beyond run-to-run noise and the outputs still match.
Faster and still correct, or dropped
-
Remember
Each kept change is stored with the kernel's hardware footprint. The next agent, on any model, starts from what worked.
Shared across kernels and models
03The proof
Pika H3 Flex: the same model, made faster.
We ran the loop on Pika's reference-to-video model when we brought it to our GPUs: agents went through its engine kernel by kernel. Same weights, same six denoising steps.
- Official build the starting point
- +15 kernel optimizations 1.20–1.58× faster generation
- +F16 softmax and FP8 attention 1.56–1.83× faster generation
End to end, against Pika's own API
Four 10-second 768p jobs, from submitting to the finished MP4, raced again at ten times speed.
14.2 s video + 2 images 1.70×
10.5 s video + 2 images 1.56×
14.1 s video + 1 image 1.43×
4 images 1.24×
Mean of two runs per task, 13–14 November 2026.
The same method for any AI workload, across GPU generations
-
NVIDIA Blackwell
up to 1.41×
Found by an agent comparing traces with cuDNN's closed-source kernel, our tile ordering is now in the official FlashAttention, measured by its maintainers on B200, B300 and GB300.
-
NVIDIA Hopper
3.9× sooner
Tuning FlashAttention-3 on an H100, an agent reading our trace reached its target in 14 iterations, against 54 with the best existing tracer.
MarsRover, September 2026
-
NVIDIA Ampere
up to 1.51×
FlashAttention-2's forward pass on an A100, at sequence lengths of 1,024 and up, in an automated run.
Report, March 2026
Trusted at hyperscale. We built a customized MarsRover for Meta, to help its engineers optimize internal AI workloads.
Read the story04Put it to work
Make your models faster.
Use the same technology on your own GPUs, on ours, or through a model we serve.