The longer it searches, the faster your code runs.
The doubleAI Drop-in Framework speeds up your code without a rewrite. Point it at a GPU workload, give it compute, and it searches for a kernel faster than the expert-written baseline. Then it proves the result is still correct. Give it more time, and it predictably finds more speed.
The models you already run. Our code already outruns the kernels inside them.
The same result, where you feel it.
Solo & small teams
You pay for every GPU hour and cannot keep a full-time performance engineer. You write a kernel once and hope it holds.
The framework writes the fast one and proves it is correct, so your compute budget goes further.
Platform teams
Every new model means new kernels to hand-tune. Latency and cost climb, and a subtle bug can hide in production for weeks.
Point it at each new workload. Faster kernels, verified, without pulling engineers off the roadmap.
Enterprise
At fleet scale, a few percent of kernel speed is a very large number. Your perf team is strong and still cannot cover every workload.
The framework covers the long tail and closes the gap toward the hardware limit, with a correctness result you can audit.
Being early compounds.
We set the record on day one and beat it on day two by going further, not by starting over. Early access runs on the same logic: the sooner the framework is on your workloads, the longer your advantage has to compound. The first cohort is small, and it moves first.
Take a seat ↓We gave it a second day. It beat our own record.
Same system, same benchmark. We just let it search longer. For context, the previous top result on this benchmark ran for three weeks and beat the baseline on 63% of problems, at 1.38× average.
How close is that to the limit?
Every workload has a physical speed of light: the fastest the hardware could ever run it. The SOL score measures how far a kernel closes the gap from the optimized baseline toward that limit. On many of these kernels the limit is not even reachable. The framework gets close.
Peak, not average: the single best kernel, an NVFP4 attention kernel, runs 14.9× faster than the reference, essentially at the speed of light. We lead with the average because the average is the honest number.
Fast is not the hard part.
A fast kernel that is quietly wrong is worse than a slow one that is right. We built the framework to verify correctness, not just chase speed. On one benchmark kernel a faster competitor passed every test the benchmark defined and still corrupted training: the loss quietly diverged and never recovered. Ours stayed correct across the same runs. That gap is the point.
See the speedup on your own code.
We are opening the Drop-in Framework to a small first group. Pick who you are, leave an email, claim a seat. Solo builders and fleet-scale teams are equally welcome: the framework has to prove itself at every size. You will get exactly one email, when it is ready. No spam, no drip, no list-selling.
You want faster inference without hiring a perf engineer. We will start you with the workloads where the budget saving is largest.
You want verified kernels for every new model without stealing time from the roadmap. Tell us your stack and we will prioritise your architectures.
You want fleet-scale savings with an auditable correctness result. We will set up a scoped evaluation on your own workloads.
doubleAI · Drop-in Framework early access · limited first cohort