Getting more from the GPUs behind AI
An introduction to Tessella and the challenge of building faster, reliable GPU programs with SOL-ExecBench.
The benchmark
Every AI model relies on small programs called kernels to carry out its calculations on a GPU. How these programs organize their work can make a substantial difference to speed, even when they compute the same result. NVIDIA’s SOL-ExecBench puts that challenge to the test: make these programs faster while keeping their answers correct. It compares performance with estimates of what the hardware can theoretically deliver, across operations and input sizes used in AI.
Tessella
We are building Tessella, a programming language and compiler for GPU computing. It provides a way to describe a calculation and control how it runs on the chip. The compiler turns that description into executable GPU code. Tessella brings performance controls together with mathematical models and validation tools, making it possible to reason about changes and check that optimizations preserve the intended computation.
How we work
We use SOL-ExecBench as a practical testing ground. AI agents explore different implementations, check their results, and measure their speed. Benchmark feedback helps identify which ideas deliver repeatable improvements. This connects the development of Tessella to concrete computing problems and measurable outcomes.
Evolving the language and kernels together
Our goal is an agentic system in which AI agents improve kernels and the language used to write them together. Kernel experiments reveal where new language features or compiler improvements could help. Agents can propose those changes, test the implementations they enable, and use correctness checks and performance measurements to evaluate the results. Useful improvements to Tessella then become available to other kernels, allowing discoveries from individual experiments to benefit a broader range of tasks.