At NVIDIA GTC 2025 we spent the week showing Artemis to developers who arrived with the same three problems: legacy code nobody wants to touch, AI-generated code carrying hidden technical debt, and a codebase that simply needs to go faster on the hardware already paid for.
The demo
The one that drew a crowd: Artemis applied to Sakana.ai's CUDA code, cutting runtime from 145 seconds to 14 — a 10x speedup, run live at the booth.
Why CUDA is a hard target
GPU code is where hand-tuning hurts most. The search space is large, the interactions between memory layout, occupancy, and kernel structure are non-obvious, and the only honest way to compare two candidates is to run both on the target hardware. That is exactly the loop Artemis automates: generate, build, benchmark, keep what measurably wins.
What it means for your codebase
A 10x on a demo is a headline. The transferable part is the method. Nothing about the approach is CUDA-specific — it works wherever performance is measurable and correctness is checkable.
