aien-dev/cudapp is a collection of tools that directly generate and execute GPU instructions on the GB10 (sm_121) of the NVIDIA DGX Spark. Aiming to work without the CUDA toolkit, it provides its own machine code toolchain, runtime, and driver client. No version is given for the latest release. Other NVIDIA GPUs are not supported yet.
The core is an implementation of a path that handles GPU instructions instead of CUDA user space components. The README covers machine code reverse engineering, instruction encoding, memory, and command submission. The code includes an instruction encoder, matrix multiplication and attention operation generators, and an execution engine. In other words, rather than presenting itself as a replacement for high-level CUDA code, it is a project that demonstrates a hardware-level approach to creating and submitting instructions on the GB10. It is written in C and licensed under Apache-2.0. The README also includes documents covering the implementation and the license transition.
However, this path depends on the NVIDIA kernel module and signed GSP firmware. It is therefore not a standalone way of driving the GPU with the CUDA toolkit removed, and its support is limited to the GB10. The performance measurements in the README are the project's own, so they cannot be widened into independently verified results. It is a technical attempt to examine a hardware-level approach for local inference, but there is not enough ground to treat it as a replacement tool for general-purpose NVIDIA GPUs.