Shikumi is a GPU runtime built around persistent processes. A process stays resident, keeps its own state, and runs at its own rate. Processes talk through typed channels: one writes, another reads, and both keep going. The interesting machine is a population of those loops, sharing the device for as long as they have work. That is a different bet from a queue of kernels that each start, finish, and hand the GPU back.

Flapping Airplanes made the public case for this kind of machine this week. Their post describes a megakernel framework that can train, serve inference, and play a game, including all three at once in a single kernel. The system in that recording is their Kittens Virtual Machine, KVM. We agree with the direction. A GPU wants programs that stay alive together. KVM approaches that as a virtual machine with its own instruction set. Shikumi approaches it as processes and channels, each process on its own clock. Same destination, another angle. Two groups arriving there is support for the idea.

Their demo is replicated on Shikumi: training, inference, and Doom in one kernel launch, on an RTX PRO 6000. The model is about 133 thousand parameters, so the decoded text is word salad. The point of the recording is the launch. One kernel, three programs, still running. A larger system is the same shape.

Training, inference, and Doom. One kernel launch. RTX PRO 6000.