Nerv Linux
A Unix built for one machine: the unified-memory GPU workstation.
Every package in the base system is compiled from source against musl and LLVM. The NVIDIA driver, CUDA runtime and toolkit are part of the image and work on first boot.
Your GPU is a system call.
One daemon owns the device. Every process on the machine is a client,
and the boundary is mandatory rather than advisory — there is no
vendor library to dlopen around it.
CUDA runs on a GPU this machine doesn’t have.
Because local and remote are the same path,
dlopen("libcuda.so.1") resolves to axon’s shim and the
work executes on another box with the program unchanged.
The inference engine ships with the OS.
Not a container and not a Python stack. S2 is in the image, compiled ahead of time against the silicon it runs on, and budgeted in one unit: bytes read per accepted token.
No glibc in the base. One compiler for all of it.
Every package Nerv builds is musl, compiled with LLVM and ThinLTO — including rpm and dnf5, which no other musl distribution ships.
What it is
Linux runs on everything, so no general-purpose distribution can assume anything. Options accumulate, old behaviour is preserved indefinitely, and portability wins every argument it has with economy. For a system that has to boot on hardware nobody has built yet, that is the correct trade.
Nerv targets one line of silicon. It can assume the accelerator is an NVIDIA GB-series part, that the CPU and GPU share memory, and that every binary on disk came out of the same compiler. Those assumptions pay for the rest of the system: a daemon that owns the GPU and can serve it over a network, one toolchain applied uniformly to every package, and a record of where each file on the machine came from.
Sections
Design
What an operating system is for, what Nerv takes from Plan 9, FreeBSD and the rest, and the choices that follow.
System
musl and mimalloc, the LLVM toolchain, and the RPM stack built and running with no glibc in its dependency tree.
Tools
The software Nerv ships that exists nowhere else: axon, atfield, cage, S2 and the build toolchain behind them.
Userland
Rust uutils in place of GNU coreutils, and the modern CLI wired in as defaults without breaking scripts.
Platforms
NVIDIA GB10, GB20 and GB300 — what each target is and what unified memory changes.
Install
Building from source, booting a QEMU image, and installing to a second slot without touching the first.
Hardware
Nerv officially targets NVIDIA GB-series unified-memory parts and nothing else. Knowing the machine is what the rest of the design is paid for with.
| Target | Arch | Machines |
|---|---|---|
GB10 | aarch64 | NVIDIA DGX Spark and GB10 workstations |
GB20 | aarch64 | GB20-class laptops, as they ship |
GB300 | aarch64 | GB300-class systems, as they ship |