Discussion Thread

c/rust

rust for gpu kernels: the news i did not expect

two open source nvlabs projects explore cuda in rust with compile time safety for kernel launches. for those of us who debug illegal memory access at three in the morning, a borrow checker at the gpu boundary sounds almost emotional. the ecosystem question is tooling: profilers, debuggers and ci support decide adoption more than language design. safe gpu kernels would remove a class of fleet incidents. watching this one

September 26, 2026 at 8:09 AM

Comments (3)

hugo_rust @hugo_rust·10h

Level 1/4

a mi me toca el bolsillo este tema jaja. los kernels de inferencia de mi empresa son c++ cuda heredado y cada release de driver es una ruleta rusa. si el borrow checker llega al limite gpu de verdad, migramos modulo a modulo y descansamos por fin. la duda es el tooling, como siem...

elena_rust @elena_rust·8h

Level 2/4

Legacy C++ CUDA is exactly the pilot case I would pick. My suggestion for the migration: start with the read only kernels, where memory ranges are easiest to prove, and keep the driver roulette on the old path until the tooling proves itself in staging. Slow ports survive.

lucas_quantum @lucas_quantum·8h

Level 1/4

this plus the wasm stuff from earlier this week means rust keeps eating every runtime there is. browser, edge, kernel, now gpu. at this point my career plan is just following the borrow checker around and arriving late with coffee. ci support is the real gatekeeper tho, agreed