Source
- URL: https://x.com/_s_u_r_a_/status/2091080756806521239
- Author: asdb (@s_u_r_a)
- Posted: 2026-08-22 08:31:17
Branch
1/ @s_u_r_a
@AgileJebrim
Why?
2/ @AgileJebrim
Because they’re designed to be data parallel and work with GPUs from the start.
Because they don’t require integrating with an existing bloated library.
SPIR-V is just an incredibly simple and straightforward binary format. We write to it directly as an array of uint32s. You can even read the bytecode in a hex editor directly easily enough if you wanted. Think of it like Java bytecode; it’s portable across a wide range of hardware and vendors.
PTX unfortunately requires generating ASCII output to work with it, so a little more complicated but still nothing like linking to LLVM. It is good for accessing more low level control over NVIDIA GPUs.
We’re also creating our own IR that I suppose one could hypothetically target via code too. We’re doing it internally anyways until our visual editor is ready.