Versions newer than 1.3.2 make Julia/CUDA/Flux slow #742
Description
Hi,
this is a weird issue I suppose. It showed up the first time when the move from 1.3.2 to 1.4.0 was made and I thought it might be a transitive quirk but it seems to have persisted even up to the current release:
Create a new temporary environment and
add CUDA, cuDNN, Flux
(Note that GPUCompiler is then added as dependency at the current release.)
For some reason this makes Julia slow. Even right after doing this, in the precompilation step, julia never takes more than ca. 20% of a cpu and the precompilation step takes a very long time. Once it's done everything with Flux also takes a long time.
If you do instead:
add CUDA, cuDNN, Flux, GPUCompiler@1.3.2
everything seems to behave normally. Precompilation utilizes cores up to 99.9% and my Flux code runs normally. This system is running NixOS and has an Nvidia RTX 3080.
In the past I think I tried repeatedly attaching GDB to see what's happening, but I got none the wiser. This slowdown does not happen on a Windows machine (different CPU, GPU and everything).
julia> versioninfo()
Julia Version 1.11.7
Commit f2b3dbda30a (2025-09-08 12:10 UTC)
Build Info:
Official https://julialang.org/ release
Platform Info:
OS: Linux (x86_64-linux-gnu)
CPU: 16 × AMD Ryzen 7 1700X Eight-Core Processor
WORD_SIZE: 64
LLVM: libLLVM-16.0.6 (ORCJIT, znver1)
Threads: 16 default, 2 interactive, 8 GC (on 16 virtual cores)
Environment:
JULIA_NUM_THREADS = 16