What is the role of CUDA_CACHE_DISABLE=1? Please explain in detail

Question

I saw some codes using CUDA_CACHE_DISABLE=1 in run.sh file.

But there is no thorough explanation on website.

What I get from the online document is something abstact, like below:

Disables caching (when set to 1) or enables caching (when set to 0) for just-in-time-compilation. When disabled, no binary code is added to or retrieved from the cache.

But what is binary cache and retrieval?

Where can be better with this command, better utilization of GPU-memory or something else? Thanks in advance.

Answer 1

It is explained for example here by Nvidia .

In short: CUDA code has to be recompiled for every major new GPU architecture. Binaries of CUDA code typically contain PTX as an intermediate representation for this. The compilation results will be cached.

Disabling cache will affect startup time and disk memory usage but not runtime performance.