Does constant cache exist in A100 GPUs?

[asterix_obelix]

I was reading the A100 white paper (https://images.nvidia.com/aem-dam/en-zz/Solutions/data-center/nvidia-ampere-architecture-whitepaper.pdf), and there is only one mention of constant memory. Specifically, “Constant memory space resides in device memory and is cached in the constant cache”. My questions -

1. Does constant cache still exist? If so, how big is it? I don’t see any information about constant cache in any of my Nsight Compute profiles.
2. If constant cache does not exist, what is the significance/use case of constant memory?

[rs277]

@asterix_obelix

Does constant cache still exist? If so, how big is it?

Yes. According to the Programming guide, under, “Cache working set per SM for constant memory”, it’s 8kB on CC8.0.

[Curefab]

The main use cache is to free the other data channels (to keep their full bandwidth) for coefficients and other data, which are small overall or for an extended time a small subset is used (to fit into the constant cache) and where the accessed data item is the same for all lanes (threads of a warp).