Constant cache

Hi,
We’re analyzing the memory hierarchy in the NVIDIA Hopper (H100) architecture and specifically looking into the placement and structure of the constant memory cache.

We understand that constant memory is stored in global memory but accessed through a small, fast on-chip constant cache (typically 8 KB per SM in earlier architectures like Volta/Ampere).

We are trying to determine:

  1. In H100, is this constant cache still a separate, dedicated on-chip unit, or is it now part of a L1/shared memory space?
  2. Is this cache physically located per-SM, or has its placement and behavior changed in Hopper?
  3. Is there any official documentation or whitepaper that details the location, size, and access latency of the constant cache in H100?

It is still 8kB/SM as documented here, (Cache working set per SM for constant memory).

Relevant profiling metrics are outlined here, see dcc, gcc, icc, idc and imc.

More unofficial details in Dissecting … Architecture papers, possibly not Hopper yet. Past architectures had an in-between level, e.g. for SM vs. for the SM Partition.