Is there any way to find Physical location through the logical SM IDs?

I’m trying to understand how logical SM IDs (as reported by the PTX %smid register) map to the physical placement of SMs on the die — specifically on Blackwell parts such as B200 and the RTX PRO 6000 Server Edition (GB202), and on Hopper H100.

Taking the block diagram below as a reference: which SM IDs would you expect to land in the top-left GPC?

  • Are SM IDs assigned contiguously per GPC, i.e. SM 0, 1, 2, … , N/8 − 1 all sit in the top-left GPC (N = total SM count)?
  • Or are they interleaved across GPCs, i.e. SM 0, 8, 16, 24, … belong to the top-left GPC?

More generally, is there any documented (or reliably reverse-engineered) relationship between %smid and the GPC/TPC a given SM physically belongs to, or does floorsweeping make this mapping chip-specific and effectively opaque to software?

You can prepare a kernel that communicates all SM unit pairs in isolation, and create a heat-map of latencies to see how close SM units are in terms of cache-coherence which should follow physical closeness somehow.

If there are 148 SM units, it would take (148 x 147) / 2 number of benchmark iterations, this could be completed within minutes. Then tell claude(or nemotron) to create heatmap from the data.