L2 sectors question!

Hello everyone,

I have a question regarding how to determine the number of sectors for the L2 cache.

My assumption was that L2 sectors should correspond to the L1 misses, which I tried to represent using these metrics:

  • lts__t_sectors_srcunit_tex_op_read.sum
  • lts__t_sectors_srcunit_tex_op_write.sum

However, when I compare these numbers, things don’t add up. Here’s what I’m seeing:
lts__t_sectors_srcunit_tex_op_read.sum (sector) 2,073,756
lts__t_sectors_srcunit_tex_op_write.sum (sector) 16,808,976
lts__t_sector_hit_rate.pct (%) 99.37
lts__t_sectors.sum (sector) 27,430,992

The total number of L2 sectors (lts__t_sectors.sum) is much larger than the sum of L1 read and write misses. I also found another metric:
lts__t_sectors_lookup_hit.sum (sector) 27,258,385

This value is also much larger than the sum of L1 read and write misses.

Can anyone explain why these metrics are so different? Why is the total number of sectors and sector hits at L2 much higher than what I would expect from just the L1 misses (reads and writes)? Is there something I’m misunderstanding about how these metrics are calculated or what they represent?

Thank you in advance for any insights!

I would suggest to read/watch Requests, Wavefronts, Sectors Metrics: Understanding and Optimizing Memory-Bound Kernels with Nsight Compute | GTC Digital April 2021 | NVIDIA On-Demand and 2. Profiling Guide — NsightCompute 12.9 documentation to get a thorough understanding about how the L2 cache is represented in the tool.