# Non-aggregated profiling result support

**URL:** <https://forums.developer.nvidia.com/t/non-aggregated-profiling-result-support/266910>\
**Category:** CUPTI – CUDA Profiler Tools Interface\
**Tags:** cuda, kernel\
**Created:** [September 19, 2023, 3:10pm UTC](https://forums.developer.nvidia.com/t/non-aggregated-profiling-result-support/266910 "2023-09-19T15:10:47Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![wlgus](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@wlgus](https://forums.developer.nvidia.com/u/wlgus)\
**Post date:** [September 19, 2023, 3:10pm UTC](https://forums.developer.nvidia.com/t/non-aggregated-profiling-result-support/266910/1 "2023-09-19T15:10:47Z")

</div>

The nvprof is not supported starts from Ampere GPU.

“aggregated-mode=off” used to be supported in nvprof, and I am looking for a nwe profiling tool that shows the non-aggregated result for the L2 cache or Memory metric/event.

For example, the result look like this:

```auto
 l2_subp0_write_sector_misses (32 instances) - - [0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 0 0 4 0 0 0 0 0 0 0 0 0 0 0 0 0 0] -
 l2_subp1_write_sector_misses (32 instances) - - [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0] -
l2_subp0_read_sector_misses (32 instances) - - [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 5 0 0 0 0 0 0 0 0 0 0 0 0 0 0] -
 l2_subp1_read_sector_misses (32 instances) - - [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0] -

```

Does CUPTI have similar functionality?

---

<div class="post-metadata">

**Author:** ![mjain](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/mjain/32/14043_2.png) [@mjain](https://forums.developer.nvidia.com/u/mjain)\
**Post date:** [November 9, 2023, 2:58pm UTC](https://forums.developer.nvidia.com/t/non-aggregated-profiling-result-support/266910/2 "2023-11-09T14:58:08Z")

</div>

nvprof events and metrics profiling is based on the event and metric APIs from CUPTI and these are supported on Volta and older GPU architectures only. These APIs are replaced by CUPTI [Profiling API](https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdocs.nvidia.com%2Fcupti%2Fr_main.html%23r_profiler&data=05%7C01%7Cmjain%40nvidia.com%7Ca0db3a9176a44a7eed1a08dbe1287310%7C43083d15727340c1b7db39efd9ccc17a%7C0%7C0%7C638351336088832449%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=exObpvlCKxOuHIffrKn99TmInXOY61CrZxwI7de6leA%3D&reserved=0) which is supported for Volta and higher GPU architectures, and this API does not support per instance metrics. For nvprof metrics listed by you, the equivalent metrics from Profiling API are _lts\_\_t\_sectors\_op\_read\_lookup\_miss.sum_ and _lts\_\_t\_sectors\_op\_write\_lookup\_miss.sum_ and these provide aggregate values across all L2 instances.  
Non-aggregated metric values for L2 cache or Memory are not supported on Ampere GPU by any tool (CUPTI or Nsight Compute).

---

<div class="post-metadata">

**Author:** ![veraj](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/veraj/32/256893_2.png) [@veraj](https://forums.developer.nvidia.com/u/veraj)\
**Post date:** [January 12, 2024, 9:38am UTC](https://forums.developer.nvidia.com/t/non-aggregated-profiling-result-support/266910/3 "2024-01-12T09:38:57Z")

</div>

Closed on January 12, 2024, 9:38am UTC