Feature request: expose globally identifiable per-SM/SMSP performance samples

[Loong0x00]

Summary

I could not find a supported NVIDIA interface that exposes time-resolved performance-monitor samples for each globally identifiable physical SM/SMSP instance (for example, <physical SM ID, SMSP ID, timestamp, activity>).

The existing tools provide valuable but different views:

• nvidia-smi reports aggregate GPU utilization.
• Nsight Compute sm__* / smsp__* metrics generally expose rollups such as sum/average/min/max.
• smsp__warpidsamp_* exposes documented <SMSP ID>:<Warp ID> instance keys, but no physical SM coordinate. On my test system, output such as 0:0, 0:4, 0:8, and 0:12 was useful for scheduler research but could not identify 680 physical SMSPs across the chip.
• CUPTI/PM sampling provides a timeline, but I could not find a supported way to retain a globally identifiable physical SM/SMSP coordinate for every sample.
• PTX exposes %smid and %warpid, but no %smspid; in-kernel instrumentation is intrusive and is not equivalent to reading performance-monitor utilization.

Reproducer / motivating experiment

Project: https://github.com/Loong0x00/rtx5090-smsp-badapple

An RTX 5090 has 170 enabled SMs and four SMSPs per SM: 170 × 4 = 680 = 34 × 20. The project launches 170 cooperative CTAs, limits occupancy to one CTA per SM, and uses four warps per CTA to encode a 34×20 binary animation as FP32 issue activity versus nanosleep.

The included placement and contention probes establish:

• simultaneous sampled coverage of 170/170 SMs;
• four distinct physical warp slots per CTA;
• on the tested GB202, warps 0/1/2/3 do not contend while 0/4/8/12 do, supporting the empirical formula SMSP = physical_warpid & 3 for this exact stack.

However, the experiment cannot observe the resulting 680-channel physical activity pattern through a supported NVIDIA monitor. The visible output has to be reconstructed through application instrumentation, which is precisely what the requested monitor would avoid.

Research notes, including PTX/SASS special-register probes and Nsight Compute sampling output:

https://github.com/Loong0x00/rtx5090-smsp-badapple/blob/main/research/SM120_HIDDEN_REGISTERS.md

Feature request

Could NVIDIA expose an authorized, documented profiler/CUPTI mode that provides:

1. a stable per-capture physical SM instance ID;
2. the SMSP/scheduler index within that SM;
3. timestamped activity, issued-instruction, active-warp, or programmable-trigger samples;
4. clear documentation of aggregation, sampling, replay, and identifier lifetime;
5. an offline/privileged mode if unrestricted access would create performance-counter security concerns?

This does not need to be a CUDA programming-model affinity guarantee, nor does it need to permit warp placement control. A profiler-only observation interface would be sufficient.

If this belongs to the user-mode Nsight Compute/CUPTI stack rather than the open kernel modules, please point me to the correct public feature-request tracker. I am posting here because the repository's issue configuration directs feature requests to the Ideas discussion category, and the necessary per-instance counter plumbing may cross the driver boundary.

Environment

• GPU: NVIDIA GeForce RTX 5090 / GB202, compute capability 12.0
• Driver: 610.57.04
• CUDA: 13.3
• Nsight Compute: 2026.2.1
• OS: Arch Linux, kernel 7.2.3-arch1-2