The output of many metrics are 0 when running metric tests

[Kepontry]

Hi, jwludzik,

I’m using the NPU on Lunar Lake with driver version 1.13.0 and Ubuntu 24.04. While running the metrics test in umd-tests and attempting workload profiling, I noticed that out of the eight metric groups, I only get output from the first metric in the NOC group. However, the outputs for ShaveIL1Cache, ShaveDL1Cache, and ShaveL2Cache are all zero, and I do not see any output for other metric groups such as ResMgr.

I tested this using the add_abc and memory copy examples in experiments.

My questions are:

How can I retrieve performance metrics for the NPU, similar to PMU counters for the CPU?
Could this issue be due to the workload being too simple?
Is it possible that the workloads do not invoke the Shave module at all?
Or is the support for these metrics not yet fully implemented?

Would appreciate any insights—thanks!

[jwludzik]

I am very sorry for late response. The "Shave*" metric groups can be ignored, they are not supported any more. They are left for backward compatibility. The only meaningful "profiling" metric is "NOC" as it allows you to measure DMA bandwidth.

How can I retrieve performance metrics for the NPU, similar to PMU counters for the CPU?

From user point of view the best way to profile a network is to use "-pc" option from OpenVINO benchmark_app to see network layer by layer performance. As well there is VTune tool to profile the different IPs (including NPU) (check https://www.intel.com/content/www/us/en/docs/vtune-profiler/user-guide/2024-2/npu-exploration-view.html )

Could this issue be due to the workload being too simple?

No

Is it possible that the workloads do not invoke the Shave module at all?

No, as the "Shave*" counters should be ignored

Or is the support for these metrics not yet fully implemented?

I am not sure about it. The implementation of gathering metrics is done between driver and firmware. The firmware is responsible to provide metrics to driver. It is highly possible that there will be more metrics added to firmware

[Kepontry]

Thanks a lot, Jozef. There are six metrics in the NOC group: NOC_noc, NOC_DMA_00, NOC_DMA_01, L2C_Riscv, and NOC_timestamp. However, I’m only seeing a non-zero value for NOC_noc — the others are all zero. Is that expected?

Also, regarding DMA bandwidth calculation — is it correct to assume it's computed as NOC_DMA_00/NOC_timestamp or NOC_DMA_01/NOC_timestamp?

[jwludzik]

Yes, it depends on workload. The short description of every metric:

• NOC_noc - total bytes transferred
• NOC_DMA_X - number of bytes transferred per DMA channel. Sum of those is NOC_noc
• L2C_Riscv - number of bytes transferred by the NPU chip
• L2C_Shave - number of bytes transferred by the NPU kernel execution unit
• NOC_timestamp - timestamp of the sample