How to export raw data of Roofline in Nsight Compute (ncu)?

[juamaros]

Hello,

I am using Nsight Compute (GUI) and I see that I can generate the Floating Point Operations Roofline (FP32) directly in the graphical interface. However, I want to be able to obtain the raw data behind this graph, including:

1. Application characterization metrics: kernel FLOPs, SM cycles, L1/L2/DRAM traffic.

2. Roofline maximum ceiling points: the peak limits that define the Roofline “breakpoints” for FP32.

My questions are:

• Is there a way to export these values directly from ncu (CLI) or from the GUI to CSV, JSON, or another readable format?

• If I cannot get the exact graph points, is it possible to obtain the underlying metrics for both the kernel and the GPU peak ceilings, even if I need to process them myself afterward?

My goal is to be able to plot the complete FP32 Roofline, including both my kernel data and the GPU peak ceilings, for a convolution kernel on Orin, without relying on the GUI visualization.

Thanks in advance!

[AastaLLL]

Hi,

Yes, please find the document below for more information:

4. Nsight Compute CLI — NsightCompute 13.0 documentation

Thansk.

[juamaros]

In case anyone needs it, the parameters and formulas required to calculate the points shown in the Roroline graph are displayed in the additional information when selecting a point at least in the 2025 version of NVIDIA Nsight Compute.