Summary

This article explains how to calculate the effective Streaming Multiprocessor (SM) usage of SupremeRAID™ AE workloads on Linux systems using NVIDIA DCGM and OpenCL (clinfo) tools.

This method helps administrators understand the actual GPU compute resources consumed by SupremeRAID™ AE and is useful for:

• GPU capacity planning

• vGPU / shared GPU resource evaluation

• Workload sizing and performance analysis

Environment

• Product: SupremeRAID™ AE

• Operating System: Linux

• GPU Vendor: NVIDIA

• GPU Type: Discrete NVIDIA GPUs (e.g., H100 / H200 / A100 / L40)

• Privileges: Root or sudo access required

Prerequisites

Ensure the following requirements are met before proceeding:

• NVIDIA GPU driver is properly installed

• SupremeRAID™ AE workload is running

• Root or sudo privileges are available

• Internet access is available (for package installation)

Procedure

Step 1 – Install Required Tools

Install the tools required to query GPU hardware information and monitor real-time GPU metrics:

• clinfo – Used to retrieve OpenCL platform and device information

• datacenter-gpu-manager (DCGM) – NVIDIA tool for monitoring GPU utilization metrics

sudo apt update
sudo apt install clinfo datacenter-gpu-manager

Note: On RHEL / Rocky / Alma Linux, package names or repositories may differ.

Step 2 – Enable NVIDIA DCGM Service

Start and enable the NVIDIA DCGM service:

sudo systemctl --now enable nvidia-dcgm

Optional verification:

systemctl status nvidia-dcgm

Step 3 – Monitor SM Active Ratio

Use dcgmi to monitor the SM Active metric.

• Metric ID 1002 (SM Active)

  • Indicates the ratio of time during which at least one warp was active on an SM

• -i <GPU_ID> specifies the target GPU index

# Example: Monitor GPU ID 0
sudo dcgmi dmon -e 1002 -i 0

Record the reported value (typically between 0.0 and 1.0).

Example:

• 0.25 → 25% SM active ratio

Step 4 – Retrieve Total SM Count

Use clinfo to retrieve the total number of hardware Compute Units (SMs) for a specific GPU.

clinfo -d P:D | grep "Max compute units"

Parameter Explanation

• P – OpenCL platform index

• D – Device index within the selected platform

This allows you to query a specific GPU in multi-GPU or multi-platform systems.

Identify Platform and Device Index

To list all available OpenCL platforms and devices:

clinfo -l

Example Output

Platform #0: NVIDIA CUDA
 +-- Device #0: NVIDIA H200
 +-- Device #1: NVIDIA H200
 +-- Device #2: NVIDIA H200
 +-- Device #3: NVIDIA H200
 +-- Device #4: NVIDIA H200
 +-- Device #5: NVIDIA H200
 +-- Device #6: NVIDIA H200
 `-- Device #7: NVIDIA H200

In this example:

• Platform index: 0 (NVIDIA CUDA)

• Device indices: 0 through 7

Example Commands

Query SM count for Device #0:

clinfo -d 0:0 | grep "Max compute units"

Query SM count for Device #3:

clinfo -d 0:3 | grep "Max compute units"

Step 5 – Calculate Effective SM Usage

Apply the following formula: