Skip to main content
This guide is for anyone who will use the Silicon Data portal and run the agent on a Linux GPU server. API details are on the SiliconMark API pages.

​
Before you start

You need a Silicon Data account with access to SiliconMark in the portal, and a Linux x86_64 or ARM64 machine with NVIDIA or AMD GPUs and current drivers (NVIDIA downloads, AMD support). On AMD, your user has to belong to the render and video groups. If it does not, add them and log out and back in:
You do not need to install Python or PyTorch yourself. The agent creates its own environment with what it needs for its benchmarks. To check that the GPU is visible before you start:

​
Create a job in the portal

  1. Open SiliconMark and click Create New Job.
  2. Choose a benchmark type (see benchmarks). QuickMark, the general speed test, is selected by default.
  3. Enter a job name you will recognize later, up to 100 characters.
  4. Set the number of nodes. Use 1 for a single machine. Raise it only when several servers should join the same job, which is what the network test between machines needs. The slider runs from 1 to 256.
  5. Add tags if you want to filter later, for example GPU model, provider, or site. You can add up to 10 tags of up to 50 characters each.
  6. Click Confirm.
The portal then shows a Job Created panel with copy-paste commands for that job. Copy them before you move to the GPU machine.

​
Commands the portal gives you

On the GPU machine, download the agent for your CPU type, make it executable, and run it with the job key from the portal. Use the exact block from the Job Created panel; the example below shows the shape, and your key will differ.
On ARM machines the portal points at https://downloads.silicondata.com/arm64/agent instead. The agent sets up an automatic private environment under ~/.local/silicon_data. The job key ties the run to your account, so results are saved and a PDF can be generated. How long you wait depends on the test. QuickMark’s speed pass often takes a few minutes, and the heat-and-cool check after it adds several more. SiliconAudit is a deep hardware scan and can run for hours. The Llama serving test may spend a long time pulling containers before the first numbers appear. In a terminal the agent shows a live progress view. In scripts or CI it prints plain logs.

​
Watch the GPUs while it runs

From another SSH session:
On a machine with several GPUs, QuickMark measures each GPU on its own and then all of them together.

​
After the job finishes

The job should show Completed. Open it for per-GPU numbers, the fleet comparison, and the PDF report.

​
Other ways to run

​
Non Registered Account Job Run

You can run the agent with no job key as a quick local check. Results are not stored on your account and there is no PDF report.

​
Extra keys for some tests

QuickMark and SiliconAudit need no extra cloud keys. The Llama tests do. Llama 3 inference needs NGC_API_KEY and a container runtime; Podman is preferred and Docker works. Llama 3 fine-tuning needs HF_TOKEN (a Hugging Face token) and Docker, and an NGC key helps it pull the training container. Set these on the machine before you start the agent. The benchmarks page has the details. QuickMark itself runs as a native program and does not need Docker.

​
Remove local agent environment

This removes the agent private environment and its binaries. Portal jobs are not affected.

​
Support

support@silicondata.com