⚡️ Feature: 'GPU Usage' in Dashboard / System Overview

[mrtimothyduong]

Feature Description

I think it would be nice to enable 'GPU Usage' as an available utilisation metric/log as part of dashboard live statistics.

Problem It Solves

Provides another level of live telemetry, useful for diagnostics on system usage of GPU

Proposed Solution

Stats could include:

• RAM Usage / RAM Total (GB)
• GPU Temp ('C/'F)
• Wattage (W)

Of course would need to look at support for different GPU integration capabilities (e.g. via docker run -it --device /dev/dri)

• Nvidia
• AMD dGPU + iGPU
• Intel dGPU + iGPU

I assume the implementation would need to be via Docker socket + Arcane Agent.

Existing UI:

Alternatives Considered

Using other GPU monitoring: e.g. Beszel

Additional Context

No response

Would you be willing to work on this feature?

• Yes, I'd like to implement this feature
• I could help with parts of this feature
• No, I'm just suggesting the feature

[kmendell]

This may be tricky to implement, as since its inside a container it would have to have the gpu pased through no? IF you want to take a inital stab at it feel free. :)

[mrtimothyduong]

Easy will give it a crack :)

[mrtimothyduong]

Got MVP working but nvidia-smi doesn't work with alpine out the box. Needed to install gcompat to get nvidia-smi working but that's too specific and adds additional variables. Once I got it working this is what it looks like as seen below.

What is the best approach for getting telemetry data from GPU cards?

1. Use Nvidia-smi: Requires gcompat installation to image or post build. Too proprietary. Does not support other methods. But it does provide good level of telemetry ✅
2. Use nv-top: More generic approach, supports intel/amd/nvidia. Downside is less granular telemetry for nvidia but that's what I'm aiming for (for now) 🚩

I've built opt. 1 MVP but will pivot to opt. 2 since it should be more generic/wider support (and I don't have non-nvidia gpus to test!)

[kmendell]

is there a way we can somehow do both? In my personal opinion id like to try and have an option for all gpus, as the nvida alone would not cover AMD and what not, so id prefer option 2, but i have to belive there is a go library ou there that can monitro both amd and nvida gps

[mrtimothyduong]

I'll implement a clean version of Opt 1. after I review the Go Libraries for gpu monitoring (Opt 1+2).

Once both are ready I'll have it online.

@kmendell , do you have a Non-nvidia GPU for testing?

Also my implementation is as follows: gpu_service.go is a tacked on internal service rather than editing the original system_service.go.

Current forked repo: https://github.com/mrtimothyduong/arcane-gpu-mon

• GPU-MONITORING.md
• backend/internal/api/system_handler.go
• backend/internal/bootstrap/router_bootstrap.go
• backend/internal/bootstrap/services_bootstrap.go
• backend/internal/services/gpu_service.go
• frontend/src/lib/types/system-stats.type.ts
• frontend/src/routes/dashboard/+page.svelte

[mrtimothyduong]

Done some research, will proceed with go-nvml for nvidia GPUs and sysfs for AMD/Intel as fallback depending on detection.

Docker Container + GPU host (Alpine)
└── gpu_service.go (Go binary)
    ├── Vendor detection (/sys/class/drm)
    ├── Metrics source:
    │     ├── NVIDIA → NVML
    │     ├── AMD → sysfs (/mem_info_vram_used, /temp1_input)
    │     ├── Intel → sysfs (/gpu_busy_percent)
    └── Polled by → Arcane Dashboard (displays name, memory used/total with opt temp/util/power)

Gaps/unknowns: Hosts with multi-gpu's or mixed gpu's

[kmendell]

Yeah i found that one too, let me know if you envounter any issues, i dont have a GPU in my docker host to test, but well cross that bridge when we get there. :)

[mrtimothyduong]

I'm not able to work go-nvml with Alpine image base due to the lack of compatibility. Tried using gcompat, libc6-compat but due to missing symbols i can't get even the frontend to boot cause glibc symbols fail to load.

At this stage, i might as well just use nvidia-smi with a post-install .sh task. (I was able to get this working)

Let me know what your thoughts are.

[kmendell]

@user00265 Shed some light on this, i know you used to run GPU's in a alpine box or you said you did at leats maybe you have some ideas.

[kmendell]

Otherwise i guess nvida-smi will work that supported all gpus right?