gpu-watch - A bashingly pretty nvidia-smi monitor

Description

A bash-based prettifier wrapper for nvidia-smi. A project vibe-coded using Claude.ai chat.

Current Status:

Author        : eabase
Date          : 2026-05-12
Version       : 1.0.5
Repo          : https://github.com/eabase/gpu-watch

Example Output:

The screenshot below is an example while running gpu-watch in an MSYS bash shell inside a Windows Terminal.
This particular state was obtianed while running Qwen3.6 35B A3B using LM Studio (lms) and Furmark.
I didn't think this was possible, but oh yeah!

In Terminal-1 (MSYS)

_{(FFS always include a screenshot in your GitHub repo!)}

In Terminal-2 (powershell)

_{Loading up Qwen3.6-35B and firing off a prompt.}

Click here for the CLI details.

Here are all the CLI commands to load the model, so you can reproduce my results.

# To see your locally downloaded models
# lms ls

LLM                                           PARAMS     ARCH            SIZE        DEVICE
qwen/qwen3.5-9b (1 variant)                   9B         qwen35          6.55 GB     Local
qwen/qwen3.6-35b-a3b (1 variant)              35B-A3B    qwen35moe       22.07 GB    Local
qwen3.6-27b-abliterated-heretic-uncensored    27B        qwen35          17.48 GB    Local
qwen3.6-35b-a3b-apex                          35B-A3B    qwen35moe       14.32 GB    Local  <--- We load this one !
unsloth/qwen3.6-35b-a3b                       35B-A3B    qwen35moe       19.52 GB    Local
...


# lms load --gpu 0.5 -c 32768 --estimate-only qwen3.6-35b-a3b-apex

Model: qwen3.6-35b-a3b-apex
Context Length: 32,768
GPU Offload: 50%
Estimated GPU Memory:   8.00 GiB
Estimated Total Memory: 14.80 GiB

Estimate: This model may be loaded based on your resource guardrails settings.

# Load using 50% GPU offload with 32 KiB context
# lms load --gpu 0.5 -c 32768 qwen3.6-35b-a3b-apex

Model loaded successfully in 12.60s.
(13.33 GiB)
To use the model in the API/SDK, use the identifier "qwen3.6-35b-a3b-apex".

# lms chat --stats qwen3.6-35b-a3b-apex

# <my chat prompt>
Can I run a 35B MoE model on a 8GB GPU by offloading sparse expert tensors to CPU?

...

# Here are the lms chat stats results, after having run 
# Furmark half of the time while the ai was responding.
# Without Furmark or any additional tuning, I get ~18 Tps.

Prediction Stats:
  Stop Reason: eosFound
  Tokens/Second: 12.90
  Time to First Token: 1.231s
  Prompt Tokens: 59
  Predicted Tokens: 2357
  Total Tokens: 2416

# /exit 

# lms unload
Model "qwen3.6-35b-a3b-apex" unloaded.

Here are the screenshots!

_{At GPU idle}

_{On Furmark}

_{On loaded qwen3.6-35b-a3b-apex (14 GB)}

_{During AI prompt response and running Furmark.}

✨ Features

• Color-coded output for the most relevant GPU metrics (utilization, VRAM, temperature, power)
• Color legend included in the display so thresholds are always visible
• Animated VRAM progress bar using pure ANSI/Unicode block characters
• Zero external dependencies - only native bash and ANSI escape codes
• Configurable polling interval (default: 2 seconds)
• Already tuned if your laptop video card is the NVIDIA GeForce RTX 4070 Mobile (and same tune specs as for the RTX 4060.)

💻 Portability

Environment	Status
Linux (bash)	✅ Supported
Windows 11 + MSYS2/MINGW64	✅ Supported
NVIDIA RTX 4070 Mobile [1]	✅ Tested
Other NVIDIA GPUs	⚠️ Requires msome manual tuning (see below)

_{[1] GPU specs for: RTX 4070 Mobile.}

Usage

./gpu-watch.sh [interval_seconds]   # default interval: 2s

Examples:

./gpu-watch.sh        # refresh every 2 seconds
./gpu-watch.sh 0.2    # refresh every 0.2 seconds
./gpu-watch.sh 5      # refresh every 5 seconds

⚙️ Required GPU Customisation

The color thresholds and limits are tuned for the RTX 4070 Mobile.
If you have a different card, adjust these values in the script:

Setting	RTX 4070M	Action
Max Temperature (throttle point)	86°C	Check your GPU specs
Max Power Draw	140 W	Check your GPU specs
VRAM*	8188 MiB	Check your GPU specs

_{* Change to GiB for cards with VRAM > 24 GB}

To see all queryable GPU metrics:

nvidia-smi --help-query-gpu
nvidia-smi --help-query-gpu | grep --color=always -iE '^".+$|$' -A5

# NOTE:
# - No spaces are allowed between items in the list!
nvidia-smi --query-gpu=timestamp,index,utilization.gpu,memory.used,memory.reserved,memory.total,memory.free,temperature.gpu,power.draw,power.max_limit,c2c.mode,mig.mode.pending,compute_mode,pstate,kmd_version,serial,persistence_mode,addressing_mode,accounting.mode,inforom.img,vbios_version

📊 RAM Progress Bar

Example usage of the percentBar() bash function:

. percent_bar_demo.sh
# The VRAM bottom border is 27 characters wide + 2 edges.
p=47; percentBar  "$p" 27 bar; printf '\U2595\e[0;32m\e[48;5;235m%s\e[0m\U258f%6.2f%%' "$bar" $p; echo
p=47; percentBar2 "$p" 27 bar; printf '\U2595\e[0;32m\e[48;5;235m%s\e[0m\U258f%6.2f%%' "$bar" $p; echo

NOTE:
For more advanced examples of colored progress-bars in bash, check out the percent_bar_demo.sh,
and the amazing bash-script collection by Felix Hauri, and various StackOverflow answers like this.

Thermal Notes

For the RTX 4070 Mobile we have the following thermal notes:

• GPU throttling begins at 86°C (core temperature).
• Hotspot temperatures can fluctuate by up to +/-10% without worry, as this reading is a "peak momentary" temp.
• Hotspot temperatures are not measured by nvidia-smi, but by other OS hardware layers.
  (Check projects like LibreHardwareMonitor, and others.)

🔗 Similar Projects

Project	Description
lablup/all-smi	Super nice Rust replacement for nvidia-smi
(yours here)	Know of a bash alternative*? Let me know.

_{* No bash alternatives have been found at the time of this writing.}

🐛 Issues & Contributing

Found a bug or want to add support for your GPU?

• Open an issue
• Submit a pull request

Respected Repos

Find all available and sorted emoji's here:

• https://phw198.github.io/github-emoji-cheatsheet/

---

No Stars For Funny Bars 😭

Or maybe I'm just on the wrong planet...