Building a LLM Powerhouse: RTX 4090 + Ollama + OpenClaw

In this post, I’ll walk through my weekend project:

• setting up a dedicated openclaw server using my old intel macbook pro (2015 version)
• setting up a dedicated local LLM server using an RTX 4090
• configure openclaw to use my dedicated LLM server

The Strategy: Compute where it counts

The core philosophy of this setup is hardware efficiency.

• The Workhorse: A Windows machine with an RTX 4090 (24GB VRAM). This stays on 24/7 to handle the heavy lifting.
• The Clients: An older Intel Mac repurposed as a dedicated terminal/gateway, this allow openclaw to use UI interface.

1. Setting up the Windows Server (RTX 4090)

The goal was to run GLM flash 4.7 (which supports images) via Ollama. Recommended by Ollama Ref

The Memory Challenge

Running ollama run glm-flash-4.7 requires about 17GB of VRAM. However, I encountered an issue where the KV Cache exploded from 5GB to 19.7GB, hitting a “model request too large for system” error. I resolved this by setting context window to 64000, which consume ~93% of the VRAM.

• Lesson learned: You must tune your context window to avoid VRAM blow-ups. If the KV cache + model weights exceed VRAM size (24GB in my case), Ollama will return error.

Configuration for Remote Access

To make the Windows machine accessible over the network:

1. Set Environment Variables:

1
2

setx OLLAMA_ORIGINS "*" /m 
setx OLLAMA_HOST "0.0.0.0" /m

2. Run as a Service: I used nssm (Non-Sucking Service Manager) to ensure Ollama starts automatically with Windows.

1

nssm install OllamaService "C:\Users\xxxxx\AppData\Local\Programs\Ollama\ollama.exe" serve

2. Repurposing the Intel Mac

I decided to give my old Intel Mac a second life as an “Always-On” bridge.

• OS Prep: Factory reset and disabled sleep so it remains reachable.

1

sudo pmset -a disablesleep 1

• Networking: Installed ZeroTier to create a secure virtual private network between my Windows machine and the Mac. It also keeps internal IP addresses stable.

3. Troubleshooting & Testing

How to verify your remote connection:

From the Mac (client), run a simple curl command to the Windows IP:

1

curl http://<windows-machine-ip>:11434/api/version

If it returns a version number (e.g., 0.1.8), your firewall and host bindings are correct.

GPU Utilization

If you’re on Windows and Ollama isn’t hitting your GPU, check your logs. You should see offloading 65/65 layers to GPU. If you see layers being offloaded to CPU, check your VRAM overhead.

Summary

This setup serves as a cornerstone for me to learn more about openclaw without spending money on LLM API. I am going to expand this setup and offload some dirty works to openclaw and see where are the limits.