Quick overview
Upgrade an existing PC for local LLMs: GPU swap, RAM, NVMe, and PSU checks before you replace the whole tower.
You may not need a new chassis to start with local LLMs. Many towers accept a GPU upgrade, extra RAM, and a larger NVMe drive while keeping the CPU for one more generation. The upgrade order should follow your measured bottleneck, not the newest launch video.
Quick answer
If inference OOMs on GPU memory, upgrade GPU VRAM or reduce model size. If offload stutters, add RAM. If downloads hurt, add NVMe. If reboots happen under load, inspect PSU and thermals before spending more.
Step-by-step upgrade path
1. Baseline tokens and temps with current hardware 2. Replace or add GPU if VRAM-bound 3. Upgrade PSU if new GPU demands it 4. Move model library to larger NVMe if needed 5. Add RAM if CPU offload is in use 6. Service cooling if thermals throttle
Compatibility checks
- Physical GPU length and power connectors
- RAM compatibility
- BIOS updates for new GPUs: safe BIOS update
When to stop upgrading
If the platform is PCIe or CPU limited and you need 70B local, a new workstation may beat stacked adapters.
How Roger N Reckon can help
Roger N Reckon performs GPU swaps, RAM installs, and diagnostics before you buy another full build. Describe your Ollama model tags when you book upgrade work.
Practical checklist
1. Write down the model size, quant level, and target context length you need. 2. Measure current GPU VRAM, system RAM, and free NVMe space before buying parts. 3. Confirm PSU wattage against the GPU vendor's required system power. 4. Plan cooling for sustained inference, not just short benchmarks. 5. Back up data before OS or hardware changes. 6. Request a written component list, testing scope, and warranty terms for any build or upgrade. 7. Re-test with the same prompt and context after changes.
Frequently Asked Questions
Can I use a laptop eGPU?
eGPU enclosures work for some users but cost and bandwidth limit value.
Will older CPUs bottleneck LLMs?
They affect offload speed more than GPU-fit models.
Should I replace thermal paste during GPU upgrade?
Replace CPU paste if temps were high; GPU paste is a specialist job.
Is 500W PSU enough after GPU upgrade?
Recalculate using the new GPU's required system power.
Can software tweaks avoid upgrades?
Smaller quants help until quality or OOM forces hardware.
Final takeaway
Upgrade the component that your logs prove is saturated.