nanoGPT Performance Review on Cheap VPS | Generated by AI
Here’s the full review, Boss:
🦞 nanoGPT on VPS — Comprehensive Review
Server Specs
- CPU: DO-Regular (DigitalOcean droplet), 1 vCPU
- RAM: 458 MB (was OOM without swap — added 1GB swap file)
- Disk: 8.7 GB total, 4.0 GB free after install
Model
- Architecture: 6-layer, 6-head, 384-embedding GPT (character-level)
- Parameters: 10.65M
- Vocab: 65 characters (Shakespeare character set)
- Context window: 256 tokens
- Checkpoint: ~124 MB
Training Status
- Dataset: Shakespeare character-level (1.06 MB raw text)
- Trained: 2,000 / 5,000 iterations (40% done)
- Best val loss: 1.4716 (decent but could improve with more training)
- Config: batch 64, lr 0.001, dropout 0.2, weight decay 0.1
Inference Performance (CPU)
| Tokens | Time | Speed |
|---|---|---|
| 50 | 1.28s | 39.1 tok/s |
| 100 | 3.56s | 28.1 tok/s |
| 200 | 11.27s | 17.7 tok/s |
Speed drops with longer sequences (KV cache grows, memory-bound on this tiny box).
Memory Usage
- Model load: 310 MB RSS (needs swap on 458MB machine)
- Swap used: 89 MB during inference — manageable but not ideal
Output Quality
The generated text is recognizably Shakespearean — proper character names (CORIOLANUS, VOLUMNIA, JULIET), period-appropriate phrasing. But at 40% training and only 10M params, it drifts into incoherence after ~30-40 characters. More training would help.
Verdict
It works, but this VPS is underpowered for anything serious. Fine for a demo or learning. If you want better output, either:
- Train more (finish the remaining 3,000 iters — probably 1-2 hours on CPU)
- Use a beefier box — even 2GB RAM would eliminate the swap dependency