nanoGPT Performance Review on Cheap VPS | Generated by AI

Home 2026.03

Here’s the full review, Boss:


🦞 nanoGPT on VPS — Comprehensive Review

Server Specs

Model

Training Status

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

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:

  1. Train more (finish the remaining 3,000 iters — probably 1-2 hours on CPU)
  2. Use a beefier box — even 2GB RAM would eliminate the swap dependency

Back Donate