Choose the Best Storage Tier for Local LLM Work

Local LLM storage tiers

Model loading and network bottlenecks

Model weights are large sequential files, so a fast HDD array can source them reasonably. Sharded models, metadata, many concurrent loads and vector databases add random access where SSDs pull ahead. The best result usually comes from copying the chosen model to local NVMe before starting inference.

1GbE, 2.5GbE and 10GbE impose progressively higher transfer ceilings, but the slowest component wins. A fast all-flash NAS behind 1GbE wastes most of its throughput; a 10GbE link cannot rescue a single fragmented disk from random-I/O limits.

Endurance, capacity and backup

Reading model weights is not usually an SSD-endurance problem. Checkpoint creation, dataset transformation, embedding generation and database compaction can write heavily, so TBW ratings and thermal behavior matter for the active SSD tier.

Price per terabyte is meaningful only within a role. HDD wins capacity economics; NVMe wins latency and local bandwidth. Presenting both on one leaderboard would encourage the wrong purchase, so the live tables remain separate.

Best choice by workflow