Published: July 3, 2026 | QSCompute
Every edge AI node generates data. Camera frames at 30 FPS. Inference logs at 100 lines per second. Model checkpoints every 1,000 iterations. Over a 30-day retention window, a single 8-camera AOI station produces 15–20 TB. Storing all of it on high-performance NVMe is wasteful; archiving it all to a remote NAS breaks real-time access. The answer is a tiered 存储 architecture — and this guide explains how to design one.
| Tier | Meaning | Access Latency | Typical Media | Cost per GB |
|---|---|---|---|---|
| Hot | Data actively being read/written by inference pipeline | <100 µs | NVMe Gen4 M.2 / U.2 | $0.15–$0.35 |
| Warm | Recent data retained for re-analysis (last 7–30 days) | <1 ms | SATA SSD / NVMe QLC | $0.05–$0.12 |
| Cold | Long-term archive for compliance or model retraining | 10–100 ms | HDD / NAS / Cloud | $0.01–$0.03 |
The hot tier is where your inference models read frames and write results. It needs low latency (sub-100 µs) and enough throughput to handle concurrent camera streams. A single 1080p YUV camera at 30 FPS writes about 90 MB/s — eight cameras push 720 MB/s, which saturates a single SATA link. NVMe Gen4 x4 provides 6,000+ MB/s of headroom.
Sizing rule: Hot tier capacity = (number of cameras × stream bitrate) × (max minutes before data promotion to warm). For most pipelines, 1–4 TB of NVMe is sufficient — data moves to warm tier within minutes to hours.
Media choice: Use TLC NVMe with ≥1 DWPD endurance. The hot tier sees near-continuous writes and cannot tolerate QLC's write cliff after SLC cache exhaustion.
Warm tier holds data from the last 7–30 days for re-analysis: "replay last week's footage and re-score with the updated model," "check if that defect pattern appeared before," or "generate the weekly quality report." Access is less frequent — perhaps a few queries per day — but latency still matters. Nobody waits 5 seconds for a file listing during a production meeting.
Sizing rule: Warm tier = hot tier × retention days. For 8 cameras at 720 MB/s total, a 7-day warm tier needs approximately 400 TB. That forces tough decisions: store keyframes only? Downscale to 720p? Retain only frames flagged by the inference model?
Media choice: QLC NVMe (Samsung PM9C1a, Micron 2650) or high-endurance SATA SSDs in RAID. QLC drives at 8 TB offer compelling $/GB for warm tier, but monitor write amplification — QLC cells wear 2–3× faster under sustained random writes.
Cold tier is long-term storage: 6–12 months of inference results, model training datasets, and regulatory compliance archives. Latency doesn't matter. Cost per terabyte does. This tier can live on spinning rust (enterprise HDDs in a local NAS), network-attached storage, or cloud object storage (S3, OSS).
Sizing rule: Cold tier = warm tier × archive months × compression ratio. With H.265 compression and frame sampling (1 frame per second for archive), a year of 8-camera data can compress to ~30 TB.
Media choice: 20–22 TB enterprise HDDs (Seagate Exos, WD Gold) in RAID 6. For multi-site deployments, a central NAS with 10/25 GbE backhaul.
| Pattern | Hot | Warm | Cold | Best For |
|---|---|---|---|---|
| Single-Box Edge | 1× M.2 NVMe 2 TB | 1× U.2 NVMe 7.68 TB | NAS (off-box) | 1–4 cameras, local-only |
| Edge Server | 2× U.2 NVMe 3.84 TB (RAID 1) | 4× U.2 NVMe 7.68 TB (RAID 5) | NAS + periodic cloud sync | 8–16 cameras, factory floor |
| Cloud-Hybrid | 1× M.2 NVMe 2 TB | Cloud block storage | Cloud object (S3 Glacier) | Multi-site with fiber backhaul |
QSCompute supplies pre-configured storage solutions for edge AI deployments with validated tiering setups:
Design Your Edge AI Storage Tiering Strategy
Contact: +86 189-9192-7716 | info@qscompute.com