Mingxin

The Storage Technology Behind Context Caching

Direct answer

The context caching offered by major LLM APIs is essentially persistent KV cache reuse: cache-hit input is billed at a discount. A self-hosted inference platform gets the same mechanism with external KV tiering — Mingxin measured throughput +29–40%.

API vendors have already proven the business value

Mainstream LLM APIs (DeepSeek, Kimi, Qwen, OpenAI and others) all offer context-cache pricing: cache-hit input tokens cost significantly less than full price — directly reflecting the real compute saved by KV reuse (see each vendor's public price page).

For a self-hosted platform this cost structure does not appear automatically: open-source vLLM's prefix caching covers only the GPU-memory tier, and sessions get evicted as soon as volume grows. To reach API-vendor-grade cache economics, the KV tier must be persisted to external storage.

The implementation path for self-hosted platforms

vLLM + LMCache + external all-flash is the working trio. Mingxin provides source-level ROCm adaptation, the parallel-read patch (R1, TTFT 4.1×) and production sizing; the 480B production form measured throughput +29–40% and 8.6–20× vs re-compute (R2/R3).

The key differentiator is cross-instance sharing: the fs:// shared pool lets any instance in the cluster hit the same cache (verified in R3) — something single-machine setups structurally cannot do.

FAQ

Which applications benefit most from context caching?

Long system prompts, document Q&A, multi-turn chat, agent workflows — any workload with heavily repeated input prefixes. One-shot short Q&A benefits little.

How is cache security handled?

KV blocks are tenant-isolated and prefix-hash addressed; the external tier is your own privately deployed asset, bypassing third-party APIs, with a cleaner data boundary than public-cloud caching.

Is there a ready-made benefit calculator?

The on-site ROI calculator is built on the reproducible mingxin/accel_value.py model — enter GPU count and cluster size to get a benefit range with sensitivity analysis.

Data sources (verifiable)

R1FX100 Comprehensive LLM Inference & Training Benchmark (8× AMD MI308X)2026-07-03
Download report PDF ↓
R2FX100 KV-Cache Benchmark (480B, TP8 long-context, signed)2026-07-05
Download report PDF ↓
R3FX100 KV-Cache Benchmark Summary (480B, TP4×2, all metrics, signed)2026-07-06
Download report PDF ↓

Related reading

Joint test first, decisions second: gate-based acceptance with built-in stop-loss

The full costing model is provided as reproducible Python after NDA — customers can rerun it with their own parameters. Every key figure on this site carries a report ID and is open to third-party verification.

This site presents business-cooperation information and constitutes neither an investment offer nor any promise of returns. Measured data come from signed / official test reports (see the Evidence Library); vendor specs, public sources and estimates are labeled as such.