Mingxin

AI Inference Storage Acceleration: The Third Way to Raise Cluster Output Without Adding GPUs

Direct answer

Inference storage acceleration uses a high-performance storage tier to raise a GPU cluster's token output: KV tiering avoids re-compute, models load in seconds, switching stops idling. Mingxin measured gains of 29–40%, 6.2–9.3× and 1.9× on the three levers respectively.

Three storage levers

Lever one: KV cache tiering — long-context cold recovery without re-compute, throughput +29–40% (R2/R3, 480B production form).

Lever two: model-load acceleration — under an NFS bottleneck, DeepSeek-32B serving load fell from 691 s to 112 s (6.2×) and 70B from 1399 s to 150 s (9.3×) (R9, measured on Huawei Ascend 910B).

Lever three: model switching and checkpoints — at 20 switches/hour, effective compute utilization rose from 46.7% to 62.8%; training checkpoint concurrent writes are 1.9× faster (measured, R1).

Why "efficiency before expansion" holds

The marginal cost of adding GPUs is linear, while storage acceleration costs within roughly 10% of the GPU budget (joint-test basis). When the cluster bottleneck is re-compute, loading, or switching idle time, new GPUs just buy more idle silicon; fixing the storage tier first yields more effective capacity per yuan.

This is not a universal conclusion: if your workload is all short chats with high cache hits, storage acceleration gains are limited. That is why Mingxin engages through gated joint tests — the G3 main gate requires TTFT reduction ≥25% and throughput inside the measured +29–40% band, with stop-loss if unmet.

FAQ

How is this different from just buying faster local SSDs?

Local drives cannot be shared across nodes, are capacity-limited, and lose data on failure. An external array offers pooled capacity, cross-instance sharing (R3) and redundancy — and still beat local drives on TTFT by 26–32% in measurements (R2).

Are domestic GPU platforms supported?

Measurements cover AMD MI308X (R1–R4), Huawei Ascend 910B (R9) and MetaX N260 (R5); the software stack is a source-level adaptation of open-source vLLM/LMCache.

How do I verify these numbers?

Every figure comes from signed/official test reports (downloadable in the Evidence Library); the R8 code export package contains the patch, load clients and raw data for independent third-party reproduction.

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 ↓
R9Mingxin FX100-HBMM vs NFS Baseline on Huawei Ascend 910B2026-05-30
Contact us for access →

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.