Storage acceleration · Domestic AI compute · Full-stack AI datacenter services

Storage acceleration that turns idle GPU time into verified throughput

Mingxin builds the FX series of all-flash NVMe-oF storage acceleration platforms. In signed benchmark reports on a production 480B-parameter LLM deployment (8× AMD Instinct MI308X), tiered KV-cache offloading delivered 29–40% higher inference throughput, 26–32% lower time-to-first-token, and 8.6–20× speedup versus recomputation without external storage. All key numbers are backed by downloadable reports.

+29–40%

Inference throughput gain from tiered KV cache

480B production deployment, long-context cold recovery: +29% at concurrency 8 (lower bound), +40% at the optimal operating point (concurrency 16)

Source: Reports R2/R3 (measured)

TTFT −26–32%

Time-to-first-token reduction

480B·TP8, three concurrency tiers: TTFT p50 down from 10.17–35.73s to 7.53–26.35s

Source: Report R2 (measured)

8.6–20×

Speedup vs. recomputation without external storage

Recompute baseline TTFT p50 149.5s (conc. 16) vs. 11.85s with FX100; throughput 4.1 vs. 74.9 tok/s

Source: Report R2 (measured)

6.2–9.3×

Model-loading speedup vs. NFS

Huawei Atlas 910B platform: DeepSeek-32B service load 691s → 112s (6.2×), DeepSeek-70B 1399s → 150s (9.3×)

Source: Report R9 (measured, Ascend platform)

Download signed test reports FX series & pricing Contact us

This site presents business-cooperation information and does not constitute an investment offer or any promise of returns. Measured, vendor, public and estimated figures are labeled as such; all measured numbers come from signed/official test reports (see Evidence).