Frequently Asked Questions
Common questions about products, performance data, pricing and collaboration. Every number cites its report — see the evidence library.
QWhat is the Mingxin FX series storage acceleration platform?
The FX series is Mingxin's all-flash NVMe-oF storage acceleration platform (FX100/FX200/FX300 in production; FX400 expected in volume production by end of 2026). Combined with a tiered KV-cache software stack (vLLM + LMCache with Mingxin's parallel-read patch), it accelerates LLM inference and training workloads on the storage side. FX100 offers 100Gb per port and 16M IOPS; FX300 is a PCIe 5.0 platform with 400Gb per port and 60M IOPS.
QHow much does tiered KV-cache storage acceleration improve inference performance?
In signed benchmarks on a production deployment of a 480B model (Qwen3-Coder-480B-FP8) on 8× AMD MI308X: inference throughput improved 29–40% (+29% at concurrency 8, +40% at the optimal operating point), time-to-first-token dropped 26–32%, and versus recomputation without external storage the speedup was 8.6–20×. Sources: test reports R2/R3, downloadable from the evidence library.
QHow can these performance numbers be verified or reproduced?
All key figures come from signed/official test reports (R1–R9; R4 was issued by a third-party testing organization), downloadable from the evidence library. Test code (LMCache parallel-read git patch, load clients, orchestration and forensic scripts) and raw data are packaged as R8, so any third party can independently reproduce the conclusions. Two independent measurement rounds at the same tier differ by ≤5%.
QHow much does a fully configured FX system cost?
Reference prices for fully configured systems (including 24 enterprise NVMe SSDs): FX100 about ¥371,200 (≈¥2,014/TB), FX200 about ¥331,200 (≈¥1,797/TB, lowest per-TB cost), FX300 with DPUs about ¥924,000 (≈¥5,014/TB, Gen5 performance tier). SSDs are priced at market rates; final pricing is set in the purchase contract, and the calculation is reproducible.
QHow do we start a collaboration?
Collaboration starts with a gated joint validation of about 10 weeks with four gates (G1 delivery acceptance → G2 single-node baseline → G3 main gate: TTFT reduction ≥25% and throughput gain within the measured +29–40% band → G4 72-hour stability). If a gate fails, the project stops — losses are cut early. The full costing model is provided as reproducible Python after NDA. Use the contact page to schedule.