Mingxin

Mingxin FX vs DDN: Route Differences from HPC Storage to Inference Acceleration

Direct answer

DDN is the incumbent power in HPC/AI training storage, with deep EXAScaler share in supercomputing. Mingxin FX does not do general HPC storage — it focuses on inference KV tiering and load acceleration, winning on reproducible measurements and domestic-platform adaptation.

Stated honestly: DDN's strengths

DDN has cultivated HPC storage for decades; EXAScaler (Lustre) has a deep installed base in global supercomputers and large AI training clusters, the new Infinia platform evolves toward AI data pipelines, and there is long-standing alignment with NVIDIA SuperPOD reference architectures (public-source basis). For large-scale training-data pipelines, DDN is a mature option.

Mingxin FX's differentiation

A different lane: Mingxin does not compete with Lustre on training-data pipelines — we fill the most-missing inference layer: KV cache tiering (+29–40% throughput, R2/R3) and model loading (6.2–9.3×, R9).

Light delivery: standard Ethernet RoCEv2 fabric, an open software stack, and a joint test that starts with a single array (the ten-week G1–G4 process) — no supercomputer-grade infrastructure rebuild.

Price structure: fully-populated reference prices published (FX200 ≈ ¥1,797/TB, lowest per-TB of the line) for predictable budgeting.

FAQ

We already run Lustre/EXAScaler — do we still need FX?

Look at the bottleneck: if training-data reads are fine but inference TTFT is high and loading slow, what's missing is the KV tiering and load-acceleration layer — complementary, not a replacement.

Can FX serve as training storage?

Checkpoint writes measured 1.9× faster (R1) and training-data reads accelerated (R9); but for massive parallel training-data pipelines we honestly still recommend a professional parallel filesystem.

Why not build general-purpose storage?

Going deep in a niche is a small company's survival rule: in KV tiering we can produce 480B production-form signed measurements; the general market belongs to general vendors.

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.