Mingxin

Mingxin FX vs WEKA: The KV-Cache Extension Scenario Compared

Direct answer

WEKA's parallel filesystem and Augmented Memory Grid lead the KV-extension direction in the NVIDIA ecosystem. Mingxin FX achieves the same class of capability with an NVMe-oF block layer + the open LMCache stack, strong on domestic platforms and reproducible verification.

Stated honestly: WEKA's strengths

WEKA's high-performance parallel filesystem has years of standing in HPC/AI training; its inference-facing Augmented Memory Grid (KV-cache HBM extension) is the same idea class as Mingxin's KV tiering, aligns tightly with the NVIDIA ecosystem (GDS, Dynamo/NIXL) and has rich international references (public-source basis). This also corroborates that external KV tiering is an industry-recognized correct direction.

Mingxin FX's differentiation

Technical route: open software stack (vLLM/LMCache) + standard NVMe-oF hardware — no proprietary client or protocol lock-in; customers can audit every layer.

Domestic compute coverage and measurements: 480B production-form measurements on MI308X (R2/R3), Ascend/MetaX measurements (R9/R5), and a parallel-read patch contributed to the open-source community (R8).

Verification style: signed reports + code export + gated joint tests, all numbers reproducible — we do not ask customers to believe, only to re-measure.

FAQ

Parallel filesystem or NVMe-oF block layer — which is better?

Each has its domain: filesystems favor unified namespaces and training data; KV tiering workloads are chunked object reads/writes where block + LMCache suffices with a thinner stack. Choose by workload; no need to take sides.

Can the two coexist?

Yes: training data on a parallel filesystem and KV tiering on FX arrays is a sensible combined architecture with no conflict.

Does Mingxin have international references?

Honestly: Mingxin's public measurements are mainly on domestic platforms (R1–R9) and international delivery is in its early stage — exactly why we made all data open and reproducible, letting evidence speak instead of references.

Data sources (verifiable)

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 ↓
R5FX100 KV-Cache Benchmark (14B, HBM efficiency, official, No.-004)2026-07-03
Download report PDF ↓
R8FX100 KV-Cache AMD Code Export + Raw Benchmark Data2026-07
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.