TECHNICAL BRIEF · VS NVIDIA CMX

Same tier.Two paths.

KV cache that leaves HBM should not be recomputed. NVIDIA CMX and Mingxin FX fill the same G3.5 context-storage tier.

480B long-context cold recovery · TTFT p50 · concurrency 16

149.5sRecompute, no external store

11.85sFX100 tiering

8.6–20× vs recompute · R2 signed

HBMHot sessions192 GB / GPU

Host RAMRecent≈1.5 TB / node

Local NVMePer node2 TB single drive

CMX or FX lives here

G3.5

Context storageShared KV across nodes184 TB / unit (FX100 full)

Network storageModels & archiveNFS / object

‹ Faster · smaller · costlier

Both fill G3.5. Neither replaces HBM. · Capacities are the R2 test-platform configuration

Larger · slower · shareable ›

Same tier. Different binding.

5× and +29–40% use different denominators. Do not cross-compare.

NVIDIA CMX

Public / vendor

Up to 5×vs “traditional storage” · baseline undisclosed

Binding: Rubin · BlueField-4
GPUs: NVIDIA accelerators
Stack: Dynamo · NIXL · DOCA Memos
Ship: Reported 2H 2026

Mingxin FX

Signed

+29–40%480B TP8 throughput · TTFT −26–32% · R2/R3

Binding: Standard 100 GbE RoCEv2
GPUs: MI308X · Ascend · MetaX
Stack: vLLM · LMCache, source-level
Ship: FX100 / 200 / 300 shipping

NVIDIA figures are vendor-public with no disclosed baseline. Every Mingxin figure carries a report ID (R1–R9) and is open to re-verification. Not a disparagement or substitution claim. NVIDIA CMX product page · NVIDIA Technical Blog