Architecture

NVIDIA STX Reference Architecture

NVIDIA's rack-scale AI-storage blueprint (GTC 2026): three mandated modules — CMX context-memory tier, Vera-CPU storage frontend, AI Data Platform system — over Spectrum-X Ethernet.

6 connections 3 resources

Summary

What it is

NVIDIA's rack-scale AI-storage blueprint (GTC 2026): three mandated modules — CMX context-memory tier, Vera-CPU storage frontend, AI Data Platform system — over Spectrum-X Ethernet.

Where it fits

The procurement gate for AI-factory storage: WEKA, VAST, Cloudian, Hammerspace, and IBM all certified within a quarter. Resolves the CMX-vs-STX-vs-AIDP naming confusion: pod tier vs rack blueprint vs RAG-pipeline design.

Misconceptions / Traps

STX is not a product — it's a reference design vendors build against; "STX-ready" claims vary widely in depth (VAST runs natively on the DPU; most others attach externally).
The mandated Spectrum-X + BlueField fabric makes alignment an NVIDIA-exclusive commitment — evaluate the lock-in alongside the performance.

Key Connections

depends_on NVIDIA BlueField-4 — the DPU is the enforcement point
enables Inference Context Memory Storage (ICMS) — CMX is module one
solves Memory Wall / GPU Starvation — the blueprint's reason to exist

Definition

What it is

NVIDIA's rack-scale reference architecture for AI-native storage — **Storage Technology eXtensions** — introduced in full at GTC 2026. STX mandates three interconnected modules: a **Context Memory Storage system** (the physical [CMX / ICMS](/node/inference-context-memory-storage-icms) KV-cache tier), a **Storage Frontend System** running on NVIDIA Vera CPUs, and an **AI Data Platform System** on Rubin/Blackwell GPUs — all connected exclusively over Spectrum-X Ethernet with RDMA. The software stack spans NIXL (open-source KV-tensor transfer), DOCA Memos (the KV API SDK for [BlueField-4](/node/nvidia-bluefield-4)), and Dynamo (KV-locality-aware inference routing). Per [NVIDIA — STX](https://www.nvidia.com/en-us/data-center/ai-storage/stx/) and [NAND Research — STX/CMX infrastructure](https://nand-research.com/nvidia-stx-cmx-infrastructure-for-agentic-ai-context-storage/).

Why it exists

AI factories were assembling storage ad hoc — every vendor pairing differently with GPUs, every deployment a custom integration. STX turns the assembly into a certifiable blueprint, the storage analogue of what DGX reference designs did for compute. The strategic effect: **alignment becomes a procurement gate.** Vendors inside the blueprint (WEKA NeuralMesh, VAST, Cloudian, Hammerspace, IBM) sell "STX-ready"; vendors outside get relegated to the cold G4 tier of the memory hierarchy. Disambiguation the market needed: **CMX** is only the pod-level KV-cache tier (renamed from ICMSP); **STX** is the rack-scale superset; **AIDP** (AI Data Platform) is a parallel end-to-end RAG-pipeline reference design (RTX PRO GPUs + BlueField + NIM/NeMo Retriever/cuVS) — three constructs, not three names for one thing.

Primary use cases

AI-factory storage procurement against a certifiable blueprint, KV-cache tier deployments at rack scale, vendor storage-platform certification (NeuralMesh AIDP, VAST AI OS, Cloudian, Hammerspace, IBM Storage Scale), disaggregated inference infrastructure with pod-shared context memory.

Recent developments

Latest signals

The vendor race formed within one quarter of GTC. WEKA shipped NeuralMesh AIDP as an appliance; Hammerspace launched its own AIDP-reference platform; IBM published Storage Scale ECE + Fusion integration; Cloudian declared itself "all-in" on STX; VAST runs its CNode natively inside BlueField-4's 64-core Grace complex — currently the only storage software in the native DPU loop, on the sole CMX-spec enclosure (AIC F2032-G6). Per Hammerspace AIDP launch, IBM Redbook, and Cloudian — all-in on STX.