NVIDIA BlueField-4 STX Drops to One Vera CPU for CMX Storage

San Jose, Oct. 19 — NVIDIA partners showed a smaller BlueField-4 STX board for the company's CMX context memory storage platform at the OCP 2026 summit, built around one Vera CPU instead of the two on NVIDIA's first STX design, ServeTheHome reported.

The Technical Details

CMX is NVIDIA's storage tier for KV cache, the working memory a language model keeps for each token of a conversation so it does not have to recompute it. NVIDIA launched the STX reference architecture at GTC in March and said partner systems would arrive in the second half of 2026.

The first BlueField-4 STX processor NVIDIA showed paired two Vera CPUs with four ConnectX-9 SuperNICs, network cards built for AI clusters. ServeTheHome said that layout behaves more like a dual-socket server than a classic DPU, or data processing unit, the small offload card that usually sits beside a server's main processor.

The new card carries a single Vera package, SOCAMM memory (a compact, swappable LPDDR module format), an M.2 slot and dense connectors for PCIe and system links. ServeTheHome did not see a ConnectX-9 SuperNIC on any of the boards it photographed on the show floor.

• CPU: 1x Vera on the OCP 2026 CMX board vs. 2x Vera on NVIDIA's first BlueField-4 STX design
• First STX design network bandwidth: 3.2 Tb/s vs. 800 Gb/s for the standard BlueField-4 DPU
• First STX design memory: 384 GB LPDDR5X plus a 2 TB M.2 SSD
• First STX design host interface: PCIe Gen 6, 96 lanes
• Standard BlueField-4 DPU: Grace CPU with 64 Arm Neoverse V2 cores, up to 128 GB LPDDR5X
• ConnectX-9 SuperNIC on the OCP 2026 boards: not seen
• NVIDIA claim, CMX: up to 5x tokens per second vs. traditional storage
• NVIDIA claim, STX: up to 4x energy efficiency vs. traditional CPU architectures for storage
• Partner availability: second half of 2026

For scale, the 3.2 Tb/s listed for the dual-Vera STX configuration is four times the 800 Gb/s of a standard BlueField-4 DPU. That gap shows how far STX already sits from the offload cards it shares a name with.

Voltgrid Analysis: Infrastructure Implications

A cheaper rung below HBM

CMX sits between GPU memory and bulk storage. Every block of context that spills from HBM to flash frees costly accelerator memory, which is why NVIDIA sells CMX on tokens per second rather than raw capacity. Our earlier report on NVHBM base dies shows how much engineering NVIDIA puts into the top of that stack.

Our read: halving the CPU count lowers the price of each context node, and that matters more than peak throughput for a tier operators will buy by the rack. More of our NVIDIA coverage tracks how Vera spreads beyond compute trays.

Networking is the open question

The missing ConnectX-9 leaves bandwidth per node unclear. If the network moves off the card, builders gain freedom to mix NICs but lose the single-board design NVIDIA pitched in March. Optics choices, like the Cisco and NVIDIA uSFP modules we covered, then shape the storage fabric as much as the board does. See our interconnect hub for the wider picture.

Thermal and floor space

ServeTheHome noted that the SOCAMM layout makes it easy to fit either a heatsink or a liquid cold plate. That flexibility lets storage racks stay air-cooled in older halls while GPU racks move to liquid.

Voltgrid Verdict

Looking Ahead

NVIDIA has promised partner STX systems in the second half of 2026, so shipping configurations should surface before year-end. Watch for whether those systems carry ConnectX-9 on the board or rely on a separate NIC.

Source: ServeTheHome, NVIDIA.