WEKA claims Nvidia CMX support plays to its strengths

As the dust settles on Nvidia GTC 2026 news about its CMX KV cache extension to external RDMA-connected SSDs, we considered how it might impact WEKA. The company says it's good news.

That’s not what we initially thought. We expected that suppliers like WEKA and Hammerspace, who had placed bets on using local SSDs in GPU servers, with Augmented Memory Grid software from WEKA and Tier 0 software from Hammerspace, could be overtaken or bypassed by CMX-supporting competitors. Before the CMX/STX scheme, WEKA and Hammerspace both had an advantage as no other supplier could feed data as fast into these SSDs – they would have had to write specific software to do it.

Now that the CMX/STX reference architecture is here, we thought WEKA would have to alter its software to write data to the CMX SSDs rather than local SSDs in GPU servers. Other Nvidia storage partners now have the framework to do this as well, leveling the playing field and allowing them to catch up with WEKA and Hammerspace, for example, Everpure and VAST Data. 

Valentin Bercovici, WEKA's Chief AI Officer, argued that WEKA would not be disadvantaged. He told us: “WEKA doesn't have to alter our software for STX/CMX.”

That’s because both the WEKA NeuralMesh client and server software modules have supported Arm-based Grace servers from Nvidia since they became available on GB200 and GB300 systems.

On Grace, WEKA has 97 percent of line-rate CX-7 400GbE performance under real-world mixed I/O workloads, and will soon demonstrate the same with CX-8/CX-9 800GbE on Vera.

BlueField-4 (BF-4) Vera-based DPUs will have roughly double the core count, enabling even better performance for the three categories of container modules in NeuralMesh:

• "Drive" queue depth management
• "Compute" NVMe Fabric metadata and lock sync
• "Front-end" protocols (NFS, S3, SMB)

WEKA's AXON is its NeuralMesh filesystem software running as containerized microservices on Nvidia's GPU servers. Bercovici said: "CMX-based systems on new BF-4 initiators/clients and targets/servers, will not only be upward compatible with AXON on GB300s today, they will provide NeuralMesh even more elbow room to stretch WEKA's performance at higher real-world levels, under a broader set of Vera-Rubin configurations (Kyber, Oberon, Ultra, HGX, MGX)."

"This will be important as cluster-wide KV Cache routing and scheduling scales with upcoming LLM architectural evolution – like Hybrid DiT, Mamba and JEPA architectures – as well as DeepSeek 4-style engrams (conditional memory layers) which are anticipated this year."

He argued: "Even today's Claude Mythos preview with 10 trillion parameters will ultimately reveal new CMX requirements, which the flexibility of NeuralMesh under load at scale will benefit frontier labs pushing the boundaries of GPU memory requirements."

And here's the kicker: "To put it more simply – CMX will be one of our hardware platform offerings for the next version of WEKApod."