What is CMX (NVIDIA BlueField-4 CMX) ?

NVIDIA® CMX™ context memory storage is an AI‑native context tier for long‑context, multi‑turn, and agentic AI inference. NVIDIA BlueField-4 CMX is a purpose-built, accelerated storage platform designed to power the next generation of AI workloads, particularly those requiring extreme performance for large-scale data processing. It is built on the NVIDIA BlueField-4 Data Processing Unit (DPU) and integrates NVIDIA's Data Center Infrastructure-on-a-Chip Architecture (DOCA) software stack. CMX is engineered to deliver unprecedented levels of storage performance, efficiency, and scalability for AI factories, enabling faster data access, processing, and model training. It leverages advanced technologies like NVMe over Fabrics (NVMe-oF) and GPUDirect Storage to accelerate data movement directly to GPUs.

Why do you need it?

As AI models grow in complexity and data volumes explode, traditional storage architectures often become a bottleneck, hindering the speed and efficiency of AI training and inference. You need CMX to overcome these storage limitations and accelerate your AI pipelines. It's essential for:

Maximizing GPU Utilization: By eliminating storage bottlenecks, CMX ensures GPUs are fed with data at optimal speeds, preventing idle cycles and maximizing the return on investment in expensive GPU infrastructure.

Handling Massive Datasets: It provides the extreme throughput and low latency required to process petabytes of data for large-scale AI training.

Accelerating Time to Insight/Model Deployment: Faster data access and processing directly translate to quicker model training, iteration, and deployment of AI solutions.

Building Scalable AI Infrastructure: CMX offers a highly scalable and efficient storage foundation for building and expanding AI factories.

Benefits of CMX

NVIDIA BlueField-4 CMX offers significant benefits for AI-driven enterprises:

Extreme Performance: Delivers industry-leading storage performance, throughput, and ultra-low latency, crucial for demanding AI workloads.

Enhanced Efficiency: Optimizes data movement and processing, reducing the time and resources required for AI training and inference.

Scalability: Provides a highly scalable architecture that can grow with the increasing demands of AI data and models.

GPUDirect Storage Integration: Accelerates data transfer directly from storage to GPU memory, bypassing the CPU and system memory, which significantly reduces latency and increases bandwidth.

NVMe-oF Acceleration: Leverages NVMe over Fabrics to enable high-performance, low-latency access to shared storage resources across the network.

Simplified AI Infrastructure: Offers a unified and optimized platform for AI storage, streamlining deployment and management.

How is ASUS helpful?

While the NVIDIA CMX product page primarily focuses on NVIDIA's core technology and general ecosystem partners, specific information reveals the direct and significant contribution by ASUS to the CMX platform. ASUS offers a validated storage server, the 2U UF920 E3 RS24, which is specifically designed to leverage NVIDIA CMX capabilities. This ASUS solution is empowered by cutting-edge NVIDIA technologies, including the NVIDIA Vera CPU, NVIDIA BlueField-4 DPU, and ConnectX-9 SuperNIC. By connecting this ASUS CMX-enabled server with NVIDIA GPUs, it effectively optimizes end-to-end AI data pipelines, significantly reducing bottlenecks and maximizing GPU utilization for demanding AI workloads. Furthermore, ASUS enhances this offering through strategic software partnerships with WEKA and IBM, integrating advanced data management and intelligent services to deliver a comprehensive and high-performance AI storage solution.