What is Computational Storage?

Computational storage is a storage architecture that integrates processing power directly into storage devices, allowing data to be analyzed where it resides rather than being transferred to a central processor. This reduces latency, minimizes bandwidth usage, and improves performance in data-intensive applications such as IoT, AI/ML, and edge computing.

Why is it Important?

  • Performance and scalability: Handles massive datasets without saturating system memory or network links.
  • Lower latency: Critical for real-time decision-making in edge and IoT environments.
  • Energy efficiency: Less data transfer means reduced power consumption.
  • Security and privacy: Keeps sensitive data on-device, lowering exposure risks.
  • Application breadth: Powers use cases from in-storage AI model inference to video transcoding and industrial sensor analytics.

How Does It Work?

In traditional systems, raw data is moved from storage to a CPU for processing, then back to storage, often multiple times due to memory limits.


With computational storage:


  1. An operation is requested, not the full dataset.
  2. The storage device processes data locally using its onboard compute resources.
  3. Only the processed output is sent to the host system.

This local processing reduces I/O bottlenecks, conserves bandwidth, and accelerates analysis.

What Are the Key Components or Features?

  • On-device processors: Multicore CPUs, FPGAs, or ASICs built into the storage device to perform computations.
  • Reduced data movement: Only processed results or relevant data subsets are sent to the host system.
  • Real-time data processing: Supports immediate analysis for time-critical workloads.
  • Device types:
    • Computational storage drives (CSDs): Storage devices with built-in compute capabilities.
    • Computational storage processors (CSPs): Compute engines that work alongside storage devices.
    • Computational storage arrays (CSAs): Multi-device systems designed for large-scale, in-place data processing.

    FAQs

    Is computational storage only for SSDs?

    No. While most implementations use SSDs, computational storage can be integrated into other non-volatile storage systems.

    How does it benefit AI/ML workloads?

    By processing large datasets locally, it speeds up training and inference while reducing infrastructure strain.

    Relevant Resources

    Related Topics

    • Edge Computing: Understand the differences between edge and cloud computing.
    • IoT Security: Key concepts and practices for securing IoT devices and data.
    • Machine Learning: Explore Arm’s role in powering AI and ML solutions.