Using Computational Storage to Solve the Issues with Embedded Computing

The last blog in our series discussed the difficulties of providing computing resources in forward-deployed environments. Between power consumption, physical footprint, and environmental hardening, utilizing standard server and storage technology for petabyte-scale real-time applications (which are rampant in forward-deployed defense and intelligence theatres) is very problematic. However, these are problems that computational storage can easily solve, improving performance while reducing physical and power footprints.

If you don’t know what computational storage is, let me give you a quick summary (we also have a great video on the subject on YouTube and on our website). Computational storage embeds processing and application acceleration capabilities inside storage devices; in our case, we embed these capabilities inside of our NVMe Express (NVMe) solid state drives (SSDs). The advantage of doing so is threefold. First, computational storage eliminates the need to move massive data sets from storage systems and devices into server memory, significantly reducing the time to process petabyte-scale data sets. Secondly, computational storage reduces the number of servers required for these problem; in most cases, a single server can provision applications across multiple computational storage SSDs. This significantly reduces both physical footprint and power/cooling requirements. Finally, computational storage provides the ability to scale compute power linearly with storage capacity, while maintaining the same level of performance.

What does this mean for a workload with a petabyte-size dataset? Here is the comparison of a 2U two-processor server with 32 x 32TB computational storage SDDs versus a cluster of 8 x 1U  two-processor servers, each with 4 x 32TB standard SSDs running the Facebook Artificial Intelligence Similarity Search (FAISS) workload with a total of 24M records (power assumes both types of SSDs consume 12W per SSD):

Configuration Physical Footprint Power Consumed Performance on FAISS Application
1 x 2U/2P Server with Computational Storage 2U 792 W 0.53 seconds
8 x 1U Servers with standard SSDs 8U 4,000 W 391 seconds (98% of which is load time)

As can be seen, the computational storage configuration took 25% of the real estate and 20% of the power of the standard server configuration, while providing performance that is over 700 times that of the standard server configuration. While the acceleration provided by computational storage does depend on the application, the advantages of computational storage for space and power-constrained deployment scenarios is clear even if no performance acceleration is provided. To find out more about how you can help your tactical users, please contact me at