A single, pre-integrated end-to-end solution for deep learning


Cisco recently expanded its UCS® product portfolio with the Cisco UCS C480ML M5, specifically designed for compute-intensive AI and ML workloads.

We’re excited to present FlashStack for AI, a new Cisco Validated Solution that combines the storage scalability and simplicity of Pure Storage® FlashBlade™ servers and the performance of the new Cisco UCS C480 ML M5 platform. Like other FlashStack™ solutions, FlashStack for AI will help IT teams reduce infrastructure management pain and enable cutting-edge development from their application teams.




Cisco Validated Design

The C480 ML M5 server brings the same highly-scalable network simplicity to AI workloads that UCS customers already know and love. By incorporating the NVIDIA NVLink architecture into their UCS form factor, the C480 ML M5 is able to provide highly-optimized GPU-to-GPU communication fronted by a simple, scalable network uplink.

Pairing the C480 ML M5 with a FlashBlade, IT teams can help data science teams seamlessly grow from experimental mode to full-scale, revenue-impacting deep learning at scale.


Prepared to Scale

We put this solution to the test with end-to-end deep learning training workloads that resemble our customers’ real-world workloads. We optimized training throughput when reading datasets from Pure Storage FlashBlade into C480MLs and ensured that the solution is able to provide linearly scaling performance.

End-to-end tests showed that:

  1. FlashStack for AI is able to provide linearly scaling performance. As C480 ML compute servers or FlashBlade storage Blades are added, there’s minimal scale-out networking overhead. AI teams will be able to grow their environments as needed and depend on scalable networking performance that stays simple to manage.
  2. Throughput is nearly identical between training on remote datasets saved to FlashBlade and training on synthetic data (generated right on the GPUs). The input pipeline from storage → CPU → GPU is near-invisible, even at multi-node scale.


FlashStack for AI is an integral component for an AI pipeline that provides the massive data throughput required by AI and ML applications all the way from storage into GPU.

Now, teams can ingest, ETL, train, and deploy with a wide range of Cisco UCS servers combined with a FlashBlade Data Hub, delivering a single platform powerful enough to efficiently share data across the entire pipeline. 


To learn more, come talk to us at GTC 2019 (booth #1217) or visit www.flashstack.com