Transform GPU Computing Infrastructure by Unlocking a New Tier 0 of Ultra-Fast Shared Storage

Organizations are deploying large GPU computing environments for AI, enterprise HPC, and other forms of unstructured data processing. Typically, these same organizations also deploy external flash storage systems to provide the data to these GPUs, and these storage systems add cost, consume power, and take time to evaluate, purchase and deploy.

Meanwhile, the flash storage in GPU servers typically goes unused for a few reasons:

• It is siloed - only available to the GPUs in the server and not other GPUs in the cluster
• It is not protected - so there is risk of data loss if the GPU server goes down
• It is difficult for data users to manually move data between local storage and shared storage

And the capacities are no longer small. A single GPU server today can have 100s of Terabytes of data, and by next year a single server could have 2 Petabytes of capacity.

So organizations today have Petabytes of available storage in their GPU cluster that ends up as ‘stranded’ capacity. And now, for the first time, Hammerspace allows you to use that ultra-high performance capacity in new ways by enabling the local storage in the GPU servers to become part of the Hammerspace Global Data Platform.

This creates a new tier of ultra-fast shared storage that we call “Tier 0.”