At present, the Open Flash Platform (OFP) initiative was launched with inaugural members Hammerspace, the Linux group, Los Alamos Nationwide Laboratory, ScaleFlux, SK Hynix, and Xsight Programs. OCP intends to handle necessities from the following wave of knowledge storage for AI.
“The convergence of knowledge creation related to rising AI purposes coupled with limitations round energy availability, sizzling knowledge facilities, and knowledge middle area constraints means we have to take a blank-slate method to constructing AI knowledge infrastructure,” OFP mentioned in its announcement.
A decade in the past, NVMe unleashed flash because the efficiency tier by disintermediating legacy storage busses and controllers. Now, OFP unlocks flash because the capability tier by disintermediating storage servers and proprietary software program stacks. OFP leverages open requirements and open supply—particularly parallel NFS and commonplace Linux—to position flash instantly on the storage community. Open, standards-based options inevitably prevail. By delivering an order of magnitude better capability density, substantial energy financial savings and far decrease TCO, OFP accelerates that inevitability.
Present options are inherently tied to a storage server mannequin that calls for extreme assets to drive efficiency and functionality. Designs from all present all-flash distributors are usually not optimized to facilitate one of the best in flash density, and tie options to the working lifetime of a processor (sometimes 5 years) vs. the working lifetime of flash (sometimes eight years). These storage servers additionally introduce proprietary knowledge buildings that fragment knowledge environments by introducing new silos, leading to a proliferation of knowledge copies, and including licensing prices to each node.
OFP advocates an open, standards-based method which incorporates a number of components:
- Flash units – conceived round, however not restricted to, QLC flash for its density. Flash sourcing must be versatile to allow prospects to buy NAND from varied fabs, doubtlessly by means of controller companions or direct module design, avoiding single-vendor lock-in.
- IPUs/DPUs – have matured to some extent that they will exchange far more useful resource intensive processors for serving knowledge. Each decrease price and decrease energy necessities imply a way more environment friendly element for knowledge providers.
- OFP cartridge – a cartridge accommodates the entire important {hardware} to retailer and serve knowledge in a kind issue that’s optimized for low energy consumption and flash density.
- OFP trays – An OFP tray suits plenty of OFP cartridges and provides energy distribution and fitment for varied knowledge middle rack designs.
- Linux Working System – OFP makes use of commonplace Linux working inventory NFS to produce knowledge providers from every cartridge.
“Our targets are usually not modest and there’s a lot of labor in retailer, however by leveraging open designs and trade commonplace parts as a group, this initiative will end in huge enhancements in knowledge storage effectivity,” OFP mentioned.
“Energy effectivity isn’t non-compulsory; it’s the one strategy to scale AI. Interval. The Open Flash Platform removes the shackles of legacy storage, making it attainable to retailer exabytes utilizing lower than 50 kilowatts, vs yesterday’s megawatts. That’s not incremental, it’s radical,” mentioned Hao Zhong, CEO & Co-Founding father of ScaleFlux.
“Agility is all the pieces for AI — and solely open, standards-based storage retains you free to adapt quick, management prices, and decrease energy use,” mentioned Gary Grider, Director of HPC, Los Alamos Nationwide Lab.
“Flash would be the subsequent driving pressure for the AI period. To unleash its full potential, storage programs should evolve. We imagine that open and standards-based architectures like OFP can maximize the complete potential of flash-based storage programs by considerably bettering energy effectivity and eradicating boundaries to scale,” mentioned Hoshik Kim, SVP, Head of Reminiscence Programs Analysis at SK hynix.
“Open, standards-based options inevitably prevail. By delivering 10x better capability density and a 50 p.c decrease TCO, OFP accelerates that inevitability,” mentioned David Flynn, Founder and CEO, Hammerspace.
