DDN targets enterprise-shaped gap in its AI storage supply
The substitute intelligence (AI) increase has but to essentially break and, when it does, enterprises will spend billions on it, based on storage array maker DDN, which has long-standing experience in delivering storage to help huge compute clusters in excessive efficiency computing (HPC) and, extra just lately, AI.
HPC compute clusters already depending on DDN storage will seemingly triple or quadruple orders by 2027, stated Paul Bloch, co-founder and president of DDN, including: “200,000 GPUs are deployed in [Elon Musk’s] xAI cluster in Memphis, and that quantity will attain 1,000,000. The expansion of efficiency of GPUs is more and more driving information storage necessities, as a result of the one technique to take care of it’s more and more to ingest information in actual time.”
Throughout a current assembly with sister publication LeMagIT as a part of the IT Press Tour in Silicon Valley, DNN CTO Sven Oehme famous that the enterprise market itself can be promising.
“AI hasn’t but achieved its potential, however enterprises can see the seemingly evolution and what’s doable,” stated Oehme. “The increase will actually come when enterprises totally launch into inference and exploit AI in all sectors of exercise.”
DDN is value $5bn and can realise a turnover of $1bn in 2025. In the beginning of this 12 months, the enterprise secured $300m in funding from Blackstone in return for six% of capital and a seat on the board.
In response to Bloch, Blackstone’s help will assist the corporate to seek out clients outdoors its conventional shopper base: “AI funding within the order of $1bn to $2bn will happen by corporations whose names we don’t even know of but.
“These selections can be taken by common administration or information administration round potential financial savings round AI, particularly in industrial functions. In that context, Blackstone isn’t simply an investor, however extra like a accomplice to whom we open our contact e book.”
A extra strong EXAscaler for enterprise clients
To realize its goals, DDN up to date its core product line. After the EXAscaler AI400X2 Turbo was launched final 12 months, it then introduced AI400X3, which comes with 24 NVMe SSDs and help for added SSD cabinets through NVMe-overTCP.
Efficiency was improved by 55% in reads and greater than 70% for writes, with throughput of 140GBps (learn) and between 75GBps and 100GBps (write). That velocity achieve comes from use of 4 Nvidia BlueField playing cards that use that vendor’s SpectrumX protocol.
“We’ve already obtained a reference structure for his or her new Blackwell GPU,” stated Oehme. “Elsewhere, all testing and growth for Blackwell is going down with DDN tools.”
Different enhancements have been in safety and pooling. The latter addresses the wants of MSPs for cloud providers that must dynamically provision components of the array for various and sometimes quite a few purchasers.
From the standpoint of a brand new buyer base amongst enterprises, the EXAscaler AI400X3 brings information compression that provides useable capability a lot better than its uncooked capability, as is already the case with suppliers equivalent to Dell, HPE or NetApp.
“And now we now have significantly improved its resilience,” stated Oehme. “We will now exchange any {hardware} or software program element with out stopping the array. The chance of downtime is decreased to a minimal. That’s elementary as a result of enterprises can’t entertain any disruption. That differs from HPC clients preferring, in case of incident, to diagnose the issue themselves earlier than coming to us.”
Infinia 2.1: Object storage accelerated for AI
Whereas the EXAscaler AI400 shares its contents in file mode, the route of enterprise AI customers is in the direction of object storage. That’s much less speedy when it comes to entry however gives capability and choices to label or filter contents whether or not for re-training or for personal information in inference through RAG. It’s for that want that DDN launched its object storage product Infinia in 2024.
Infinia has now reached model 2.1, for which DDN claims velocity multiplied by 100x. Right here it competes with different object storage options which have additionally just lately accelerated efficiency to help GPUs, equivalent to HyperStore from Cloudian or Ring from Scality.
DDN contrasts its supply to AWS’s S3 Specific, which permits enterprises to hold out AI duties in its cloud. Infinia 2.1 can be 10x sooner when it comes to entry time and 25x sooner when it comes to requests.
The DDN product has connectors to most AI software program stacks, equivalent to TensorFlow, Pytorch and Apache Spark, plus NIM modules from Nvidia’s AI Enterprise suite. All that was current in model 2.0. The two.1 replace added integrations with observability platforms Datadog and Chronosphere, and with those who relaxation on the open supply OpenTelemetry customary. It could possibly additionally now connect with Hadoop-based information lakes.
Infinia 2.1 comes on DDN {hardware} or can be utilized through digital occasion on Google Cloud Platform.
DDN additionally gives two different {hardware} merchandise, xFusionAI and Inferno – xFusionAI is an EXAscaler AI400X3 that runs the Lustre file system in addition to Infinia 2.1, so it may possibly entry the identical information through S3 object storage, which brings the benefit of having the ability to entry metadata that indexes contents.
Inferno supplies a community proxy put in between the Infinia equipment and servers to speed up communication between object storage and Lustre file storage. The gadget has a change based mostly on Nvidia BlueField, which makes use of its SpectrumX protocol and GPUdirect for direct communication with GPUs, plus NVMe SSD cache.
11,000 clients, 4,000 in AI
DDN stated that it has deployed storage that helps 700,000 GPUs at 7,000 clients, of whom 4,000 are engaged in AI workloads. The corporate employs 1,100 individuals, of which 600 are engineers, with 200 to 250 extra hires deliberate for this 12 months.
At its origins in 1998, DDN equipped storage arrays for the few supercomputer centres then in existence. Its key know-how comes from feeding quite a few compute nodes in parallel, and that’s what satisfied Nvidia to accomplice with it since 2010.
“Think about 10,000 GPUs,” stated Bloch. “For those who wished to get 1 GBps on reads per GPU, you’d want throughput of 10TBps for the entire array. There are few techniques on this planet able to offering 10TBps on reads and on writes. That’s what we now have mastered.”
