Everpure dives deeper into AI

Everpure is triple-boosting its AI credentials at Nvidia’s GTC event with full AI data pipeline automation; benchmark proofs of FlashBlade AI hardware performance; and SLA-driven, public cloud-like, AI storage capacity provisioning.

The AI pipeline automation is based around its Data Stream technology, a software suite running on FlashBlade and Nvidia Blackwell GPU hardware. Data Stream automates and accelerates the ingestion, transformation, and optimization of data for enterprise AI pipelines.There are benchmarks for Everpure’s FlashBlade//EXA product, and the Evergreen//1 offering is being extended to Everpure’s AI product portfolio.

Everpure says there is no one single AI workload, and traditional, dual-controller style storage can’t cope. With AI training, inference, and agent use increasing, the AI storage infrastructure has to have mission-critical reliability and HPC-class performance across each stage in the AI pipeline: across file, block, and object data, covering both large sequential and small random read and write performance, and everything in between.

Kaycee Lai, Everpure VP for AI and Analytics, tells us: “Successful companies really treat AI as a tier one application, they make sure that you've got production grade, availability, reliability, performance, uptime. You make sure there's governance in there just like you would if it was your Oracle environment, your SAP environment.”

It’s developed an Enterprise Data Cloud (EDC) for AI to provide this and today’s Evergreen//1, Data Stream and FlashBlade//EXA benchmarks each play a role in it.

The Evergreen//One option arrived in 2018 and related to on-premises storage. An Everpure array’s actual ownership was retained by Pure with customers paying – subscribing – for consumption of its services; paying for the storage they use, as with public cloud storage services. This changed the customer’s focus away from specific hardware as they paid for set levels of storage services, not for a specific array model or array configuration.

They can scale storage capacity up or down flexibly based on changing needs without over-provisioning. Everpure delivers extra capacity to the array as needed if the customer’s workload increases its storage demands.

Everpure originally provided its all-flash arrays in classic dual controller style but realized that the growing need for combination of extreme performance and extreme capacity needed a parallel filesystem-type approach with scalable metadata storage controllers linked to scale-out all-flash storage nodes and separate control and data paths; a disaggregated approach coming from the Lustre-influenced HPC world. 

The resulting FlashBlade//EXA used existing FlashBlade array technology for the metadata controllers and NVME fabric-connected JBOFs for the data nodes. EXA’s performance credentials are now being published, with a SPECStorage Solutions run and now ML Perf 2.0 results, with EXA scoring twice as high as its nearest competitor.

FlashBlade//EXA is Nvidia-certified with NCP Certification coming later this year. There is a co-engineered Everpure-Supermicro AIDP configuration “narrowing the barrier to entry to organizations just starting their AI journey.”

Everpure has also released an IO500 result for its FlashBlade//S500, designed before EXA; a unified file and object array with disaggregated compute (controller) and storage node features. It scored 7.2 million IOPS and is cited as an additional EXA performance credential, not as a pure IO500 test.

Lai is enthusiastic about EXA’s AI performance: “On the preparation side, RAG pipelines can require 300,000 to a million IOPS each time you run it. Great. Flashblade delivers 400 million, 450 million IOPS. On the training, you require multiple 100 gig per second throughput at scale. Great. Flashblade delivers 220 gigabytes per second. And then the metric we're really proud of - on the inference side, where you go to inference in production, you need 100,000 to a million metadata operations per second. EXA delivers mind-blowing 4.6 billion metadata operations per second.”

The Data Stream offering supports Nvidia’s AI Data Platform, and Lai emphasized Everpure;s closeness to Nvidia: “We're partnering very, very closely with Nvidia. I would say in the last year, we've really re-ignited that partnership. So we're now seen as a tier-one storage partner with Nvidia. We're included in all their marketing launches and we're keeping up with all the certifications.”

Everpure says it delivers more than 10 TBps read performance, surpasses its closest competitors in ML Perf 2.0 and SPECStorage Solution 2020 benchmarks, supports more than 10,000 GPUs, and can decrease total cost of ownership by up to to 58 percent compared to its competition – which we understand to mean Dell, DDN, Hammerspace, HPE, Hitachi Vantara, IBM, NetApp, VAST Data, and WEKA.

The DataStream offering is in beta test mode.

 Everpure has neocloud customer wins for EXA in the AI space with beyond.pl, options and STN. 

Lai tells us: “AI is the fastest growing business within Everpure now. So last quarter alone, it was a massive triple-digit growth year-on-year, where customers are buying Everpure specifically for AI first, which is great. And of course, we still have our existing customers who are extending their platform into AI with us, so that will always be there. It's a great one-two punch.”