TECHunplugged analysts Max Mortillaro and Arjan Timmerman were recently invited to attend Pure //Accelerate 2024 in Las Vegas, from 17-Jun-24 to 20-Jun-24. This article will cover AI-related announcements (except the AI Copilot, which was covered in our previous article).
TECHunplugged also covered the announcements in two TECHunplugged Take videos viewable at the end of this article.
Preamble: Pure Storage’s Vision of AI
Before starting, it makes sense to briefly recall some of the statements made by Pure Storage CEO Charlie Giancarlo during the FY25 Q1 Earnings Call (end of May 2024), quoting the excellent article written by an authority in enterprise IT journalism, Mr. Chris Mellor:
Giancarlo sees three other opportunities ahead: “The recent advances in AI have opened up multiple opportunities for Pure in several market segments. Of greatest interest to the media and financial analysts has been the high-performance data storage market for large public or private GPU farms. A second opportunity is providing specialized storage for enterprise inference engine or RAG environments. The third opportunity, which we believe to be the largest in the long term, is upgrading all enterprise storage to perform as a storage cloud, simplifying data access and management, and eliminating data silos, enabling easier data access for AI.”
At this time, TECHunplugged commented as follows:
“It’s reassuring to see that despite the overhype around AI and LLMs (albeit real opportunities exist), Pure Storage is lucid on the long run that the challenges lie more with data management, improving data quality, and turning storage into a seamless enabler rather than assuming hardware only will solve the problem.”
This should set up the proper context about where the company is aiming at. Some of those aspects (notably data management or, as Charlie Giancarlo stated during the press conference, dataset management) is being partially handled by Pure Storage Fusion and is a topic where the company will dive deeper. Some tangible announcements were made during the event:
- Evergreen//One for AI
- Secure Application Workspaces
- NVIDIA SuperPod Certification
Evergreen//One for AI: The Spice Must Flow!
Pure Storage assesses today’s AI environment as very fast-paced (they for example rewrote their AI Copilot three times and changed the security framework twice within 6 months of development), with complex workflows and unclear use cases. On top of that, many organizations rely on fragmented data storage environments with inflexible storage architectures, and no way to resize or upgrade performance without risks or complexity.
Pure intends to address this with EverGreen//One for AI, an AI-optimized addition to its Evergreen//One Storage-as-a-Service (STaaS) offering catalog. As with every other Evergreen//One offering, this service removes the management complexities and allows organizations to focus on delivering business outcomes, with one important differentiator.
Evergreen//One tiers usually focus on capacity and performance but when it comes to AI, another parameter comes into play: bandwidth. GPUs are expensive, and organizations don’t want to have GPU sitting idle when data throughput becomes the bottleneck, hence a new offering which delivers guaranteed performance based on maximum bandwidth requirements for GPUs. The service is charged on a GB/s bandwidth base rate, with a marginal rate also used for capacity usage. In other words: the spice (data) must flow!
Worth noting, the solution is certified for NVIDIA OVX servers, DGX BasePOD, and will be certified for DGX SuperPOD once available (see below).
Also worth mentioning Pure Storage introduces a new high-performance SLA enhancement: Evergreen//One customers on UBF Ultra and Premium tiers gain a 50% increase in performance, unrelated to the Evergreen//One for AI offering.
Secure Application Workspaces
One of the other key announcements of Pure //Accelerate 2024 is Secure Application Workspaces, a technology based on Portworx that will combine Kubernetes container management, secure multi-tenancy, and policy governance tools to enable advanced data integrations between enterprise mission-critical application data AI clusters.
This technology integrates Portworx with the Purity OS API on Flasharray systems to ensure workloads are deployed to specific, isolated tenants with fine-grained access controls. Currently, storage administrators will create the tenants, but a future release will allow to request tenant creation via a self-service portal.
Secure Application Workspaces is currently in testing with select customers and is expected to be Generally Available in August 2024.
NVIDIA SuperPOD Certification
Finally, Pure Storage announced the expected completion of its NVIDIA SuperPOD Certification by the second half of 2024’s calendar year, bringing a fully validated design with performance guarantees to its customers. DGX SuperPOD deployments are the absolute endgame in terms of performance, integration, and speed of deployment for no-compromise large-scale AI training and inference.
TECHunplugged’s Opinion
Perhaps the most important announcement when it comes to AI is the launch of Evergreen//One for AI, a purpose-built STaaS offering with AI in mind that guaranteed storage performance and throughput for GPUs to support training, inference, and HPC workloads. It als introduces a new way of charging for the service, ensuring that GPUs are kept busy and regularly fed with data while making sure organizations are only charged the right price (no need to purchase the highest storage performance tier to achieve the desired throughput).
The NVIDIA DGX SuperPOD certification is also relevant: although the number of SuperPODs deployed may be significantly lower than any other solutions (for example AIRI or DGX BasePOD), this isn’t only a matter of prestige or “me too”. Pure already works with customers in the AI field and there are real opportunities to be explored here. The massive current and future capacity improvements (new 150 TB DFM modules, upcoming 300 TB modules and a roadmap beyond), are shifting the balance in favor of Pure Storage in the company’s discussions with hyperscalers.
They key differentiators here for Pure Storage revolve around its solution’s power efficiency, which derives from the current and future storage densities that will be achieved by the solutions. When the DFM module capacity doubles, the effective storage capacity of the array also doubles, and the power efficiency also increases twofold since there are practically no changes to the power consumption of an array, whether fully populated by 75 TB, 150 TB, or 300 TB DFMs. With each DFM generation, organizations get twice the previous storage capacity packed in the same RU size, and twice the capacity for the same power draw. These linear and predictable improvements will lead to massive storage consolidation in the datacenters, freeing up rack space and power for hungry GPU systems.
Additional content
TECHunplugged published the following video content during the event: