BUSINESS WHITE PAPER
The Pure Storage
Platform for AI
Pure Storage® accelerates and simplifies AI deployments,
enhancing their value to the enterprise.
BUSINESS WHITE PAPER
With readily available generative artificial intelligence (GenAI), AI has
“The playing field is poised to
become a sine qua non of information technology (IT) operations.
Enterprises in finance, medicine, manufacturing, transportation, security, become a lot more competitive,
and others all realize that AI is now a survival issue for them. Those that and businesses that don’t
use AI to identify trends, make accurate predictions, serve clients faster deploy AI and data to help them
with less effort, and so forth have distinct competitive advantages over innovate in everything they do
those that don’t.
will be at a disadvantage.”
The increasing importance of AI-based solutions makes reliable, PAUL DAUGHERTY, CHIEF TECHNOLOGY
AND INNOVATION OFFICER, ACCENTURE
easy-to-use IT services a must for production deployments. This brief QUOTED IN: HTTPS://WWW.SALESFORCE.COM/
surveys AI needs throughout the project lifecycle (primarily from a BLOG/AI-QUOTES/
storage perspective) and shows how the Pure Storage® portfolio of
storage system, data services, and workflow management products for
Kubernetes promote efficiency both for AI and IT infrastructure teams as
well as developers and MLOps engineers who design, implement, and run
AI applications.
The AI Project Lifecycle
Organizations undertake AI projects to support mission objectives such as,
• More accurate medical diagnoses
• Acceleration of genomic research
• More predictable market fluctuations
• Bank card fraud detection
• Rapid identification of security threats
• etc.
Whatever their objectives, in-house AI projects tend to follow a trajectory
similar to that shown in Figure 1, from conception, development, production,
to evolution. FIGURE 1 AI Project Life Cycle
AI projects generally start with a proposed model (algorithm) and train (refine) it iteratively using steadily increasing amounts
of available or easily acquired input data for which outcomes are known until it reliably produces inferences (outcomes) that
support the mission objective. For example, a medical diagnosis model might be trained using thousands of MRI scans with known
diagnoses. The finished model would then take live scans as input and suggest diagnoses to medical practitioners.
Models are typically envisioned by small groups of data scientists, often in functional or business organizations rather than in
IT teams. Data scientists start with modest IT resources (e.g., public cloud virtual machines and storage) to experiment with
model variations.
Uncomplicate Data Storage, Forever 2
BUSINESS WHITE PAPER
As model refinement progresses, development is typically taken over by
Machine Learning Operations (MLOps) engineering organizations which:
As projects progress, the
intellectual property embodied • Assume on-going model training responsibilities from
in models and the curated data scientists.
training data become increasingly
• Manage the transition of project resources and collateral to IT
valuable. Along with performance
environments with the scalability, performance, and reliability
and reliability, protection against
needed for late-stage training and production.
data loss and theft of intellectual
• Introduce automated workflows that enable self-service training f
property become important
or data scientists and other developers, and ultimately support
considerations.
robust production.
When models produce reliable inferences, they move into production. Production models use live data to produce inferences that
assist with business decisions. As with any mission-critical IT application, both AI models and the environments in which they run
must be stable. Reliable access to a model and the data it needs to function is key to stability.
Models in production are usually monitored for “data drift” to ensure that results continue to meet mission objectives. With GenAI,
open source or proprietary large language models (LLMs) are often combined with retrieval augmented generation (RAG) using
context-specific vector databases that are updated frequently to include new data and remove what has become less relevant.
In some cases, AI models may be completely reimagined as business needs change and/or if new types of training data
become available.
Information Technology in AI Projects
While AI projects typically begin using modest in-house or public cloud IT resources, most are destined for eventual production.
Planning for production computing, storage, and software needs, and designing production workflows early in the development
process minimizes mid-stream infrastructure and procedure changes and accelerates return on investment.
Planning for production is particularly important with storage. Ideally, an AI storage infrastructure should provide:
• Non-disruptive expansion to meet rapidly growing data and changing I/O needs.
• Seamless data sharing among developers, training jobs, and production.
• Decade-long “24 365” duty cycles.
• Security to protect intellectual property from intrusion and theft.
Uncomplicate Data Storage, Forever 3
BUSINESS WHITE PAPER
How AI Projects Utilize IT
From an information technology perspective, AI projects can be thought of as consisting of two general types of tasks:
Data Curation
The best models are trained using input data from multiple disparate sources—event records, documents, images, sensor
readings, etc. Data curation is a blanket term for the acquisition, storage, and pre-processing of data for use in model training, and
later in production.
Projects usually preserve raw input data to avoid the time and cost of recreating or re-acquiring it. In most cases, it must be
curated (preprocessed) for model training. Data must be anonymized; timelines, measurement units, graphics resolutions, and so
forth must be reconciled; and items must be transformed into the file or object formats that training and production tools require.
Most projects preserve both raw and curated data.
What Data Curation Means for Storage Infrastructure
A large project might use petabytes of curated training data. Its storage infrastructure must be able
to scale from a few hundred gigabytes of files to petabytes contained in billions of files
and/or objects.
Raw data is used as input to curation but is then largely idle until a model is reimagined. Storage
used for it should be low-cost, but highly scalable.
Data is typically curated by processing large batches of raw data sequentially, but items are
accessed randomly during training and production. Flexible systems that perform well with both
sequential, random, and mixed access I/O workloads are key.
Training and Production
The fundamental assumption of AI is that iterative training with large amounts of input data with known outcomes converges on
a model that reliably makes useful inferences when presented with input1 for which outcomes are unknown. In a medical imaging
scenario for example, thousands of images with known diagnoses might be used to train a model that, when presented with new
images, would propose diagnoses used to advise physicians. Subject matter-specific applications often combine application-
specific data sets with readily available general-purpose large language models (LLMs) and use retrieval-augmented generation
(RAG) to satisfy natural language queries in their subject matter area.
As training progresses, the amount of curated data and the intensity with which it is used increase rapidly. In addition, the data,
files, and other collateral that comprise the evolving model become increasingly valuable.
When a completed model moves into production, its I/O needs change from the very intensive demands of a relatively small
number of training jobs to the agility required to service thousands of concurrent client transactions making many unrelated
I/O requests.
Uncomplicate Data Storage, Forever 4
BUSINESS WHITE PAPER
Finally, it is common to preserve production inputs and the corresponding output inferences for retraining and other techniques
that adapt models to changing conditions.
What Training and Production Mean for Storage Infrastructure
Storage for curated data must be scalable, both in capacity and performance, and must be easily
sharable by many concurrent training jobs. As the number of concurrent training jobs grows,
automated scheduling and data sharing are musts.
The business value of a production model requires storage with production-grade reliability,
performance, and administrative simplicity. The value of the intellectual capital embodied in it
and its training and production data requires robust security for “data at rest.”
In Summary: The Ideal Storage for an AI Infrastructure
Perhaps the most important attribute for AI project storage is agility—the ability to grow from a few hundred gigabytes to
petabytes, to perform well with rapidly changing mixed workloads, to serve data to training and production clients simultaneously
throughout a project’s life, and to support the data models used by project tools. The attributes of an ideal AI storage solution are:
Performance Agility Space and Power Efficiency
• I/O performance that scales with capacity. • Low space and power requirements that free data center
resources for power-hungry computation.
• Rapid manipulation of billions of items, e.g.,
for randomization during training. Data Models
Capacity Flexibility • Support for block, file, and object data models and common
network protocols.
• Wide range (100s of gigabytes to petabytes) with easy, non-
disruptive expansion. Security
• High performance with billions of data items. • Strong administrative authentication.
• Range of cost points optimized for active and seldom- • “Data at rest” encryption.
accessed data.
• Protection against malware (especially ransomware) attacks.
Availability & Data Durability
OperationAL Simplicity
• Continuous operation over decade-long project lifetimes.
• Non-disruptive modernization for continuous
• Protection of data against loss due to hardware, software, long-term productivity.
and operational faults.
• Support for AI projects’ most-used interconnects
• Non-disruptive hardware and software upgrade and protocols.
and replacement.
• Autonomous configuration (e.g. device groups,
• Seamless data sharing by development, training, data placement, protection, etc.).
and production.
• Self-tuning to adjust to rapidly changing mixed random/
sequential I/O loads.
Appendix A lists potential pitfalls to be avoided when designing an AI storage infrastructure.
Uncomplicate Data Storage, Forever 5
BUSINESS WHITE PAPER
Pure’s AI Product Portfolio
Storage Systems
The Pure Storage® portfolio of all-flash storage systems,
illustrated in Figure 2, includes three FlashArray™ scale-up
Unified Block and File (UBF) servers, a FlashBlade® scale-out
Unified Fast File and Object (UFFO) server, and two Unified
Data Repository (UDR) servers, one based on FlashArray and
the other on FlashBlade, for low-cost large-scale storage.
All systems support broad ranges of capacity that is easily
expandable online. Each is optimized for specific capacity
ranges, data type(s), and cost/performance targets. With
these servers, Pure Storage can satisfy virtually any AI
storage requirement from project conception through model
training, and on into production. FIGURE 2 Pure’s AI Storage System Portfolio
Features Common to All Systems
All Pure Storage systems share key properties:
Reliability and Availability Continuity and Longevity
Systems are designed to continue operating when any internal Systems are designed for lifetimes of a decade or more of
component fails. For example, they survive at least two (in most continuous operation with no planned downtime or service
cases more) rare DirectFlash® Module2 (DFM) failures that outages, even during software and hardware upgrades
overlap in time without loss of data or client access to it. and modernizations.
Efficiency Evergreen®
Systems optimize capacity utilization by thin provisioning— Evergreen subscriptions that include regular software and
deferring space allocation until clients write data. They allocate hardware updates and periodic modernizations are perhaps
space autonomously to balance utilization and load. DFMs’ high Pure’s most important benefit for AI projects of long duration.
density minimizes system “footprint,” power consumption, and
3
The company offers subscriptions both for purchased systems
ultimately, e-waste. and for storage delivered as a managed service (Evergreen//
One™). Technical briefs TB-230601f and TB-230601o, available
Simplicity at https://support.purestorage.com or from a Pure Storage
Systems minimize administrative tasks to the greatest extent representative, describe the company’s Evergreen offerings in
possible—there are no “device groups” to manage, and no more detail.
data placement or protection decisions to make. Systems
report status and events to the Pure1® Cloud frequently; Pure1
analyzes behavior and proactively initiates any necessary
service operations.
Uncomplicate Data Storage, Forever 6
BUSINESS WHITE PAPER
Product-Specific Features
Spectrum of Performance and Cost Options
From latency-optimized FlashArray//XL™ for rapid response in production to throughput-optimized FlashBlade//S for training with
very large data sets, to cost-optimized FlashArray//E™ and FlashBlade//E™ for less-active data (e.g., raw data, feature stores, etc.),
Pure’s product line offers cost/performance options that span AI project requirements.
Pure systems’ very high-performing metadata operations on large numbers of files and objects make them particularly suitable for
AI project training.
Capacity Flexibility
With maximum capacities ranging from FlashArray//C50’s 1.6PB (effective4) to FlashBlade//S’s nearly 20PB (physical), Pure’s
systems can support many in-house AI projects from concept through production with a single project-wide data hub. Systems
can “start small” in a single chassis with minimal physical capacity and be expanded up to a model’s maximum supported capacity
without interrupting service to applications.
Where multiple systems are required, for capacity, performance, cost, or data model reasons, Pure1 centralizes storage
management and supports AI-based “what if” capacity planning for users’ entire “fleets” of Pure systems.
Data Reduction
Pure’s systems optimize flash utilization by compressing data prior to storing it. In addition, FlashArray systems, achieve further
efficiency by deduplicating blocks of data, replacing duplicate blocks with links to a single stored instance. Deduplication works
well with structured data (e.g., databases, tables, etc.).
NVIDIA Collaboration
With approximately 80% market share,5 NVIDIA Corporation is the
acknowledged leader in AI computation. Since 2017, Pure Storage has
collaborated with NVIDIA to develop solutions for AI. The collaboration
has resulted in the jointly-developed AIRI® Pure Storage NVIDIA DGX
BasePOD Reference Architecture6 for AI, based on NVIDIA’s DGX
servers and network fabric coupled with Pure’s FlashBlade//S™ storage
systems. As an NVIDIA BasePOD certified reference architecture,
AIRI eliminates the design, deployment, and management complexity
inherent in custom-crafted AI infrastructures. FIGURE 3 Pure-NVIDIA Collaboration
More recently, the Pure Storage-NVIDIA collaboration has resulted in:
• Pure’s implementation of the NVIDIA GPUDirect7 storage protocol to transfer data directly between FlashBlade//S storage and
NVIDIA’s GPUs, bypassing control CPU memory.
• Storage partner validation for FlashBlade//S in NVIDIA-certified OVX L40S reference architectures offered by major server
vendors. When combined with FlashBlade//S storage, OVX-certified servers are complete AI platforms that accelerate small
model training and fine-tuning, as well as GenAI RAG and production inference workloads.
Finally, when used in conjunction with Portworx®, NVIDIA’s device plugin for Kubernetes8 provides comprehensive management
and scheduling of both GPU and storage resources at all AI project stages.
Uncomplicate Data Storage, Forever 7
BUSINESS WHITE PAPER
Portworx: Pure’s Secret Weapon
Portworx by Pure Storage is the company’s Kubernetes data services platform that provides persistent storage, data sharing and
protection, workflow automation, and (optional) disaster recovery for containerized applications.
Portworx accelerates development of IT environments for the containerized applications used in most AI projects with a software-
defined storage model that enables infrastructure-neutral access to data. Portworx supports any type of block storage, whether
located on-premises or in a public or private cloud.
Portworx presents standardized virtual block or
file storage devices to applications, regardless
of the on-premises or cloud technology used to
instantiate it. It does this by making architect-
defined storage classes available to developers.
Storage classes standardize storage properties,
simplifying self-service job creation and promoting
stable, reliable development and production
environments. In addition, Portworx includes
templates that assist with setup for applications
like Apache Kafka, Zookeeper, Elasticsearch, as
well as for popular databases including SQL Server,
MongoDB, Postgres, and Cassandra, all of which are
commonly used in AI projects. It provides consistent
development environments while enabling
self-service job creation by data scientists and
other developers.
The software-defined storage model enables
Portworx to share data among multiple Kubernetes FIGURE 4 Using Portworx with a Database App
pods running separate jobs. This makes it particularly
useful for model training, where running many concurrent jobs that share the same input data is key to rapid implementation. As
an example, Figure 4 shows how Portworx can simplify deployment of a training application that utilizes data from a database.
Finally, Portworx provides fault tolerance by replicating its virtual storage devices to either on-premises resources or to a public
cloud. It can also protect entire project environments by copying them to S3 objects in a public or on-premises private cloud.
Uncomplicate Data Storage, Forever 8
BUSINESS WHITE PAPER
Table 1 suggests ways in which Portworx can be used to advantage in in-house AI projects.
Portworx Feature Application to AI
Up to 640PB per cluster Adequate storage capacity for any size AI project.
Software-defined Presents on-premises, private cloud, and public cloud storage to applications as
(“cloud-native”) storage feature-consistent block volumes or NFS file systems.
Public and hybrid cloud storage Supports mix of AWS or Microsoft Azure and on-premises storage for a Kubernetes
cluster.
Data Sharing & Resource / Workflow Management
Dynamic provisioning Enables convenient self-provisioning of virtual storage resources for greater data
scientist and developer agility.
Easy data migration Minimizes public cloud access charges by copying data to on-premises storage for
use by applications in multiple clusters.
Portworx
Vector database support Facilitates Retrieval Augmented Generation (RAG) for training customized GenAI
models managed by Kubernetes.
Python libraries for Simplifies common operations in data scientists’ Jupyter notebooks for self-service
Portworx services resource instantiation and management.
Disaggregated architecture Enables independent scaling of computing and storage eliminates need to provision
unnecessary resources.
Storage class, resource, and Standardizes Kubernetes resources for consistency while eliminating needs for
application templates developer awareness of implementation details.
Data replication and backup Transportable snapshots and backup of Kubernetes environments protect data
protection transparently to developers.
Application checkpoints Enables easy-to-use preservation of progress in long-running jobs.
TABLE 1 Portworx Features That Optimize AI Projects
While Portworx is designed to utilize any storage platform, combining it with Pure’s systems creates
a reliable, scalable, high-performing storage, data protection, and resource provisioning and
workflow management environment for containerized AI projects.
Uncomplicate Data Storage, Forever 9
BUSINESS WHITE PAPER
At the End of the Day...
The Pure Storage system portfolio includes storage for all phases of AI projects, large and small. Pure’s systems relieve IT, data
scientists and MLOps teams from most common storage management tasks. With them, data scientists can concentrate on
modeling and MLOps teams can provide reliable, scalable, high-performing environments for project data with sharable storage
that expands to meet both training and production needs without disruption.
Available in performance-optimized and capacity-optimized models that scale on demand, The Pure Storage platform accelerates
and enhances AI projects for healthcare, genomics, exploration, financial and many other fields. It can do so while sharing storage
capacity and I/O bandwidth with other data-intensive applications such as analytics, database backup and restore, software
development, media and entertainment post-production, electronic design automation (EDA) and more. .
Portworx takes the guesswork out of creating robust, scalable Kubernetes environments for containerized AI training and
production applications. Its templates simplify configuring and implementing applications and databases commonly used in AI
projects; its built-in storage services enable data sharing and backup of both project data and entire project environments.
For Further information…
Additional information on topics discussed in this brief is available at:
Pure Storage in Artificial Intelligence
https://www.purestorage.com/solutions/analytics-and-ai/artificial-intelligence.html
Overview of Pure Storage Portworx
https://portworx.com/
Uncomplicate Data Storage, Forever 10
BUSINESS WHITE PAPER
Appendix A:
AI Storage Pitfalls
Data storage can have a profound impact on in-house AI project cost, development speed, and ultimately, project success.
This appendix lists pitfalls to be avoided when planning and developing project storage infrastructures.
Configuration Rigidity Inflexibility
It is virtually impossible to predict lifetime storage needs I/O requirements vary throughout the life of an AI project.
accurately at the start of an AI project. Storage systems that Storage that is highly optimized for narrow use cases (e.g.,
are awkward or impossible to reconfigure, expand, or upgrade by data set location, block sizes, file types, etc.) can make it
should therefore be avoided. Even upgradable systems difficult to respond as a project evolves. Manual storage “tuning”
can be problematic if upgrading entails service outages, interrupts development and production and requires expensive
especially in the production phase when inferences are driving expertise. And even experts don’t always get it right.
business decisions.
Data Model Rigidity
Fragility At the outset, AI projects typically utilize files and/or structured
If an AI project is important enough to invest in, it’s important databases, hosted either on block storage or in file servers.
enough to keep available. Data scientist productivity is As models develop and training data scales, many tools use
important, but as projects move to intensive training and data in object form to simplify organization and processing and
production, availability is vital. Storage that can’t survive construct vector databases that maximize the quality of search
component failures, be repaired or upgraded online, or recover results. Storage systems that support only a single data model
from user and administrative errors isn’t up to the AI job. (e.g., file servers, block storage systems) may make converting
input data into forms usable by training jobs and in production
Data Isolation awkward and time-consuming.
Data that can’t be easily shared among developers and between
containerized training apps and production tasks must be copied
from where it is to where it’s needed. As data sets expand,
copying large data sets becomes disruptive. In addition to being
time and resource-consuming, data set copying initiated by
humans can be error-prone.
Uncomplicate Data Storage, Forever 11
BUSINESS WHITE PAPER
Appendix B:
Using Pure’s Products for AI
The Pure Storage portfolio of all-flash systems meets AI capacity, I/O performance, and orchestration needs from project
conception through model development, on into production. Table 2 suggests the optimal applications for each of Pure’s systems
in AI projects.
Product Training Production Infrastructure
• Vector databases
very large objects I/O-intensive production
UFFO
0.35PB—19.6PB • Retrieval Augmented
(images, video,…) (e.g., GenAI)
(physical) Generation (RAG)
• Ecosystem streaming
(Kafka, etc.)
large data sets
0.96PB—5.5PB (effective) • Kubernetes volumes
& VMs
Latency-sensitive
production • Vector databases
• Retrieval Augmented
UBF
medium-large
Generation (RAG)
0.31PB—3.3PB (effective) data sets
• Kubernetes volumes
& VMs
• Feature stores
• Inference
1.6PB—8.9PB (effective)
input & output
1.0PB—4.0PB (effective) Raw training data Feature stores
Exception logging
UDR
seldom-accessed after Inference
Project archives
initial use input & output
4.0PB—8.0PB (physical)
TABLE 2 Pure Storage Systems in AI Projects
Uncomplicate Data Storage, Forever 12
BUSINESS WHITE PAPER
Storage for Every AI Project
As Table 2 suggests, Pure’s systems offer a breadth of capacity, cost/performance, and protocol support that are a fit for all
aspects of an AI project. In terms of the desirable AI storage properties listed on page 5:
Performance Agility Availability & Durability
The portfolio includes a range of cost/performance options to Pure’s systems provide reliable “24x7” storage for AI model
meet lifetime project needs: training and production:
• All systems utilize reliable, high-performing, power- • Systems provide continuous availability over lifetimes that
efficient flash-based DFMs exclusively. exceed a decade.
• Latency-optimized UBF systems support quality-of- • Systems are highly available—internal redundancy
service (QoS) limits for prioritizing performance among allows them to continue operating if components fail.
data sets. Purity software is designed to sustain a minimum of two
concurrent DFM failures (in many cases, more) without
• Throughput-optimized UFFO systems scale out—adding
loss of data availability.
blades to a system increases both storage capacity,
network performance, and processing capability. • All updates and modernizations are non-disruptive,
performed while systems are serving client I/O requests
• Cost-optimized UDR systems make retaining large
and with minimal performance impact.
amounts of infrequently-accessed data affordable.
• DFM service lifetimes are over 10 years—far longer than
• Rapid File Toolkit for UFFO systems accelerates bulk
typical SSD specifications. Data durability (i.e., statistical
operations on millions of files.
mean time to data loss, or MTTDL) due to DFM failure is on
Capacity Flexibility the order of millions of years.
All Pure systems are easily expandable from minimum to
Space and Power Efficiency
maximum capacity by non-disruptive addition of DFMs (and
Pure’s high-density all-flash system architectures:
blades in scale-out UFFO systems). AI project architects can
specify systems based on production expectations with initial • Hold up to 2 petabytes of physical storage in five
configurations sized to meet early-stage capacity needs and rack units.
add capacity incrementally as projects progress: • Consume up to 90% less power per terabyte than
competitive storage systems.
• Result in much less “e-waste” due to longer lifetimes
of devices and storage systems.
Uncomplicate Data Storage, Forever 13
BUSINESS WHITE PAPER
Data Models Operational Simplicity
AI development and production tools utilize block storage, As AI projects evolve from concept through training and
file servers, or object stores. Pure’s systems support block, into production, a smooth operating environment becomes
file, and object protocols for easy integration: increasingly important. Pure storage systems’ ease of operation
• UBF systems support client-side file systems and object can simplify a project’s operating environment in several ways:
stores with virtual block devices as well as providing file • They are typically installed and ready for use in a few
services for hundreds of file systems containing as many hours rather than days.
as a half-billion files.
• With support for all common I/O protocols (NVMe, NFS,
• UFFO systems support petabyte-scale data sets SMB, S3, iSCSI, GPUDirect), they integrate easily into data
containing billions of files or S3 objects in thousands of file centers. UBF systems can provide block and file services
systems or object buckets. simultaneously; UFFO systems can simultaneously serve
• UDR systems store less-frequently accessed data, files and objects.
whether blocks, files, or objects, economically. • They eliminate many common management operations—
specifying protection, placing data sets, reserving spare
Security capacity, tuning for performance, etc. are all autonomous.
Many AI projects utilize proprietary data. As models mature, • Both software/hardware upgrades and system
both curated training data and the intellectual property modernizations are non-disruptive (i.e., are performed
in models’ digital representations become increasingly while systems are operating) and have negligible impact
valuable. Pure’s systems provide a secure environment for on performance. Data migration is never required.
digital information: • They are self-monitoring; they upload status and
• They use strong administrator authentication to event logs9 every few seconds to the Pure1 monitoring
control access. and consolidated management service (available to
all customers with active Evergreen subscriptions).
• They automatically encrypt all stored data and metadata.
Pure1’s AI-based behavioral models identify and report
• They can be configured to take periodic snapshots potential issues and provide experience-based capacity
of selected data to protect against application and planning advice.
administrative errors.
• Users can enable SafeMode™ to protect against malware
(e.g., ransomware).
Accelerate AI Adoption With The Pure Storage Platform
1 For GenAI, “input data” usually takes the form of natural language queries.
cf: https://www.techtarget.com/searchdatamanagement/feature/Vector-search-now-a-critical-component-of-GenAI-development
“LLMs…are trained with extensive vocabularies and can determine the meaning of a question even if it isn't phrased in the business-specific language…”
2 DFMs are Pure Storage-designed flash memory modules that provide the persistent storage in all Pure systems. They fulfill the role played by SSDs in conventional flash-based storage systems.
3 Up to 75 terabytes per module at publication time.
4 Effective capacity is that available for user data, net of erasure code overhead and system metadata. Users typically experience up to 2:1 data reduction with FlashBlade compression and as much as 5:1
with FlashArray deduplication and compression. Data reduction depends strongly on the nature of data. For example, structured text and tabular data usually reduces by at least 2:1, whereas images,
streams, and encrypted data are essentially uncompressible.
5 https://www.britannica.com/topic/NVIDIA-Corporation
6 https://www.purestorage.com/docs.html?item=/type/pdf/subtype/doc/path/content/dam/pdf/en/reference-architectures/ra-airi-nvidia-dgx-basepod-architecture-config-guide.pdf
7 GPUDirect is a registered trademark of NVIDIA Corporation.
8 https://catalog.ngc.nvidia.com/orgs/nvidia/containers/k8s-device-plugin
9 Except where customer policies or regulations prohibit external connections.
purestorage.com 800.379.PURE
©2024 Pure Storage, the Pure Storage P Logo, AIRI, DirectFlash, Evergreen, Evergreen//One, FlashArray, FlashArray//E, FlashArray//XL,
PS2571-02-en 05/24
FlashBlade, FlashBlade//E, FlashBlade//S, Portworx, Pure1, SafeMode, and the marks in the Pure Storage Trademark List are trademarks
Uncomplicate Data Storage, Forever
or registered trademarks of Pure Storage Inc. in the U.S. and/or other countries. The Trademark List can be found at purestorage.com/
14
trademarks. Other names may be trademarks of their respective owners.
