The global DRAM market reached $26.02 billion in the third quarter of 2024, marking a 13.6% quarter-on-quarter increase, according to research from analyst TrendForce.
The rise was driven by growing demand for global DRAM and HBM products in datacenters, and came despite a decline in LPDDR4 and DDR4 shipments due to inventory reduction by Chinese smartphone brands, and capacity expansion by Chinese DRAM suppliers.
ASPs (annual selling prices) continued their upward trend from the previous quarter, with contract prices rising 8 percent to 13 percent, further supported by HBM’s displacement of conventional DRAM production.
Looking ahead to 4Q24, TrendForce projects a QoQ increase in overall DRAM bit shipments. It says the capacity constraints caused by HBM production are expected to have a “weaker than anticipated” impact on pricing. Capacity expansions by Chinese suppliers may prompt PC OEMs and smartphone brands to “aggressively deplete inventory” to secure lower-priced DRAM products. As a result, contract prices for conventional DRAM and blended prices for conventional DRAM and HBM are “expected to decline”.
Server and PC DRAM contract price increases lifted revenues for the top three DRAM manufacturers in Q3. Samsung retained the top spot with revenue of $10.7 billion, up 9 percent QoQ. By strategic inventory clearing of LPDDR4 and DDR4, bit shipments at Samsung remained flat compared to the previous quarter.
Over at SK hynix, there was reported revenue of $8.95 billion, a 13.1 percent QoQ increase, and it maintained its second-place position. Although its HBM3e shipments ramped up, a 1 percent to 3 percent QoQ decline in bit shipments from weaker LPDDR4 and DDR4 sales offset these gains.
Micron saw its revenue surge by 28.3 percent quarter-on-quarter to $5.78 billion, driven by “strong growth” in server DRAM and HBM3e shipments, which led to a 13 percent QoQ increase in bit shipments.
Taiwanese DRAM suppliers saw their revenues decline in Q3, falling significantly behind the top three manufacturers. Nanya Technology faced a more than 20 percent QoQ drop in bit shipments due to weaker consumer DRAM demand, and intensified competition in the DDR4 market from Chinese suppliers. Its operating profit margin further deteriorated from -23.4 percent to -30.8 percent, reflecting losses from a power outage incident.
Winbond experienced an 8.6 percent QoQ decline in revenue, falling to $154 million, as consumer DRAM demand softened and bit shipments decreased. And PSMC reported a 27.6 percent QoQ decline in revenue from its in-house consumer DRAM production. However, including foundry revenue, its total DRAM revenue rose 18 percent QoQ, driven by ongoing inventory replenishment from its foundry clients.
Dell revenue growth for Q3 FY25 fell short of expectations, driven primarily by new AI and traditional servers.
In the quarter ended November 1, Dell missed the midpoint $24.5 billion of its outlook, bringing in $24.4 billion, with GAAP net income up 12.3 percent year-over-year at $1.13 billion. The year-ago numbers were $22.25 billion in revenues and $1 billion profit.
Vice-chairman and COO Jeff Clarke answered an earnings call question about the outlook shortfall, saying: “The PC refresh continues to move out. It was the source of our underrun in Q3 … The second element … is the shift in demand in AI to Blackwell.” Nvidia Blackwell GPU shipments are now on backlog.
Yvonne McGill
CFO Yvonne McGill stated: “We continued to build on our AI leadership and momentum, delivering combined ISG and CSG revenue of $23.5 billion, up 13 percent year-over-year. Our continued focus on profitability resulted in EPS growth that outpaced revenue growth, and we again delivered strong cash performance.”
The two main business groups fared differently. Infrastructure Solutions Group (ISG) revenue of $11.4 billion rose 34 percent year-over-year, although it declined sequentially from last quarter’s $11.65 billion. Servers and networking revenue was $7.4 billion, up 58 percent year-over-year, and down 3.5 percent sequentially.
Storage revenues increased at a much slower rate – just 4 percent year-over-year to $4 billion, up 0.7 percent sequentially. Demand for both the PowerStore and PowerFlex products grew in double digits. Clarke said: “The PowerScale F910 and F710 continue to ramp nicely, driving double-digit growth in the unstructured all-flash portfolio.” But, in storage overall, “the demand environment continues to trail traditional servers.”
Dell is well ahead of its competitors in the storage market:
1) IDC Quarterly Server Tracker, 2024 Q2, based on CY13 – 2Q24 TTM revenue. Mainstream Server is based on OEM vendor type and includes: Large System, Standard Rack, Tower, and Blade. 2) IDC Quarterly Enterprise Storage Systems Tracker, 2024Q2, based on 2Q24 TTM revenue. Midrange refers to systems with ASP between $25k and $250k and High End refers to systems with ASP > $250k. 3) IDC Quarterly Converged Systems Tracker 2024Q2, based on 2Q24 TTM revenue. 4) IDC Worldwide AI and Generative AI Spending Guide, V1 2024 (Aug). Hardware includes Servers and Storage only
Clarke said: “AI is a robust opportunity for us with no signs of slowing down. Interest in our portfolio is at an all-time high, driving record AI server orders demand of $3.6 billion in Q3 and a pipeline that grew more than 50 percent with growth across all customer types.
“Increasingly, enterprises see the disruptive nature and the innovation opportunities with GenAI resulting in growing GenAI experimentation and proof-of-concepts. … Beyond the AI servers, we like the profit pools that surround them, like power management and distribution, cooling solutions, network switches, network cables, optics, storage, deployment, maintenance, professional services and financial services … The opportunity in AI is enormous.”
Dell said ISG operating income improved to 13.3 percent of revenue, up 230 basis points sequentially, and traditional server demand improved double digits in the quarter, making the fourth consecutive quarter of traditional server demand growth.
The Client Solutions Group’s (CSG) revenue was $12.1 billion, down 1 percent year-over-year with client revenue up 3 percent at $10.1 billion, but consumer revenue down 18 percent year-over-year to $10.1 billion. People aren’t buying AI PCs in any great numbers yet but, Clarke said: “We are seeing an indication that customers are lining up their upgrade cycles with new AI PCs in the first half of next year – a clear signal that enterprises are balancing their need to refresh and their desire to future-proof their purchases.”
McGill mentioned: “Enterprise demand was promising, though less than expected as we saw some demand push into future quarters.” And: “Our Consumer business was weaker than expected as demand and profitability remain challenged.”
Dell experienced the third consecutive quarter of both year-over-year and quarter-over-quarter commercial PC demand growth. Clarke said: “We are optimistic about the coming PC refresh cycle as the install base continues to age, and with Windows 10 reaching end-of-life in 46 weeks.”
Dell has the number one share position in the commercial AI PC market, according to IDC.
The outlook for the fourth fiscal 2025 quarter is that revenues will be $24.5 billion +/- $500 million – a 9.8 percent increase year-over-year at the midpoint. Dell expects the combination of ISG and CSG to grow 13 percent at the midpoint, ISG to be up in the mid-twenties driven by AI and traditional servers, and CSG to be up low-single digits.
Its general expectations for next fiscal year are for more robust AI demand with a strong five-quarter pipeline going into next year. There will be an aged installed base in both PCs and traditional servers – prime for refresh. It expects ISG growth driven by AI servers, followed by traditional servers, then storage, and expects CSG to grow as enterprise companies refresh a large and aging installed base.
Bootnote
Storage for Dell is not seeing any AI-led demand increase, but Dell thinks it will happen. Clarke was pragmatic: “Look, we’re pretty excited about the opportunity in AI to expand beyond the individual node. I think … the opportunity is beyond the node into full rack scale integration. And in full rack scale integration, it’s the networking opportunity, the storage opportunity, mundane things like cooling.”
Jeff Clarke
He added: “We expect the storage marketplace to grow next year and we expect to take share with our Dell IP storage portfolio as we’ve invested into new solutions, new capabilities, making it more competitive.”
Clarke expanded on this, saying: “The AI opportunity for storage is immense simply because GPUs devour data. I mean, you have to feed the beast, and they’re not very effective without a lot of information.
“Storage, it typically tends to be unstructured information. So we think scale out, file and object capabilities for training, tuning and inference are essential. We think parallel file systems for this kind of end-use transient data and these large training environments is essential. And remember, 80 percent of the data is on-prem. So we think AI is driving new needs in the storage architecture, which really drive to a three-tier architecture.”
Dell is well positioned, he claimed. “The Dell IP portfolio is a three-tier architecture moving towards disaggregated that allows us to scale CPU and storage and networking independently to optimize for performance. All that said, our PowerScale platform, we believe, is absolutely the right platform where we deliver high performance file, soon high performance object, the ability to take our Project Lightning that we announced at Dell Technology World, a parallel file system built exactly and purposeful for AI, and then take our Dell Data Lakehouse, which then allows us to do some data management and have the data platform around this and specifically metadata management. That combination of capability, we believe, is absolutely the future state of storage for AI and we’re very well positioned.”
Specifically, with regard to PowerScale, “the F710 and the F910. One has 614 terabytes of capacity. If I remember correctly, the 910 is 1.46 petabytes of capacity. Both of those double as we go to higher density drives in the first half of next year that will be first to market.”
Nutanix revenues grew 16 percent year-on-year to $591 million as it comfortably beat its $575 million high point outlook.
It made a GAAP profit of $29.9 million in its first fiscal 2025 quarter, ended October 31, the second profit in its history; the previous one being $32.8 million in the second fiscal 2023 quarter.
Rajiv Ramaswami
Nutanix president and CEO Rajiv Ramaswami said in the earnings call: “We’re happy to report first quarter results that came in ahead of our guidance. We continue to see steady demand for our solutions driven by businesses prioritizing their digital transformation and infrastructure modernization initiatives and looking to optimize their total cost of ownership or TCO.”
He added: ”We also saw another quarter of strong year-over-year growth in new logos and solid free cash flow generation.”
CFO Rukmini Sivaraman said: ”Our first quarter results demonstrated a good balance of top and bottom line performance with 18 percent year-over-year ARR growth and strong free cash flow generation. We remain focused on delivering sustainable, profitable growth.
”The outperformance in revenue was driven by good renewals execution … We saw strength in landing new customers onto our platform, helped by more leverage from our OEM and channel partners from the various programs we have put in place to incentivize new logos, and from a general increase in engagement from customers looking at us as an alternative in the wake of industry M&A.”
The last point is a reference to Broadcom’s acquisition of Nutanix competitor VMware.
A relatively minor point was: “Our US Fed business performance was lower year-over-year relative to the strong comparison from Q1 a year ago.”
Nutanix revenues by quarter by fiscal year showing solid sustained growth for 12 consecutive quarters
Financial summary:
Gross margin: 86.0 percent vs 84.0 percent last year
Free cash flow: $151.9 million vs $132.5 million
Operating cash flow: $161.8 million vs $145.5 million
Cash, cash equivalents and short-term investments: $1.08 billion compared to $994 million at the end of the prior quarter and $612.5 million a year ago.
Nutanix gained 530 new customers since the last quarter, taking its total to 27,160. It said new logo additions grew more than 50 percent year-over-year. Ramaswami pointed out: “Cisco has been a good contributor to our new logos this quarter as well as last quarter.”
Ramaswami talked about the Nutanix Dell partnership concerning Dell’s PowerFlex offering, saying: “We expect to have that in the market in the first half of calendar ’25. And we expect to see some revenue contribution from it in FY ’26.”
This deal involves PowerFlex external storage integrated with Nutanix’s AHV hyper-converged infrastructure (HCI) software system. More similar deals are in the air, with Ramaswami saying: “We do have plans to expand that selectively to other storage arrays over time … We also want to make sure that whatever we do is actually focusing on expanding the market opportunity for us and doesn’t cannibalize the existing HCI opportunity … We are focused specifically on working with storage array vendors with workloads that would not necessarily be easier to capture with HCI.”
Nutanix said it is seeing continued land-and-expand opportunities and a growing pipeline for its products. However, it expects uncertainty in the timing, outcome, and deal structure from the growing mix of larger deals in the pipeline, and ongoing elongation of average sales cycles relative to historical levels, which it believes is due to the uncertain spending environment.
Sivaraman said: “We believe the larger opportunities in our land-and-expand pipeline continue to involve strategic decisions and C-suite approvals, causing them to take longer to close and to have greater variability in timing, outcome, and deal structure.”
On that basis, next quarter’s outlook is for revenues of $640 million +/- $5 million; 13.2 percent year-over-year growth. The full year revenue outlook is $2.45 billion +/- $15 million – 14 percent growth year-over-year.
Bootnote
A point that came out in the call is that Nutanix’s HCI systems, with their server-attached storage, compete with external storage arrays. When customers with three-tier architecture approach hardware renewal time, particularly storage arrays, Nutanix’s selling partners will say that the customer can save money in a TCO sense by moving to Nutanix HCI and junking the external storage arrays. Even though Nutanix is integrating external storage into AHV, with the Dell PowerFlex deal, for example, it hopes to move such customers to HCI eventually.
Hitachi Vantara has agrrements with Hammerspace and WEKA, two companies supplying fast file data delivery services. Why does it need both?
Hitachi Vantara became Hammerspace’s first major systems supplier earlier this month with Hitachi Vantara using its Global Data Platform (GDP) global namespace, in which data is placed, orchestrated, made available for use, and accessed as if it were local. Hitachi Vantara customers can now use GDP to access other suppliers’ storage in a single environment, and use it in GenAI training and inferencing applications. GDP uses parallel NFS-based file system software to manage file and object data in globally distributed locations, across SSDs, disk drives, public cloud services, and tape media.
The company also partners with parallel file system supplier WEKA, using rebranded WekaFS file system software (Weka Data Platform) in HCFS (HItachi Content Software for File). Hitachi Vantara includes HCFS in its recently launched Hitachi iQ, a set of industry-specific AI systems using Nvidia DGX and HGX GPUs and Hitachi Vantara storage.
Jason Hardy
Having set the background, we can move on to the three questions we asked Jason Hardy, Hitachi Vantara CTO for AI, about the Hammerspace and WEKA partnerships and its Nvidia server systems certification.
Blocks & Files: How does this Hitachi iQ, which uses WEKA parallel file system software in Hitachi Content Software for File, relate to the partnership with Hammerspace for its Global Data Environment?
Jason Hardy: As part of Hitachi Vantara’s ongoing strategy to streamline data management and equip our customers with the tools necessary to accomplish their AI goals, the Hammerspace partnership equips Hitachi iQ with capabilities to provide more robust data management and access methodologies, supporting many of the complex AI workflows including Retrieval Augmented Generation (RAG) and other analytical workloads.
The new relationship with Hammerspace also extends the company’s capabilities for AI workloads, including intelligent file-granular data orchestration, multi-site global namespace, and hybrid cloud support.
Just like VSPOne is a key piece of the Hitachi iQ offering, many other storage technologies and data management capabilities will be included in the Hitachi iQ portfolio as we further develop and streamline the robustness of our capabilities.
Blocks & Files: How does this (Hitachi iQ) relate to Hitachi Vantara’s previous Nvidia BasePOD certification? I think it’s effectively a BasePOD DGX to HGX support upgrade.
Jason Hardy: Our commitment remains the same, to support those customers who choose to utilize the DGX eco-system and the Nvidia extended software capabilities, while also providing flexibility to choose alternative options depending on the customer’s requirements.
Hitachi iQ’s certification and participation in the Nvidia BasePOD program allows Hitachi Vantara to support our customers who have utilized Nvidia’s DGX compute offering and BasePOD framework, to support their AI ambitions.
The Hitachi iQ BasePOD architecture consists of the Nvidia DGX, Hitachi Content Software for File, and the necessary connectivity and software (including Nvidia AI Enterprise).
Matching much of what was designed for the BasePOD reference architecture, customers who would like to utilize the same capabilities provided by DGX but may not have access to a DGX channel partner or may be looking for a solution that is more flexible for their requirements, Hitachi’s HGX offering provides the same GPU processing capabilities while providing more flexibility to how it can be deployed.
Blocks & Files: How does this relate to Nvidia SuperPOD certification?
Jason Hardy: The release of our Hitachi iQ HGX offering doesn’t preclude us from participating in the Nvidia SuperPOD program. And while that is still a strategic goal, this is a long-term objective that requires coordination with Nvidia.
The creation of our HGX offering will allow our customers to deploy the same capabilities more flexibly while maintaining support and services through Hitachi Vantara.
Data protector HYCU has teamed up with three other suppliers to help financial institutions meet the European Union’s DORA requirements.
The EU Digital Operational Resilience Act (DORA) is a set of regulations designed to enhance the cyber resilience of financial institutions. Its intent is to ensure they can function during cyber attacks or other potentially disastrous IT incidents, with standards for managing cyber security risks, incident reporting, and digital resilience for banks, insurers, payment service providers, and other financial entities. DORA goes live in January next year.
DORA is not just another bureaucratic tangle of pesky point regulations. According to HYCU CEO Simon Taylor: “There’s personal liability. Now within DORA, this is no longer, oh, the company will pay the fine. Now the CIO, or the operating board member, is responsible for the fines and for personal prosecution.”
Nathan Chantrenne
IT consulting and services firm Valiantys has launched a Governance, Risk, and Compliance (GRC) service with a focus on DORA. It’s composed of elements from HYCU, Lansweeper, and Appfire. This is a complicated area and the four suppliers need to work together to provide the various components of an overall DORA compliance system.
Nathan Chantrenne, Valiantys chief solutions officer, stated: “By integrating advanced compliance management capabilities into Jira Service Management, we are enabling organizations to embrace a more resilient, streamlined approach to regulatory challenges.”
Subbiah Sundaram, HYCU SVP for Products, said: “We’re proud to contribute our data protection expertise to this innovative GRC solution. Our partnership with Valiantys, Appfire, and Lansweeper allows us to offer customers a seamless, integrated approach to meeting DORA’s stringent requirements.”
Valiantys is an Atlassian Platinum Solutions Partner. Lansweeper provides asset intelligence. Appfire’s software provides enterprise collaboration workflow tools across Atlassian, Microsoft, monday.com, and Salesforce. It has more than a million users and the most widely adopted portfolio of Atlassian apps across tens of thousands of customers worldwide.
Subbiah Sundaram
The Lansweeper platform discovers and optimizes over 80 million connected devices from 20,000-plus customers – including Mercedes, Michelin, Liverpool FC, Carlsberg, Nestle, IBM, and Samsung, along with governments, banks, NGOs, and universities.
Valiantys claims its GRC service “significantly reduces the time and resources required for DORA compliance, eliminating the need for hundreds of hours of manual asset cataloging, multiple backup tools, and extensive documentation management.” It provides:
Automated asset discovery and visualization of as-a-service applications;
Automated data protection and offsite data recovery for over 80 ICTs (Information and Communication Technologies) with continuous expansion;
One-click resilience testing capabilities;
DORA-focused dashboards, notifications, and insights.
Doug Kersten, Appfire CISO, said: “As the regulatory landscape becomes more stringent and complex, organizations have a responsibility to work together to help customers navigate these emerging regulations. Collaborating with Valiantys, Lansweeper, and HYCU allows us to leverage each partner’s unique expertise to help organizations meet and adhere to the complex compliance requirements posed by DORA.”
For more information about the Valiantys GRC service, click here.
Analyst house Forrester has issued a data resilience Wave report looking at nine suppliers, and Commvault pips Veritas to the lead position.
Forrester’s Wave diagram shows a set of two concentric quarter circles placed in a square space defined by a vertical current offering axis and a strategy area space outside them to the left. These areas are named Contenders, Strong Performers, and Leaders from left to right with suppliers ranked and placed in each category area:
The study explains: “Data resilience solutions encompass backup and recovery functionality as well as data security and governance needs.” These needs have evolved substantially in the past couple of years due to cloud infrastructure-as-a-service (IaaS), software-as-a-service (SaaS), Kubernetes and new virtualization platform adoption.
There are just three leaders. It’s noticeable that Veritas is quite close to Commvault with Veeam some distance away due to its lower offering strength, despite the highest strategy rating of all the suppliers. Veeam was not a leader in the 2022 issue of this report, nor was Veritas, whereas Rubrik and Cohesity were both ranked as leaders. Now they are categorized as strong performers along with Druva and Dell. Like Veeam, Veritas has improved its ranking, as a look at the 2022 Data Resiliency Wave diagram illustrates:
This older Wave diagram has three concentric circles for Contenders, Strong Performers, and Leaders, with the area to the left set aside for Challengers – a weaker category of supplier absent from the 2024 study.
This 2022 Forrester Data Resilience study looked at virtually the same nine suppliers but ranked them quite differently, apart from Commvault being the leader. Zerto has now been eliminated and OpenText just scrapes in.
In the latest Data Resilience Wave report, Commvault received the highest possible scores in 13 criteria within the current offering category – including hyperscale cloud/IaaS, regulatory and compliance features, directly supported SaaS platforms, protection for generative AI models/data, and recovery to alternate infrastructure. Forrester also cited its revamped core platform, Commvault Cloud, powered by Metallic AI, and added capabilities from acquisitions like TrapX and Appranix, forming Threatwise and Cloud Rewind features respectively.
Commvault’s chief product officer, Rajiv Kottomtharayil, stated: “Commvault is committed to delivering industry-leading cloud-first cyber resilience solutions that help our customers anticipate, mitigate, and recover from cyber attacks. This is where our strength lies, and we believe the industry agrees. We have been recognized as a leader in this Forrester Wave, as well as the 2024 Gartner Magic Quadrant for Enterprise Backup and Recovery Software Solutions and the GigaOm Sonar Report for Cloud-Native Data Protection.”
With Cohesity buying Veritas, it will likely leap up the rankings.
Veeam CEO Anand Eswaran commented: “In a digital world, Data Resilience – ensuring data is always available no matter what happens – is critical to keeping your business running. More than 550,000 customers and 34,000 partners around the world rely on Veeam as the number one global leader to deliver the industry’s most innovative and trusted Data Resilience solutions. With close to 15 significant product releases in the past 12 months, including the launch of Veeam Data Cloud, and Veeam Data Cloud Vault, Veeam is pioneering the next wave of innovation to ensure data is protected and available.”
Commvault has just announced Clumio Backtrack – a capability that will enable enterprises to use automation to rapidly revert billions of objects stored in Amazon Simple Storage Service (Amazon S3) to a specific version at a specific point in time. This follows on from its Clumio acquisition.
Download and read the full Forrester Wave Data Resiliency report by registering at the Commvault website here.
45Drives has announced the Stornado F16, its most powerful storage server to date. It has a 2RU chassis with:
AMD EPYC or Intel Xeon CPUs, expandable to 4TB of memory, and ready for 100GbE networking
Sequential read speeds of 56.5 GB/s, write speeds of 55 GB/s, and up to 10.2 million IOPS.
16 U.3 NVMe drive bays supporting up to 243.2 TB of raw storage.
Tri-mode backplane for seamless compatibility with U.3 NVMe, SATA, and SAS drives.
It claims this is a ‘”leap forward from the acclaimed Stornado F2, doubling the number of PCIe Gen5 lanes per drive for unmatched data throughput.”
…
Commvault has announced Clumio Backtrack to enable enterprises to use automation to rapidly revert billions of objects stored in Amazon S3 to a specific version at a specific point and time. Organisations can currently use Amazon S3 Versioning to recover specific objects; so, if a piece of data in Amazon S3 is lost, deleted, or altered, users can go back in time and revert to a good copy of that data. Commvault’s technology takes that premise to new heights. Backtrack can revert billions of objects to earlier versions, radically changing how quickly and easily large-scale datasets can be rolled back – even seconds after an issue comes to light.
Via its serverless architecture, Backtrack will uniquely empower organisations to recover datasets of practically any size, from individual objects to whole Amazon S3 buckets where billions of objects may be stored. Commvault will offer early access to Clumio Backtrack in December, with global general availability planned for early 2025.
…
According to The Information, Databricks has told investors it aims to raise $7 – 9 billion at a ~$61B valuation, to cash out employees that hold restricted stock units.
…
Rami Douenias has joined DDN as VP of Global Sales Engineering. He was Quantum’s AVP – Head of Presales / Solution Engineering – America’s for a year, and at Cohesity before that, for just 8 months, coming from an 8 year stint at Pure Storage.
…
DDN is enhancing the speed of its Infinia object storage. At AI STAC in London, UK, December 4, 2024, it will “share groundbreaking advancements that make object storage a Tier 1 data platform for latency-sensitive applications.” It will be “reducing object storage latency to sub-millisecond levels” and thereby “accelerating real-time decision-making, backtesting, and model training” and “optimizing Spark-based analytics for large-scale AI workloads.” We think DDN will be supporting NVIDIAs GPUDirect-like S3 over RDMA scheme.
…
Data manager Denodo launched Denodo Platform 9.1, adding new AI capabilities and tools to Platform 9. It has an AI-powered Denodo Assistant, which uses the semantic layer of the Denodo Platform to automate data engineering tasks and deliver contextualised insights and recommendations to data analysts and other business users. It features;
Query Wizard Recommendations: Builds on Denodo Platform 9’s support for natural language queries to provide step-by-step guidance through query creation.
Intelligent Query Autocompletion: Offers context-aware suggestions while users write queries.
Data Preparation Wizard: Enables data products to be tailored to different use cases,.
Suggestions for View and Column Descriptions: Automatically generates business-meaningful descriptions for views and columns.
Enrichment of Text-based Unstructured Data: Automatically summarises and classifies text, identifies and extracts data entities, analyzes sentiment, identifies and redacts sensitive data, and translates text using LLMs, all available as a single function call.
Platform 9.1 also includes the new Denodo AI SDK, an open-source toolkit that accelerates the development of AI-powered applications and agents, especially when implementing RAG (retrieval augmented generation). The Denodo AI SDK simplifies the integration of both structured and unstructured data into GenAI models, for higher accuracy and better performance.
…
This sounds interesting: Cloud backup supplier Eon has raised a $70 million Series C funding round led by BOND, growing its valuation to $1.4 billion. Returning investors included Sequoia Capital, Greenoaks, and Lightspeed Venture Partners. Founded in January 2024 by CEO Ofir Ehrlich and Gonen Stein, of the CloudEndure founding team (acquired by Amazon Web Services in 2019), and Ron Kimchi, former general manager of AWS migration and disaster recovery services, Eon launched in October with $130 million in funding. It has already filed dozens of patents for cloud storage and data management technologies and raised $200 million in funding overall in under a year. Eon’s platform eliminates manual backup tasks by autonomously scanning, mapping, and classifying cloud resources.
Eon says its mission is to provide instant access to all backed-up cloud data, through a next-generation platform with the first backup autopilot. Visit https://www.eon.io/ for more information.
…
Cloud data manager Informatica announced new capabilities for Microsoft Azure, to help customers build, deploy, and activate AI and analytics-driven innovations with trusted data:
Gen AI Blueprint for Azure OpenAI Service – Accelerates development and deployment of enterprise-grade GenAI and copilot experiences with reference architectures and templates using Informatica’s Intelligent Data Management Cloud (IDMC) platform and Azure OpenAI Service.
Enhanced SQL ELT for Microsoft Azure – Enables no-code data pipeline definition and execution with in-database SQL-based processing for increased performance and scale.
Open Table Iceberg Support for ADLS Gen2 – Supports data migration and integration to Azure Data Lake Storage (ADLS) Gen2 in Iceberg format, unlocking price-to-performance benefits of Iceberg in storing and accessing large-scale data sets for AI.
…
Sources like Reuters and Bloomberg are saying a Kioxia IPO will take place around Dec 18 with pricing initially set at ¥1390/share suggesting a market cap of close to $5B. Wedbush analyst Matt Bryson tells subscribers “we see this result at somewhat disappointing for WDC vs. prior suggested valuations for Kioxia or our prior assertions that a valuation of ~$10B should be considered a bottom for WDC’s flash segment.” Regarding the to-be-separated off Western Digital SanDisk flash+SSD business, Bryson sees “the expected split of the company as creating substantial value for shareholders at a valuation commensurate with Kioxia’s IPO.”
…
Netlist has won a $118 million damages award against Samsung Electronics, Samsung Electronics America, and Samsung Semiconductor (together “Samsung“) in the US District Court for the Eastern District of Texas. The award resulted from a jury trial which involved three Netlist patents: U.S. Patent Nos. 7,619,912, 11,093,417 and 10,268,608. The infringing products were all Samsung DDR4 RDIMMs and DDR4 LRDIMMs. Netlist filed the complaint against Samsung in August 2022. The federal jury’s unanimous verdict confirmed that all three Netlist patents had been infringed by Samsung, that none of the patents were invalid, that Samsung willfully infringed those patents, and that money damages were owed to Netlist for the infringement of all three patents.
In April 2023, Netlist won a $303 million damages award against Samsung, and in May 2024, Netlist won a $445 million damages award against Micron Technology, Inc., Micron Semiconductor Products, Inc. and Micron Technology Texas LLC. C.K. Hong, Netlist’ CEO, said, “In the past 19 months three separate juries have awarded Netlist $866 million in damages for the willful infringement of our patents. These verdicts are among the largest in the semiconductor industry during this period.”
…
Storage supplier Nexsan has signed up CASA Software to distribute its products in the sub-Saharan Africa region. CASA Software says it’s a “ digital transformation organisation comprised of a highly skilled team of technology professionals. The company has over three decades experience in the South African and sub-Saharan ICT industry.”
…
Quantum’s shares have soared 620.8% to $21.77 since Nov 20 and a $3.01 price. Some one is buying lots of its stock to make such a difference. Is a bid underway?
…
David Goeckeler
The Semiconductor Industry Association (SIA) announced Western Digital CEO David Goeckeler has been elected Chair of the SIA Board of Directors. SIA represents 99% of the U.S. semiconductor industry by revenue and nearly two-thirds of non-U.S. chip firms.
…
Distributed cloud object storage supplier Storj launched its Channel First program and portal with:
Simplified deal registration and partner precedence support.
Easy collaboration with clear, predictable pricing and reserved capacity options.
Comprehensive marketing and sales playbooks, including co-branded materials and training.
Integrations with leading technologies, supported by thorough testing and documentation.
Commitment to partner success with no additional costs or minimum thresholds.
…
Zion Gabai
StorONE announced Zion Gabai has joined as VP and GM of International Sales. He has expertise across networking, storage, cloud, big data, analytics, IoT, AI, and security, having worked for Dell EMC and NetApp among other companies. With 25 years experience, Gabai most recently was Senior Director for Western and Central Eastern Europe for Dell EMC and also served as Senior Manager for the strategic solutions group for NetApp.
…
Bulgaria-based Tiger Technology is looking to raise A-round funding to accelerate its growth. The company provides a hybrid multi-cloud file namespace for Windows Servers and enables space-saving file tiering from on-prem servers to cheaper file and object stores with ancillary backup, archive, file sync, business continuity, and disaster recovery benefits. It has several products: Tiger Store on-premises file sharing; Tiger Pool combines multiple volumes into a single pool; Tiger Spaces file sharing among work group members; and the Tiger Bridge cloud storage gateway, syncing and tiering product.
With 65 percent Y/Y revenue growth, 300 percent pipeline increase, and most of the revenue coming from US-based customers, the company says it addresses the cloud storage services at a turning point when it’s expected to reach $137 billion by 2025. Iraván Hira replaced founding CEO Alexander Lefterov in April this year, with Lefterov now the CTO.
…
TrendForce said the global DRAM industry revenue reached US$26.02 billion in 3Q24, marking a 13.6 percent QoQ increase. The rise was driven by growing demand for DDR5 and HBM in data centers, despite a decline in LPDDR4 and DDR4 shipments due to inventory reduction by Chinese smartphone brands and capacity expansion by Chinese DRAM suppliers.
SK hynix reported $8.95 billion in revenue—a 13.1 percent QoQ increase—maintaining its second-place position. Although its HBM3e shipments ramped up, a 1 percent – 3 percent QoQ decline in bit shipments from weaker LPDDR4 and DDR4 sales offset these gains. Micron saw its revenue surge by 28.3 percent QoQ to $5.78 billion, driven by strong growth in server DRAM and HBM3e shipments, which led to a 13 percent QoQ increase in bit shipments.
Nanya Technology faced a more than 20 percent QoQ drop in bit shipments due to weaker consumer DRAM demand and intensified competition in the DDR4 market from Chinese suppliers. Winbond experienced an 8.6 percent QoQ decline in revenue, falling to $154 million, as consumer DRAM demand softened and bit shipments decreased. PSMC reported a 27.6 percent QoQ decline in revenue from its in-house consumer DRAM production.
…
More information has come out from Wedbush analyst Matt Bryson about the separation of Western Digital (WD) from its SanDisk subsidiary which will likely take place in Q1 2025. He predicts some 80 percent of current WD shares will go to the new WD HDD-only business with c20 percent going to the spun-off SanDisk flash+SSD business.
SPONSORED FEATURE: Ransomware is everywhere. The FBI and CISA just issued yet another advisory about it.
The average payouts associated with this type of cyber attack are huge. The 2024 Cost of a Data Breach report – conducted by Ponemon Institute and sponsored, analyzed and published by IBM – pegs it at USD4.88m.
The situation is so bad that the White House has just hosted its second multi-national task force meeting to address the problem. While that’s a laudable effort, it did not offer any concrete solutions to stem the flow of damaging attacks.
User education is often the go-to remedy for ransomware prevention. “Just get them to stop clicking those malicious links”, some experts say. Clearly, that alone isn’t enough. People are aware of the threat but infections and ransoms continue to grow.
What’s needed is a multi-layered defense. For some, that stops at appliances to scan email, police URL access, and monitor client devices. For IBM, that wasn’t enough. It had a bright idea: why not do it in storage?
IBM had already introduced elements of ransomware protection to its FlashSystem NVMe-based flash storage in 2022. IBM FlashSystem integrates with its cloud-based Storage Insights storage management and optimization system, which scans for anomalies and potential threats, enabling an organization to recover data from immutable snapshots in the event of a breach or data corruption.
Snapshots play an important part in this recovery process. The snapshot is a point in time image of a disk’s data that is immutable; it can’t be altered or deleted, and it can’t be directly mapped to a host, providing a reliable source for recovery.
Using its Safeguarded Copy feature, IBM adds the ability for the user to set access controls and retention policies to govern permissions around the snapshot management process. It also features an elevated security mode that requires two people to change or remove Safeguarded snapshots. This separation of duties makes it more difficult for any one person to subvert the system.
In 2022 IBM announced Storage Sentinel, which is a system that complements Storage Insights. It scans snapshots to identify signs of corruption by ransomware. Sentinel tags snapshots to highlight a validated and verified point of restore. Armed with this information, it can help IT staff quickly find clean data copies to restore from without reinserting the ransomware threat. Sentinel is now part of IBM’s software suite for data resiliency, Storage Defender.
The next step: Computational storage
While these features and offerings can help shave valuable time from the recovery process, IBM wanted to go further by moving threat detection as close as possible to the point of the ransomware attack in the storage ecosystem. For that, it turned to its computational storage technology, the IBM FlashCore Module.
There’s only so much extra computational muscle that you can squeeze into a server CPU with each iteration. Storage devices are good places for handling storage-specific tasks, and so moving storage-related computing operations into the FlashSystem’s FlashCore Module seemed like a no-brainer, explains Philip Clark, Program Director for FlashSystem at IBM. This is the idea behind computational storage.
IBM has already moved basic functions such as encryption and compression into the flash drive, offloading it from the storage controller. This can increase the efficiency of IBM’s storage, but it felt that it could go further. IBM wanted to make FlashSystem part of a broader drive for cyber resilience. Why not transfer some of the ransomware scanning tasks into the storage devices themselves?
“The whole area of cyber resilience has really become an important focus not just in FlashSystem, but across IBM,” Clark says. Teams across IBM with functions from security to mainframes share cyber resilience knowledge and technology between them.
Computational storage gave IBM an option to put some added value into flash storage, which was becoming a more commodity product category. “Having some pretty unique technologies to address this has been something that stood out as not just the run of the mill technology,” Clark says, explaining that it has moved the endgame for computational storage beyond mere I/O efficiency.
“We’ve gone beyond speeds and feeds to having a much broader story.”Scanning incoming data for ransomware signals was an obvious choice. It allowed IBM to look for digital toxins at the block level rather than the file level targeted by more traditional malware and ransomware scanning solutions. “Where we’re doing it, at the block level in the raw operating system, is unique,” he says. “We’re not just reading the bits and then comparing it to an existing database scanner, we’re processing each IO pattern in real time, right as they’re coming in.”
Watching for suspicious bits
IBM’s Storage Insights observability platform already had the ability to detect some suspicious signals by looking for changes in compression and entropy statistics. In February 2024, IBM enhanced the ARM-based FlashCore Module in the FlashSystem to power its inline ransomware threat detection capability.
Hardware-assisted computational storage makes it easier to manage ransomware scanning across a growing storage ecosystem. Scaling out traditional server-based file system-level scanning can mean adding more of those servers to the rack. The FlashSystem storage hardware scales its scanning capabilities automatically because every additional drive comes with its own computing capabilities built in. Even though these drives search for ransomware anomalies independently, they can be combined into a single management system for visibility and convenience, Clark adds.
The ransomware detection algorithm is based on machine learning. While the sophisticated AI model trains on IBM’s servers, the inference model runs entirely in the FlashSystem hardware. The FlashCore Module collects and aggregates samples of what’s happening to the data, passing it to the inference engine every two seconds. This means it can trigger a ransomware alert after six samples, which translates into raising the alarm in as little as twelve seconds of detecting a ransomware attack.
IBM regularly updates the inferencing model automatically, or on demand, as it retrains the data on new emerging malware patterns. The AI isn’t scanning for individual ransomware hashes. Instead, it detects patterns associated with ransomware activity on data, even if it hasn’t seen the specific ransomware before.
If the inference model detects false positives, it will send information about that back to the model for further training, but that data isn’t an actual file with business-related content. Instead, it’s statistical data about block-level activity that enables IBM to update the training model without compromising client privacy.
Sounding the alarm early
This in-drive ransomware scanning function doesn’t need to replace traditional file system-level scans. It’s a different animal altogether. File system scans have the advantage of context, because they can focus on file-level content and metadata. What block-level scanning lacks in that area it makes up for in responsiveness. Together, the two form a powerful anti-ransomware proposition.
Companies can take days or weeks to find out about an attack that is identified at the file system level. Introducing another layer of defense closer to the storage and scanning at a lower level demands a different kind of scan that doesn’t rely on the context of a file. It heightens sensitivity to attacks and increases the chance of catching a nascent ransomware threat. “An early warning system is ideal,” Clark says.
That early system is all very well, but only if the warning goes somewhere and something gets done. Integration with external systems is key, and IBM accomplishes this in a couple of ways. At the most basic level, it can be integrated with anything that supports syslogs, meaning that any tool supporting these can read FlashSystem’s warnings about malicious data.
However, IBM’s integration with Storage Insights and Storage Defender means that administrators can create automated recovery processes when FlashSystem triggers an alert. Storage Insights is engineered to restore a snapshot quickly to minimize downtime from a malware infection. The security team still has to contain and eliminate the infection, but the storage software also provides integration opportunities with other third-party tools to help facilitate that process.
Webhooks from Storage Insights enable other programs to access its alerts in near real time. IT Service Management Tools can subscribe to these, giving them structured information about block-level events that can feed straight into their monitoring and operations systems. Webhooks enable FlashSystem to talk with a range of systems, ranging from SIEMs (including IBM’s own QRadar) to file scanning tools, to surface suspicious events as they happen.
Ransomware recovery in action
This in-drive detection capability surprised Sam Wheatley, a technical presales consultant at Swedish value-added distributor TD Synnex. He took the FlashSystem model 5300 for a spin, loading up a virtual machine with PDF and Excel files and then letting the REvil ransomware loose on the sandboxed system. He noticed alerts lighting up Storage Insights right away with reports of mass decompression and encryption activities.
“With fast ransomware threat detection alerting, you have a chance to save data before it gets encrypted,” he says. “Imagine the relevant data that you could save instead of having to restore terabytes of potentially infected data after the fact in attempt to find it.”
In the fight against ransomware, the time it takes to detect an attack can impact the cost and effort of remediation. The closer you can get to the point of malicious encryption and take action, the fewer headaches you’ll have later.
Computational storage is a novel way to close the gap to the malicous encryption point. Its integration with the rest of the storage management ecosystem, and beyond, makes it possible to action automated responses along the incident response chain.
Will we eliminate ransomware as a major threat anytime soon? Keep wishing. But at least with more responsive detection systems, businesses can mitigate the impact of the threat when it does strike.
Quantum has launched a high-end DXi9200 target backup appliance as its sole hybrid (flash+disk) hardware accompanying the two all-flash T-series and two V5000 virtual appliances.
As a consequence, its DXi datasheet no longer features the previously available hybrid DXi4800, DXi9000, and DXi9100 appliances, with the new generation DXi9200 taking over their roles. Quantum reckons that, with the advent of the all-flash T-Series, which provide “ultra-fast ransomware recovery,” the role of hybrid appliances is evolving to use cases where cost efficiency is a primary concern. The DXi9200 has NVMe SSDs to store metadata and 20 TB disk drives to store backup data. It offers double the price/performance of the prior DXi9100, we’re told, with more than 30 percent faster ingest and restore performance. It features up to 68 percent denser packaging, with 2.2 PB usable capacity in 12 rack units, and it also has 25 percent lower power consumption.
Quantum DXi9200
Sanam Mittal
Sanam Mittal, Quantum’s VP for DXi, said in a statement: “The DXi9200 is a powerful new solution for strengthening any organization’s cyber resilience. Coupled with highly optimized data reduction, replication, and cloud tiering, plus all-inclusive software, capacity-on-demand licensing, and flexible as-a-Service subscription service options, the DXi9200 dramatically lowers costs and increases IT efficiency.
“DXi9200 is the ideal choice for enterprise backup and recovery services, consolidation of offsite immutable copies for disaster recovery and long-term retention, and as the central hub of modern edge-core-cloud data protection fabrics.”
It scales capacity in 55 TB increments from a 110 TB usable entry level up to 2.2 PB. It can achieve up to 462 PB logical capacity, assuming a 70:1 data reduction rate and Cloud Share tiering to public and private clouds.
The DXi9200 system has protection against unauthorized access with secure connectivity to backup software, replication partners, and cloud tiering destinations based on at-rest and in-flight data encryption, secure multi-factor authentication, and role-based access control. There is offsite protection of data copies through bundled replication, cloud tiering to public and private clouds, cooperating DXi virtual and physical appliances, and Direct-to-Tape capabilities, for disaster recovery.
Quantum’s DXi product range. The now superseded hybrid products are shown with a gray background
Monitoring and alerting capabilities include real-time status and anomaly detection of critical events or irregularities. Backup data integrity against alteration, deletion, or corruption is enhanced with offline immutable snapshot data copies, frequent native data integrity health checks, parity-protected RAID, and a compatible ecosystem of data protection and malware scanning software, including Veeam, Veritas, and Commvault.
Get a DXi range datasheet here. DXi9200 appliances are available immediately, either as a capital purchase or via the Quantum GO pay-as-you-go subscription offering.
Hammerspace claims it has set new records in the MLPerf Storage 1.0 benchmark with its Tier 0 storage beating DDN Lustre more than 16x at the same client count.
The MLPerf Storage 1.0 benchmark is produced by MLCommons and v1.0 results were made public in September. The benchmark tests combine three workloads – 3D-Unet, Cosmoflow, and ResNet50 – and two types of GPU – Nvidia A100 and H100 – to present a six-way view of storage systems’ ability to keep GPUs over 90 percent busy with machine learning work. Hammerspace scored 5,789 MiB/sec on the 3D-Unet workload with 2x H100 GPUs, and 28,826 MiB/sec with 10x H100s.
DDN scored 99,018 MiB/sec with 36x H100s, using an EXAScaler A1400X2T system with 18 clients. You can see the benchmark table here. It has a # Accel column, which is the number of GPUs (accelerators) supported. There is an edited extract of it shown below, with the DDN and Hammerspace entries in it:
Since then, Hammerspace has refined its Global Data Platform technology to embrace locally attached NVMe SSDs in the GPU servers. It calls them Tier 0 storage and uses them as front end to external GPUDirect-accessed datasets, providing microsecond-level storage read and checkpoint write access to accelerate AI training workloads. It has now repeated its MLPerf Storage benchmark runs with this updated Tier 0 software and achieved significantly better performance, as its chart shows:
The Hammerspace results have been submitted but not verified by the MLCommons MLPerf organization and should be regarded as tentative.
Hammerspace says its single client system using its Tier 0 software supported 33 accelerators compared to its DDN rating of 36 for its 18-client system. With 18 clients, Hammerspace says its extrapolated results supported 594x H100 accelerators – a great deal more than DDN, or anyone else for that matter.
David Flynn.
David Flynn, founder and CEO of Hammerspace, stated: “Our MLPerf 1.0 benchmark results are a testament to Hammerspace Tier 0’s ability to unlock the full potential of GPU infrastructure. By eliminating network constraints, scaling performance linearly and delivering unparalleled financial benefits, Tier 0 sets a new standard for AI and HPC workloads.”
Hammerspace claims its MLPerf testing also showed that Tier 0-enabled GPU servers achieved 32 percent greater GPU utilization and 28 percent higher aggregate throughput compared to external storage accessed via 400GbE networking.
Overall, it claims that extrapolated results from the benchmark confirm that scaling GPU servers with Tier 0 storage multiplies both throughput and GPU utilization linearly, ensuring consistent, predictable performance gains as clusters expand.
DDN’s SVP for Products, James Coomer, told us: “We’re proud of our proven and validated performance results in the MLPerf Storage Benchmark, where the DDN data platform consistently sets new records for throughput and efficiency in real-world machine learning workloads. With the release last week of our next-generation A³I data platform, the AI400X3, that drives performance up to 150GB/s from a single 2RU appliance which supports performance of up to 150 simulated GPUs, we’re reinforcing our commitment to delivering innovative and high-performance data solutions.
“While we welcome innovation across the storage landscape, our results—unlike speculative or extrapolated claims—have been rigorously tested and verified by MLCommons, highlighting our commitment to transparency and excellence.”
Interview: PEAK:AIO says on-prem small language model (SML) AI training is practical for specific subject areas because it serves locally stored data to GPU servers.
Contrary to a common idea that Gen AI model training needs access to hundreds if not thousands of GPUs, PEAK:AIO argues that this level of GPU support may be necessary for generalized large language model (LLM) training but is definitely not needed when training models for specific subject areas – such as interpreting scans in the health market. And it has customers doing exactly this.
This became clear during a briefing with founder Mark Klarzynski, based in the UK, and US-based president and CEO Roger Cummings. A big picture view here is that PEAK:AIO developed and has proved its technology in the UK and is now taking it to the much larger US market and growing fast there. The software-based technology features re-written NFS stack and Linux RAID code to decrease the latency and increase the speed for NFS data transfers to a GPU server. In effect a PEAK:AIO user gets parallel file system access speed while using standard NFS and not a complex and HPC-style parallel file system with a possible need to re-factor NFS-using applications.
PEAK:AIO’s software can get a small 1RU server with a PCIe 5 bus sending 80GB/sec of data to a single GPU client server for AI processing.
We started out the interview session by asking if PEAK:AIO’s technical lead is sustainable.
Blocks & Files: PEAK:AIO is basically selling an NFS box built on commodity servers and so competes with every other NFS-based file system. Then you’ve got this highly specific, AI-focused data feeding, which gives you entry into a niche. However, there’s an issue in that many of the large file companies are supporting GPUDirect and saying they pump data fast to GPU servers. So up you come along with PEAK:AIO and say, yes, but we pump it twice as fast, three times as fast, five times as fast. Is this a sustainable advantage that you’ve got?
Second, Hammerspace has just come up with its Tier 0 concept, which is feeding data from external storage into a GPU server’s locally attached NVMe SSDs, and then feeding data from them to the GPUs at much higher speed than GPUDirect. How does PEAK:AIO’s technology relate to this Tier 0 idea of Hammerspace?
Mark Klarzynski.
Mark Klarzinski: Yes, we just made everything fast by getting close to the actual hardware, close to the drives, close to the Mellanox stack, all that sort of stuff. And that makes us super fast, which also makes us super fast for people like Hammerspace. I’ve known David Flynn for many, many years, and we made a specific version for their Flex Files [see bootnote].
We actually rewrote some parts of the NFS stack so that we work just that bit better. Because, of course, Hammerspace … still needs NFS to scale, and they don’t really do that. They rely on other people. … So we have a version, and they’ve just released [software which] actually led to a mini white paper with us, a solution with us. We work a lot with Hammerspace.
What I would say about the tier 0 idea is that, actually we did this three years ago, and it got messy. It sounds really great, okay, but it’s not replacing GPUDirect. It’s not bad, and maybe they’re doing it better. They’re actually saying that, when you buy a DGX or HGX, you’ve got some storage in it, why not use it? And they do, but then include it in a Hammerspace layer. That’s actually part of the kernel stack anyway. The difficulty with it is, is the way Nvidia works … it gets messy because that’s on a RAID server.
The entire Tier 0 storage has one purpose for the GPUs, which is scratch-based. … Maybe they’ve done this better, but Ceph has been doing this for a while. We did it in the early days, and what we decided to do was to let the Nvidia use their local storage for what they designed it for. And, I’m honest, we don’t see DGXs going out with that much local storage. Normally, we’ve got a handful of terabytes, so pulling it into the stack is not really adding much.
Maybe, if you’ve got 10,000 GPU servers, [it] makes a difference, but with the people that we deal with, it became more of a headache, because you’re constantly dealing with Nvidia doing updates and playing around with their RAID server.
If you make that part of your virtualized stack, it gives you problems because every time Nvidia makes an update, you then have to work with your customers to decide whether you’re going to implement it or not.
Mark Klarzynski: We work amazingly on that. What separates us from every other NFS server is not just that we’re rocket fast and very focused on that GPU world. We realized that energy density was a bigger problem, and getting what would traditionally be in, say, six or ten RU down to 2RU was what we wanted to do for pricing, which also meant that we reduced the energy and increased the density.
We’ve been working with Solidigm for the last year to get six petabytes in 2RU and doing some quite funky stuff on the byte path. What we’ve really been doing with that is not just using the endurance and speed of SLC to buffer QLC [in the drive]. We’re actually using it to stage data so that we can power down the QLC and reduce the power needed.
When you’ve got a few hundred or thousands of these drives, and most of the time they’re in idle state. That actually has a dramatic difference on the amount of power. And just about every datacenter we walk into these days is really struggling with power. Because the one thing GPUs do is they get bigger and they just demand more power, and there’s no stopping the GPU, so there’s only the storage that can come down on the power consumption.
I’d love to say this was an amazing strategy of mine, but, but in honesty, it just sort of got organically adopted. We found ourselves in large labs in the USA, like Los Alamos and National Labs, etc. Do you know that we’ve worked with HPE, working with Dell, etc., and that’s why Roger came on because that was certainly beyond the scale of me. I’m sort of the wacky guy at the back end who just comes up with ideas.
Mark Klarzynski: Yes. We were also looking at the 256 last week.
Blocks & Files: I’ve got the impression from some of the material I’ve read about PEAK:AIO, that a typical customer has single digit GPU servers, not 1000s of GPUs. And that suggests to me that you’re used mainly for AI inferencing rather than AI training.
Mark Klarzynski: That would seem the case. But let me give you some examples. Possibly the largest medical installation in the world, King’s College London, is running on six DGXs, actually similar to Serological Society of London’s only two at this moment. Most of those really large DGX clusters are in hyperscalers. … I can think of two clients that I know of that actually, genuinely have a SuperPOD themselves, and that’s not our play.
We distribute through PNY, who are one of Nvidia’s main distributors. Something like 87 percent of their customers are below 10 DGXs. A handful are bigger, and most of those are GPU-as-a-service type people.
What we what we really see, and this has been our focus, is that we see AI going more horizontal than up. So for instance, when we think about King’s College London, and I stay with healthcare, just as an example, they create an amazing model on brain scans. For instance, that spurs another five institutions to do their [own] specific training on that model.
Blocks & Files: This is lower scale training than 100,000 GPUs building a new OpenAI version. I didn’t realize this was going on.
Mark Klarzynski: It’s gigantic. So, we’re in the UK. Every university that I know has probably ten projects, if not more, that have all got less than ten DGXs and they’re training models.
Roger Cummings.
Roger Cummings: This week at the Supercomputing Show was very eye opening for us. Now that I’m on board and talk to some of the folks in the US, they’re desperately searching for a solution that can help them with the 20 or under kind of scenario of GPUs. We talked to one research institution that was supporting 1,000 or 1,100 labs, and each one of them had maybe one or two GPUs, but there was not a solution he could offer them.
Blocks & Files: Is this like training models for specific niches? This is not massive, OpenAI, Chat GPT-type training, generalized training. This is specific?
Roger Cummings: Yes. That is the work that we’re doing with some of the – staying in the healthcare – work that we’re doing on MONAI, the organizations working with medical imaging. You can imagine having data from an MRI machine and running algorithms against that. There’s a lot of applications for that. It is a horizontal use case across many, many healthcare institutions, life science institutions. That’s a great example of this kind of horizontal growth that Mark is referencing.
***
Part 2 of this interview will follow in a second article.
Bootnote
Hammerspace uses NFS v4.2 and its client has support for parallel NFS (pNFS) using a Flex Files feature. This has parallel read and write access to the data servers (cluster nodes).
Panmnesia says it raised over $60 million in Series A funding as it advances its fabless CXL chip designs for the datacenter and AI workloads.
The company has now raised a total of $80 million since its inception in 2022, and says its valuation now stands at over $250 million.
A total of 15 investment institutions participated in the Series A funding, led by InterVest, and joined by new investors such as Korea Investment Partners, KB Investment, Woori Venture Partners, BSK Investment, Nvestor, Murex Partners, Daesung Private Equity, and TS Investment. Existing investors including Daekyo Investment, GNTech Venture Capital, SL Investment, TimeWorks Investment, Quantum Ventures Korea, and Yuanta Investment Korea also participated in this round.
Panmnesia is currently focusing on developing semiconductor IP and switch chips for CXL (Compute Express Link), the next-generation memory interconnect technology. CXL 3.0 technology enables a pool of external memory, accessed through CXL switches, to be shared between host computers with coherent caches and the CXL memory endpoints or expanders.
Myoungsoo Jung
Panmnesia introduced the world’s first CXL full-system framework, including the CXL Switch chip, at the 2022 USENIX Annual Conference. More recently, at the OCP/OpenInfra joint event last month, the company said it “surprised the industry” by unveiling the world’s first CXL 3.1 IP with two-digit nanosecond latency, and the world’s first CXL-GPU designed for AI workloads.
Panmnesia was founded in 2022 by CEO Myoungsoo Jung, a tenured professor at KAIST (Korea Advanced Institute of Science & Technology) in South Korea, along with other researchers there. “We will allocate this investment to successfully complete key projects, such as manufacturing of the high-bandwidth CXL 3.1 Switch silicon chip, advancement of next-generation CXL IP, and collaboration projects with global IT companies,” said Jung.
