Fast access all-flash array vendor Pure Storage has indicated an interest in long-term archival storage by investing in ceramic storage startup Cerabyte and gaining a board seat.
Cerabyte is developing archival ceramic-coated glass tablet storage with binary ones and zeros indicated by the presence or absence of femtosecond laser-punched pits in the ceramic layer. These are written as QR codes and read by scanning microscopy. The glass tablets are stored in a robotic library system, akin to a tape library. Unlike tape media, which needs replacing every decade or so, the data on these glass tablets is immutable, and the tablets can hold more data than a tape cartridge.
Pure is making a strategic investment in Cerabyte – the amount has not been disclosed – and John Colgrove, Founder and Chief Visionary Officer at Pure Storage, is joining Cerabyte’s board.
John Colgrove
Colgrove said in a statement: “Pure’s investment in Cerabyte and joint partnership will allow us to offer our customers sustainable and immutable data storage solutions that are revolutionizing the industry. By disrupting the archival storage market, we are paving the way for longer lasting and easier to manage long-term storage.”
Cerabyte is based in Munich. It has developed a functioning end-to-end data prototype system and aims to build a commercial product. Cerabyte set up a US operation in 2023 and recruited Silicon Valley-based Steffen Helmold as director. Earlier this year it opened two US offices in Santa Clara, California, and Boulder, Colorado.
The US Cerabyte operation has landed a big fish in the shape of Pure Storage. Pure says of the move that it’s delivering the industry’s best platform to store, manage, and protect the world’s data. This strategic investment in Cerabyte will enable the company to extend this mission further. Cerabyte claims its tech can store all data virtually forever and preserve today’s digital records for future use through the use of its ceramic data storage.
Cerabyte prototype system showing laser optics
John Monroe, Chief Analyst at Furthur Market Research, said: “The complex global needs of zettabyte-scale archival storage have been poorly served with expensive solutions that consume an inordinate share of the world’s available energy. The storage industry is ripe for transformative disruption. In concert and conjunction with tape, new technologies such as Cerabyte’s will be required to provide viable and cost-effective solutions to enterprise customers’ crucial challenges with the security, immutability and sustainability (SIS) of their vital data.”
Cerabyte cartridge and ceramic-glass media
Cerabyte CEO Christian Pflaum said: “As the industry is heading towards the Yottabyte Era, sustainable data storage – which eliminates the need for data migration and thereby scales down the energy footprint and TCO – will be critical to harness the data tsunami ahead. We are thrilled to partner with Pure Storage to commercialize ceramic data storage and welcome John to our Board of Directors.”
The whole thrust of Pure’s Cerabyte involvement is to have a glass tablet-based archival storage library that it can sell to its customers, thus disrupting the archive tape-library market. This is a radical move, and would enable Pure to fulfill all the storage hardware system needs of its customers, from real time (FlashArray), near real-time and near line (FlashBlade) to long-term archives. We are looking at a multi-year effort, we think, with a possible demonstration product in 2026 or so. It’s great news for Cerabyte and will prompt tape library system vendors like IBM, Quantum and SpectraLogic to take stock of their tape-based technology and ponder a media evolution.
Micron has announced the 2650 client SSD, its first product built from 276-layer 3D NAND, 18 months since it launched the 2550 client SSD using 232-layer NAND.
The 276 layers are a record for Micron, which presents this as its 9th generation 3D flash. The 2650 is a TLC (3bits/cell) product and uses a PCIe gen 4 interface, like the 2550. It has capacity points ranging from 256 GB to 1 TB and Micron claims it offers best-in-class performance.
MIcron G9 276-layer 3D NAND chip and 2650 M.2 variants
The Gen 9 NAND is claimed to have the industry’s fastest IO speed at 3.6 GBps, beating competitive products from SK hynix, Solidigm, Kioxia, Western Digital, and Samsung with up to 99 percent faster read and 88 percent better write bandwidth. Micron also claims that its Gen 9 NAND is up to 73 percent denser and 28 percent smaller in board size than their products.
3D NAND supplier layer counts. The italicized Micron internal Gen 8 and Gen 9 entries will be branded Gen 10 and 11 externally
The 2650 comes in the M.2 gumstick format, with 2230, 2242, and 2280 sizes. Micron hasn’t disclosed its IOPS and bandwidth numbers, but has revealed PCMark 10 benchmark results versus products from SK hynix (BC901), Kioxia (BG6), Western Digital (SN740), and Samsung (PM9B1):
It claims up to 38 percent better average score, 36 percent average bandwidth, and 40 percent faster access time.
This 2650 contrasts with Micron’s 9550 datacenter SSD, which was launched a few days ago. It uses Micron’s Gen 8 232-layer NAND and, with its PCIe gen 5 interface, has very high performance numbers.
We asked Micron some questions about the new 276-layer NAND.
Blocks & Files: What was the thinking behind such a relatively modest layer count increase, from 232 to 276 layers?
Micron: The optimal customer design points, not layer count, drove our decisions for Micron G9 NAND. Primary among these was the need to create a 1 Tb die that fits in an 11.5 x 13.5 mm BGA package while delivering 3.6 GBps. This focus helped us deliver breakthroughs in performance and package size along with 44 percent higher bit density compared to our prior generation. These advancements collectively contribute to a superior NAND solution that addresses the needs of our customers.
Blocks & Files: Is the die single or multi-stacked (two or three stacks)?
Micron: Our G9 TLC NAND uses a single die and class-leading six-plane architecture, allowing us to deliver the necessary performance and requisite costs to meet our customers’ needs.
Blocks & Files: Has it been scaled down in physical size from the 232-layer NAND product, and, if so, how?
Micron: We maintain our best-in-class 11.5 x 13.5 mm BGA package size, which uses up to 28 percent less board area compared to the larger 12 x 18 mm package still being used by most of our competitors. This enables optimal PCB layout, signal routing, and thermal performance, which is especially important for newer SSD form factors, like E1.S and E3.S.
Blocks & Files: Does Micron anticipate producing a QLC version of the G9 276L die?
Micron: While we generally do not comment on future announcements, we maintain our commitment to QLC NAND innovation in the industry.
According to Forward Insights, Micron’s mix of QLC NAND bit shipments in CQ1-24 was the highest among major NAND providers.
Hazelcast is adding vector search capability to its flagship Hazelcast Platform, which combines distributed compute, in-memory data storage, and integration for enterprise AI applications.
GenAI technologies, including large language models (LLMs), are transforming the landscape for data storage suppliers like databases, data lakes, and lakehouses, with virtually all of them scrambling to adopt vector storage and search as these underlie the semantic search used by LLMs. Hazelcast says its vector-enhanced v5.5 product can “generate vector data structures and embeddings from text plot summaries.” It now “enables enterprises to deploy a high-performance end-to-end pipeline to query structured and unstructured data.”
Adrian Soars
Hazelcast CTO Adrian Soars said in a statement: “The integration of vector search in Hazelcast Platform provides the core functionality and foundation upon which developers can modernize business-critical applications and innovate for the AI era.”
The company claims the Hazelcast Platform provides significant performance gains over most competitors, especially when factoring in vector embeddings and retrievals. It outperformed competitors in Hazelcast internal benchmark tests of 1 million OpenAI angular vectors, “consistently delivering single-digit millisecond latency when uploading, indexing, and searching vectors with 98 percent precision.”
In Hazelcast’s view, vector search is “applicable to all industries.” In particular, it can immediately benefit transaction authorization applications. For example, in financial use cases such as know-your-customer (KYC) and anti-money laundering (AML), “vector search can augment and expedite the verification process with semantic search across text, imagery, and other sources to improve the accuracy and speed of determining whether a transaction is legitimate or fraudulent.”
The company points out that three-quarters of the world’s payments pass through Hazelcast-powered applications.
The company has added two more updates to its product:
Jet Job Placement Control enables customers to separate the compute functionality of Hazelcast Platform nodes from the data store component to provide further flexibility and resilience for compute-intensive workloads.
Client Multi-Member Routing improves resilience, performance, and control for applications connecting to geographically dispersed clusters.
It’s also introducing a three-year, long-term support (LTS) deal to help customers with simplified upgrades.
Hazelcast Platform 5.5 is generally available today with the vector search integration available as a preview in the Enterprise Edition with projected availability for production use in Q4.
Both IBM and Pure Storage make all-flash storage systems with their own proprietary flash drives. How do they stack up?
The two storage stars have rejected commercially available SSDs, choosing instead the expense of developing their own flash drive hardware and software to gain flash capacity, endurance, and operational advantages. In general, IBM’s flash drives are drive controller-centric while Pure’s drives, by contrast, are relatively dumb – for want of a better word – at the drive controller level, but intelligent at the system level.
IBM’s drive is called the Flash Core Module (FCM). It’s currently in its fourth generation and available with 4.8, 9.6, 19.2 and 38.4TB raw capacities inside its 2.5-inch, 15mm enclosure. These drives have a PCIe 4 interface.
Onboard hardware compression is used to give the 4.8 and 9.6TB models an effective 21.99TB capacity, the 19.2TB drives become an effective 43.98TB drive, while the top-end 38.4TB product is a 87.96TB drive with compression, 2.3x larger.
FCM generations 1–3 use floating gate NAND technology whereas FCM 4 moved to charge trap technology (Micron 176-layer NAND dies).
Inbuilt Ransomware Threat Detection “is a process that identifies and responds to security threats before they can damage data or systems. The FCM 4 collects detailed statistics on every I/O operation (IOP) for each virtual disk (Vdisk). This data is then intelligently summarized for efficient processing. The FCM 4 transmits this summary to Storage Virtualize, which relays it to an AI-powered inference engine. This engine can identify unusual activity, like potential ransomware attacks, in under a minute. Upon detection, an immediate alert is sent to IBM Storage Insights Pro, allowing for swift action. Additionally, the information can be shared with IBM Storage Defender if available, further strengthening your security posture.”
Pure Storage 150TB DFM teased at Pure Accelerate earlier this year
Pure Storage builds DirectFlash Modules (DFMs) and mounts up to four of them on storage blades. The DFMs use QLC flash dies. In 2022, a FlashBlade storage blade had DDR4 memory DIMMs, PCIe 4 connections to the DFMs, and a 100GbitE data plane. A Futurum report states: “FlashBlade includes inline compression for all data stored on an individual blade done in software. Pure Storage expects a 3:1 reduction across most file data with exceptions in some vertical market data types. Compression is handled differently between the S200 blade and the S500 blade. The S500 blade reduces compression to reduce computational overhead and increase performance.”
Pure claims its system-level management of DFMs can improve reliability up to 5x compared to conventional SSDs, and extend the drive’s lifetime. Garbage collection, sparing, and wear levelling are all carried out at a system rather than drive level, “eliminating duplicate efforts and processes that happen across every drive in a traditional system.”
IBM FCM Internals graphic
This means that its DFMs don’t need much DRAM as they “are effectively just a collection of flash chips with FTL done at the system-wide level in Pure’s controller and its software, instead of at the SSD level as in other all-flash array storage systems.”
It also claims its DFMs “consume from 39 percent to 54 percent fewer watts per terabyte than our closest competitors today.”
Pure Storage was shipping 24 and 48TB DFMs in 2023 and added a 75TB drive earlier this year, with 150TB drives due to follow them in a year or less time – possibly later this year – and 300TB drives after that, probably during 2026. Its roadmap envisages 300 and 600TB DFMs, even extending to 1.2 petabyte drives. A supplier may introduce a 128TB SSD at the August FMS 2024 event. Pure Storage will beat that at the raw capacity level with its 150TB DFM.
IBM has not published a public FCM roadmap. It has historically doubled capacity, and on this basis the next capacity point will likely be 76.8TB with a maximum effective capacity of 230.4TB, assuming no compression advances or other added data reduction methods.
The Pure FlashArray DFMs do not have onboard compression. Instead, Pure arrays, like its FlashArray, provide data reduction from their Purity OS and have a system-level compressing FPGA (DCA card) to do compression work.
Pure’s data reduction includes thin provisioning, zero detection, and unmap. Its FlashArray is claimed to deliver “an industry-leading 5:1 data reduction with a total efficiency of 10:1 (including thin provisioning).” That would mean a 48TB DFM would have an effective capacity of up to 240TB at 5x data reduction.
The FlashBlade//S systems have DFM blade-level compression working at the 3:1 level – rather less.
The DFMs do not have onboard ransomware threat detection. Pure provides immutable snapshots, so data can be recovered if there is a ransomware attack.
Summary
Get an IBM FlashCore Module product guide here. Here’s our summary table of the IBM and Pure Flash drives:
NOR, NAND’s poor relation in nonvolatile memory, could be getting a capacity jump with Macronix’s 3D technology.
Taiwan-based Macronix provides niche NAND and NOR and ROM products and led the NOR market in 2020 with a 24 percent share over Infineon (15 percent) and Micron (5 percent). The NOR market had a 15 percent CAGR last year. Unlike NAND, where the 3D revolution has resulted in vastly increased chip capacity using 200+ layers, NOR had remained in the planar, 2D era.
NOR flash has faster read time than NAND but slower write speed – ~100ns vs TLC NAND’s ~100us – and larger cell size. It is often used for storing application code in systems that need fast app load times, such as smartphones. Macronix said it was working on a 32-layer 3D NOR product back in 2022. Such a product could significantly increase NOR die capacity.
Macronix marketing VP Anthony Le told the EE Journal that Macronix was developing a 32-layer 3D NOR product with up seven times greater density than planar 2D NOR. Its target market was the non-volatile memory product one in the embedded, industrial and automotive areas.
We understand that the starting point for this 3D NOR is a semiconductor AND circuit with bit (cell) level access and a gate-all-around design. It fabricates a multi-layer structure of oxide and nitride layers, etching holes between them and then filling the holes with tree so-called plugs; two N+ doped and one silicon nitride (SiN) insulator column between them with a split channel taking up the rest of the hole space.The two N+ doped plugs act as the source and drain-connected bitline channels for the stacked transistors in the structure. They are orthogonal to the word lines.
It showed an updated version of this 3D NVM diagram at the 2023 FMS event and is using the same diagram for its FMS 2024 teaser webpage. Macronix will update attendees on its 3D NOR progress at this Santa CXlara event.
It has already been sampling 1, 2, 4 and 8Gb 3D NOR product as a 2023 chart shows:
We understand that its latest 3D NOR product features:
32-layer 3D NOR – single-die of 1Gb, 2Gb or 4Gb
100,000 program/erase cycles – similar to existing 2D NOR devices
AECQ-100 (automotive standard) reliability and ISO 26262 ASIL (Automotive Safety Integrity l Level) B compliance
Latency of ~100ns
400MB/sec throughput (200MHz DDR)
Low power draw – either 3V or 1.8V
QSPI and Octal interfaces
Backwards compatibility with 24BGA-package flash
We expect this product to be sampled later this year with possible general availability in 2025.
IBM revenues for the second 2024 quarter rose 2 percent year-on-year to $15.8 billion, beating Wall Street estimates, as its hybrid cloud and AI strategy paid off with a 7 percent rise in software sales helping to lift profits 16 percent to $1.83 billion.
The software business pulled in $6.74 billion, consulting revenues languished in comparison at $5.2 billion, down 0.9 percent. The smaller infrastructure business brought in $3.65 billion, up 0.7 percent on the year, with financing earning $169 million, 8.3 percent down on the year.
There is a pronounced seasonal pattern with Q4 peaks evident in the Software and Infrastructure segment
CEO Arvind Krishna said in prepared remarks: “We delivered a strong quarter, exceeding our expectations, driven by solid revenue growth, profitability, and cash flow generation. We had strong performance in Software and Infrastructure, above our model, as investment in innovation is yielding organic growth, while Consulting remained below model. Our results underscore the continued success of our hybrid cloud and AI strategy and the strength of our diversified business.
”Technology spending remains robust as it continues to serve as a key competitive advantage allowing businesses to scale, drive efficiencies and fuel growth. As we stated last quarter, factors such as interest rates and inflation impacted timing of decision-making and discretionary spend in Consulting. Overall, we remain confident in the positive macro-outlook for technology spending but acknowledge this impact.”
Financial Summary
Gross profit margin: 57.8 percent, up 190 basis points year-on-year
Operating cash flow: $2.1 billion, down $0.6 billion year-on-year
Cash, restricted cash and marketable securities: $16 billion, up $2.5 billion from year end 2023
Debt: $56.5 billion, flat year-on-year
EPS: $1.96
IBM has “infused AI across” its business, with Krishna saying: “In software, our broad suite of automation products like Apptio and watsonx Orchestrate are leveraging AI and we expect to do the same with HashiCorp, once the acquisition is complete. Red Hat is bringing AI to OpenShift AI and RHEL AI. In Transaction Processing we’re seeing early momentum in watsonx Code Assistant for Z. In Infrastructure, IBM Z is equipped with real time AI inferencing capabilities. And in Consulting, our experts are helping clients design and implement AI strategies.”
Its hybrid cloud area is now AI-centric as well: ”Hybrid cloud remains a top priority for clients as flexibility of deployment of AI models across multiple environments and data sovereignty remain a key focus.”
Generative AI is surging, claims the biz: “Our book of business related to generative AI now stands at greater than $2 billion inception to date. The mix is roughly one-quarter software and three-quarters consulting signings.”
This GenAI interest in its customer base is driving IBM’s compute (z16 mainframe) and storage business: “z16’s Telum processor is a unique differentiator, driving real time, in-line AI inferencing at unprecedented speed and scale for applications like real-time fraud detection. Our storage offerings are also benefiting from generative AI as clients address data readiness and need high-speed access to massive volumes of unstructured data.”
SVP and CFO Jim Kavanagh, using constant currency percentages, said: “Software grew by 8 percent, with solid growth across Hybrid Platform & Solutions and Transaction Processing, and strong transactional performance. Infrastructure had great performance, up 3 percent, delivering growth across IBM Z and Distributed Infrastructure. Consulting was up 2 percent and continued to be impacted by a pullback in discretionary spending.
“Transaction Processing delivered 13 percent revenue growth. This performance demonstrates the innovation and value of our mission critical hardware stack across IBM Z, Power and Storage.”
Tracking IBM revenues by quarter by year shows a decline from 2020 with signs of growth since Q3 2023
The infrastructure business segment is divided into a hybrid infrastructure segment and an infrastructure support (hardware and software) business. The hybrid infrastructure is further divided into the Z mainframe and OS business, and distributed infrastructure: Power server hardware and OS, storage hardware, IBM Cloud IaaS, and OEM asset recovery service.
Hybrid infrastructure revenues were $2.4 billion in the quarter with infrastructure support earning $1.3 billion.
For hybrid, Kavanaugh said: “IBM Z revenue was up 8 percent this quarter. We’re now more than two years into the z16 cycle and the revenue performance continues to outperform prior cycles … Increasing workloads translates to more Z capacity or MIPS, which are up about threefold over the last few cycles.
“In Distributed Infrastructure, revenue grew 5 percent driven by strength in both Power and storage. Power growth was fueled by demand for data intensive workloads on Power 10 led by SAP Hana. Storage delivered growth again this quarter, including growth in high-end storage tied to the z16 cycle and solutions tailored to protect, manage and access data for scaling generative AI.”
As ever with IBM results, we don’t know the storage hardware and OS revenue number, nor the storage software revenues, which would include Storage Scale, Ceph, data protection and so forth, and be included in IBM’s Software segment’s revenue number.
Kavanaugh’s summation of the quarter was: ”We are pleased with our performance this quarter and for the first half, driving confidence in our updated expectations. We are positioned to grow revenue, expand operating profit and grow free cash flow for the year.”
Krishna said: ”Given our first-half results, we are raising our full-year view of free cash flow, which we now expect to be more than $12 billion.” The free cash flow full (FCF) year estimate was previously about $12 billion.
Bootnote
IBM has a strong focus on FCF, the cash remaining after it’s paid for business operations and maintaining its capital assets. FCF is the source of dividends, debt payments, reinvestment in IBM’s business, and acquisitions. Investors and analysts can prefer FCF as a business health metric over profit (net income) because it can be a more accurate measure of financial health.
DRAM and NAND fabricator SK hynix is looking at a US IPO for its Solidigm subsidiary, according to a report in Korea’s Hankyung media outlet.
Update: SK hynix comment added. 29 July 2024.
SK hynix bought Intel’s NAND and SSD division in October 2020 in a two-phase deal. The first saw SK hynix acquire Intel’s SSD business and NAND fabrication plant in Dalian, China, for $7 billion. The second has SK hynix paying Intel an additional $2 billion in 2025 for IP concerning NAND flash wafer manufacture and design, R&D employees, and the Dalian fab workforce. SK hynix named its acquired business Solidigm and has developed and shipped successful products such as the D5-P5336 61.44 TB QLC (4bits/cell) SSD.
Hankyung’s report says that Solidigm has made its first profit after 12 loss-making quarters.
SK hynix has just reported record earnings driven by strong demand for high bandwidth memory (HBM) chips needed by Nvidia for GPU servers. The company is increasing its memory manufacturing capacity, and will invest about ₩9.4 trillion ($6.8 billion) in building an HBM fabrication plant at the Yongin Semiconductor Cluster, located at a 4.15 million square meter site in Wonsam-myeon, Yongin, Gyeonggi Province, Korea. The fab construction will start in March 2025 and finish May 2027. The intention is then to add three more plants one after the other to the cluster.
This represents a significant capital investment and the $2 billion due to be paid to Intel next year adds to SK hynix’s capital outflow in 2025.
VP Kim Young-sik, Head of Manufacturing Technology at SK hynix, stated: “The Yongin Cluster will be the foundation for SK hynix’s mid to long-term growth and a place for innovation and co-prosperity that we are creating with our partners. We want to contribute to revitalizing the national economy by successfully completing the large-scale industrial complex and dramatically enhancing Korea’s semiconductor technology and ecosystem competitiveness.”
Rendering of planned Yongin Semiconductor Cluster
Commenting on the rumoured IPO, Wedbush analyst Matt Bryson tells subscribers: “We see this news as plausible with SK hynix having previously had plans to separate Solidigm, with the entity’s recent recovery making such an option far more viable. We see SK’s success in such an endeavor as likely to be governed by how the new organization is broken out (e.g. what assets are included within Solidigm vs being kept by SK) and how hynix/Solidigm talk to future technology plans, particularly given the current roadmap for NAND at Dalian (including the QLC parts that have enabled Solidigm’s recent success with high capacity enterprise SSDs) seemingly ends at 196 layers.”
A Solidigm IPO would enable SK hynix to receive cash for some of its holdings in the company and help cover the planned capital outgoings. We have asked both SK hynix and Solidigm to comment on the IPO report and will add any responses to this story when they come in.
An SK hynix spokesperson told us: “Solidigm is exploring various growth strategies, but no decision has been made at this time.”
StorMagic has received fresh funding to grow its SvHCI edge replacement product for VMware.
The UK-based company’s SvHCI product combines a hypervisor, virtual networking, and SvSAN virtual storage software. StorMagic has received “a significant investment” to back its “fast-growing, edge computing” business from the Palatine Growth Credit Fund.
Dan Beer
CEO Dan Beer stated: “StorMagic is thrilled to have Palatine’s support as the company recently entered a new market segment with the introduction of SvHCI. Since Broadcom’s acquisition of VMware, edge and SMB customers have seen massive price increases and many are looking for alternative solutions to help them run on-site applications reliably while reducing costs. SvHCI is the ideal replacement solution and can save SMBs and edge customers up to 62 percent over VMware alternatives.”
Palatine is a private equity business with two existing funds. The Buyout Fund invests between £10 million and £30 million in “dynamic and visionary management teams looking to drive their business through their next phase of sustainable growth.” The Impact Fund invests £5 million to £20 million in “commercially driven businesses with a mission to positively impact on society or the environment.”
Its new Growth Credit Fund “supports ambitious and innovative tech businesses based across the regions of the UK.” This, StorMagic told us, is “a brand new fund that recently closed,” and “StorMagic is the very first investment they made from this new fund.”
Palatine says the fund’s strategy “is to help businesses reduce equity dilution as they grow by partnering with fast-growing B2B businesses that are Venture Capital-backed.” We were intrigued by the reducing equity dilution angle so asked some questions about it.
Blocks & Files: How was this Palatine investment non-dilutive?
StorMagic: Palatine has made an investment in StorMagic by lending money (debt), hence this is a non-dilutive way to gain capital for growth. The main reason they invested was their understanding of the market dynamics, the opportunity presented by Broadcom’s changes that have impacted the SMB and Enterprise edge markets, and their belief that our new product direction (including SvHCI – full stack HCI software product) is a great fit for what customers are looking for.
Blocks & Files: How much cash did StorMagic receive?
StorMagic: StorMagic received the full proceeds of the investment. Since we are a privately held company, we are not disclosing the amount – however, it was a significant investment that will help fuel our growth as we develop new products that include hypervisor and virtual networking (along with virtual storage, which we’ve been selling for many years).
Blocks & Files: Does Palatine get board-level representation at StorMagic?
StorMagic: No – they will not be a member of our board.
***
Although Palatine is a private equity business, it is not investing in StorMagic in the normal sense of taking an ownership position. Palatine is acting as a provider of debt funding and the loan package is supposed to help the company grow. The company did not make clear how large the sum is, nor what the repayment terms are.
Hitachi Vantara has announced the general availability of its first Hitachi iQ AI infrastructure offering, enabling its customers to use Hitachi Vantara storage with Nvidia’s BasePOD GPU server system.
The Hitachi iQ AI systems portfolio concept was revealed in March. Four months later, the first deliverables center around Nvidia BasePOD certification, meaning Hitachi Vantara storage hardware and software can be used in mid-range AI workloads. High-end ones need Nvidia’s bigger and more powerful SuperPOD, which is not yet supported by Hitachi iQ. The DGX BasePOD was launched in 2022, including up to nine Nvidia DGX A100 servers with 8 x A100 GPUs, 12 storage servers (from Nvidia partners), and three networking switches. The larger DGX SuperPOD is composed of between 20 and 140 DGX A100 systems.
Octavian Tanase, Hitachi Vantara’s chief product officer, stated: “Nvidia DGX BasePOD certification not only tells you that this solution has met rigorous standards for reliability and performance, but it also represents a significant upgrade to handle the extensive bandwidth and speed required by today’s networks.”
Hitachi Vantara Hitachi iQ diagram
Tony Paikeday, senior director of DGX systems at Nvidia, said: “The combination of solutions built on Nvidia’s DGX platform and software, coupled with Hitachi Vantara’s expertise in AI discovery and planning, will help provide customers with a foundation to turbocharge their generative AI capabilities.”
In Tanase’s view: “Hitachi iQ is equipped to handle the most demanding AI workloads, helping our customers enjoy seamless, high-speed AI operations that keep them ahead of the curve.”
Hitachi Vantara is also launching an AI Discovery Service to help customers identify the most valuable AI use cases, assess their data readiness, determine ROI, and create a strategic roadmap for AI implementation. Customers can select from a range of AI consultative services from a three-week Discovery program to an Advisory and Jumpstart program lasting up to 12 weeks, featuring a technology assessment, proof-of-concept scoping, production planning, and more.
The three-week Discovery offering has three phases:
Opportunity Analysis (Week 1): Immersion in company priorities, inventory of internal data sources, risk identification and opportunity prioritization
Use Case Selection (Week 2): Identify priority use cases to optimize the business and pinpoint a candidate for initial pilot based on value creation opportunity, risk and data availability
Planning (Week 3): Plan a proof-of-concept vision, approach and timeline, aligned to value creation goals, hypotheses to validate and success criteria, then Hitachi Vantara recommends a high-level roadmap for next actions
Jeb Horton, SVP of Global Services at Hitachi Vantara, stated: “AI solutions require a combination of tools, technologies, platforms, and frameworks that facilitate model development, deployment, and management. By combining the industry expertise of Hitachi and Hitachi group company partners such as GlobalLogic with the Hitachi iQ solutions portfolio, the company offers a unique blend of infrastructure and services capabilities to provide the market with customized, industry-specific tools.”
Jason Hardy.
We asked Hitachi Vantara CTO for Artificial Intelligence & X-Portfolio Jason Hardy some questions about Hitachi iQ.
Blocks & Files: Will Hitachi Vantara get certified with Nvidia’s SuperPod?
Jason Hardy: We are celebrating the success of our BasePOD certification, while simultaneously planning for what’s next. There are obvious goals including the development of additional Hitachi iQ solutions, and we are also in the process of planning what is next for our Nvidia partnership. While we can’t speak for Nvidia in terms of their plans for SuperPOD, examples like SuperPOD, additional compute capabilities, and planning for the next generation of GPUs are obvious points of discussion.
Blocks & Files: Will Hitachi Vantara support GPU Direct?
Jason Hardy: Yes, Hitachi Content Software for File has supported GPUDirect capabilities since 2021. The inclusion of the Nvidia Ethernet and InfiniBand switching portfolio as a part of our Hitachi iQ release means that we are now able to provide our customers with an end-to-end solution. Because GPUDirect requires specific integration beyond just “plugging it in,” Hitachi iQ engineering teams are actively working on an integration guide to help our customers simplify deployment and integration into their environments and workflows, which we expect to make available as an added resource in the coming months.
Blocks & Files: Does Hitachi Vantara support LLM dataset storage, including vector embeddings?
Jason Hardy: Yes, Hitachi Vantara’s broad range of storage capabilities have created the perfect portfolio to support the requirements of a GenAI workload. Whether you are utilizing the Hitachi Content Platform or VSP One File as a repository for raw data targeted for RAG/Vector processing, or the Hitachi Content Software for File platform for high-speed vector database storage to support the high-performance query and embeddings workload that RAG pipelines generate, Hitachi Vantara’s storage capabilities are especially well suited to support various AI workloads.
Blocks & Files: Does Hitachi Vantara support RAG workflows?
Jason Hardy: Yes, Hitachi Vantara’s portfolio supports both the computational workload required for RAG and the storage requirements that both source data acquisition from either NAS, Object, or Block, and the high-speed requirements for the GPU workload required to perform the embeddings and queries.
Additionally, Hitachi’s capabilities to support our customers in implementing their GenAI pipelines, including the necessary RAG processing, allows for us to be a unique position for both the infrastructure requirements and the engineering requirements to build these solutions.
Blocks & Files: Does Hitachi Vantara do anything to aid structured information be use used in RAG workflows or AI generally?
Jason Hardy: Today, using Hitachi Content Intelligence and our Pentaho software suite, we can help our customers understand their data ecosystem, including the detection and transformation.
***
Hitachi iQ and the AI Discovery Service are now both available globally. Get more information from a Hitachi V webpage, Horton blog, and solution brief document.
Cloud storage provider Backblaze has appointed its first CRO, Jason Wakeam, whose past experience includes serving as the VP of global sales and OEM at SnapLogic. He’s also worked at Cloudera, Microsoft, and Hewlett-Packard. Wakeam will spearhead the company’s overall sales strategy, with a focus on expanding market share and driving new revenue opportunities. He succeeds co-founder Nilay Patel, who previously served as VP of sales, and has transitioned to oversee the company’s recently established New Markets team with a special focus on AI. Backblaze says it has a commitment to attracting, retaining, and growing with larger mid-market customers.
…
David Trossell.
WANrockIT and PORTrockIT WAN acceleration supplier Bridgeworks has signed its first global distributor – Nuvola Distribution, a technology and services distributor based in the UK and Republic of Ireland. Bridgeworks’ WAN Acceleration technologies use artificial intelligence, machine learning and data parallelisation to increase the speed and volume of data flows over wide area networks (WANs). They are data agnostic and mitigate latency and packet loss to permit encrypted data to be sent and received over thousands of miles, while significantly increasing bandwidth utilization by up to 98 percent, the company claims. CEO and CTO David Trossell says the initial goal is to enable half a dozen strategic reseller partners in the UK, and to expand into each of the major European markets in year one. “We would also like to establish the Bridgeworks name and portfolio across EMEA,” he says.
…
Ceramic nanoscale etching archival storage startup Cerabyte, which recently opened US offices, has joined the Active Archive Alliance. Rich Gadomski, co-chairperson of the Active Archive Alliance and head of tape evangelism at FUJIFILM North America Corp., Data Storage Solutions, said: “In this era of explosive data growth and the impact of AI/ML, the demand for active archive solutions is more critical than ever. Cerabyte’s cutting-edge, ceramic-based data storage technology aligns perfectly with our mission of providing sustainable, cost-effective strategies to address data growth challenges now and in the future.”
…
Cohesity announced v7.2 of its unified cloud data management platform, Cohesity Data Cloud, with enhancements around speed, security, scale, and simplicity. It says it has has better security thanks to OAuth 2.0 support, stronger control and recovery of Kubernetes workloads, and more flexibility and efficiency when deploying in the three major cloud platforms AWS, Azure, and Google. With this latest release, customers can achieve a lower TCO and better performance offering multiple configurations, up to 200 TByte per node, that simplify operations and reduce overall storage costs. Compared to prior versions of the cloud edition, the new version can help organisations realise over 100% ingest throughput, require up to 50 percent fewer nodes per cluster, and reduce costs by up to 66 percent per terabyte. Check out the release notes here.
…
Pranay Ahlawat.
Data protector Commvault has hired Pranay Ahlawat as its first CTO and AI Officer (CTAIO). He brings a background in AI, cloud and infrastructure software, business strategy, product development, and operations, including experience at Boston Consulting Group (BCG), advising software and SaaS companies on cloud and infrastructure, brings to Commvault. He’ll report to Chief Product Officer Rajiv Kottomtharayil.
…
DataStax will use NVIDIA’s NeMo Retriever as its default vector embedding service in its cloud service Astra DB. Langflow and Astra DB now integrate NVIDIA NeMo Retriever NIM embedding microservice; the Hyper-Converged Database for GenAI apps now includes NeMo Retriever NIM embedding microservice as default. More info here.
…
Analyst research firm DCIG has published a report naming its top 5 alternatives to VMware. They are HiveIO Hive Fabric, Microsoft Azure Stack HCI, Nutanix Cloud Infrastructure, Scale Computing Platform (SC//Platform) and VergeIO VergeOS.
…
SW-defined storage survivor Falconstor announced certification of its StorSafe and StorGuard products with IBM Storage Ceph. IBM Storage Ceph provides an efficient way to build a data lakehouse for IBM watsonx.data and next-generation AI workloads. StorSafe is the only IBM-certified, sold, and supported solution for backup optimization with IBM Power Virtual Server and now Ceph, available via the IBM Cloud Catalog. StorGuard offers storage virtualization, business continuity, high availability, system failover, and data migration.
…
Jennifer Temple.
HPE CMO Jim Jackson is retiring after 7 years in post. He will be succeeded by Jennifer Temple who is being promoted after 7 years of being Chief Communications Officer to Chief Marketing & Communications Officer (CMCO). She will oversee all aspects of HPE’s global brand, marketing, and communications strategy and said: “HPE is a brand executive’s dream because purpose comes together with bold innovation and passionate team members. To fully become the AI solution partner for any enterprise so we can keep solving the world’s biggest challenges, we are going to need to show up and connect in powerful new ways. This is going to be fun.”
Financial systems: 70% are covered by data protection strategies.
E-commerce and HR Management Systems: 50 percent are covered.
CRM and ERP systems: 48 percent and 42 percent respectively.
Critical transaction-based systems, custom applications, and collaboration and productivity tools: Are lagging behind with only between a third and a quarter of systems covered.
…
Kioxia, the inventor of NAND flash memory when it was part of Toshiba, is the recipient of the FMS: the Future of Memory and Storage Lifetime Achievement Award for 2024. The Kioxia engineering team, consisting of Hideaki Aochi, Ryota Katsumata, Masaru Kito, Masaru Kido, and Hiroyasu Tanaka, will accept this award for its pioneering work in developing and commercialising 3D flash memory.
…
MKW Ventures analyst Mark Webb will present a paper, “Memory Technologies : How Chiplets Change Everything”, on emerging memories at FMS 2024 [FMS OMEM-203-1 Heterogeneous Solutions for Performance Session – Wed Aug 7, 3:00pm – 4:05pm Ballroom C]. He reckons emerging memories – MRAM, RRAM, PCM, Crosspoint, FeRAM, UltraRAM – will not emerge to any material impact on the DRAM or NAND Market. The emerging memory market will not be $30 billion or even $3 billion in the next 5-10 years. It is well below $300 million today. In summary he says there is no universal memory coming. We have SRAM, DRAM, NAND, NOR, MRAM, RRAM, PCM available today. Let’s focus on integrating those as embedded or in chiplets.
MKW Ventures chart
…
MaxLinear, which supplies radio frequency (RF), analog, digital and mixed-signal integrated circuits and storage accelerators, saw its Q2 2024 revenues fall 50 percent y/y to $92 million with a loss of $39.3 million, compared to the year-ago $4.4 million loss. It said the broadband market recovery is slower an anticipated and it’s taking further actions to align with market conditions. It hopes to see revenues between $70 million and $90 million in Q3. This weak guidance prompted a share sell-off.
…
Momento, a real-time data platform startup founded by the engineering team that built DynamoDB, has raised $15 million in a Series A round led by Bain Capital Ventures. Momento delivers cloud primitives like cache, storage, and pub/sub so developers don’t have to write this stuff themselves. Daniela Miao, co-founder and CTO of Momento, said: “The AI revolution requires enterprises to move faster than ever, which leaves them with even less time to build, harden, and operate their infrastructure. Shipping better products faster than everyone else is now table stakes for survival. Your engineers need a rocketship that’s ready to fly, now.”
…
Percona tells Kubernetes v1.31 is due out in August and has some storage changes:
All cloud service provider-specific in-tree volume plugins are removed and now live within the Container Storage Interface world.
Persistent Volume last phase transition time feature moved to GA in v1.31. This feature adds a PersistentVolumeStatus field which holds a timestamp of when a PersistentVolume last transitioned to a different phase. This allows you to measure time e.g. between a PV Pending and Bound. This can be also useful for providing metrics and SLOs.
Always Honor PersistentVolume Reclaim Policy moved to Beta in v1.31. This feature prevents volumes from being leaked by honoring the PV Reclaim policy after the Bound PVC is also deleted.
Kubernetes VolumeAttributesClass ModifyVolume moved to Beta in v1.31. This feature extends the Kubernetes Persistent Volume API to allow users to dynamically modify volume options (such as IOPs and throughput), after a volume is provisioned.
Project quotas for ephemeral storage moved to Beta in v1.31. Using project quotas to monitor resource consumption for ephemeral volumes provides much better performance for huge volumes and lots of Pods utilizing those. Valid use case for ephemeral databases using local NVMe storage at scale.
…
AI Vector database startup Pinecone says Pinecone Connect is now GA. This offers a widget within your platform for developers to sign up or log in, choose or create an organization and project in Pinecone, and generate an API key instantly. It allows developers to manage Pinecone resources directly from another platform via a simple authentication flow. Partners including Twilio and Matillion are already using Pinecone Connect to streamline AI workflows for their users. Those who are new to vector databases can easily import their data into Pinecone and begin their Gen AI journey for free.
…
Cloud data warehouser Snowflake will host the Llama 3.1 collection of multilingual open source large language models (LLMs) in Snowflake Cortex AI for enterprises to arness and build powerful AI applications at scale. This offering includes Meta’s largest and most powerful open source LLM, Llama 3.1 405B, with Snowflake developing and open sourcing the inference system stack to enable real-time, high-throughput inference and further democratize powerful natural language processing and generation applications. Snowflake’s AI Research Team has optimized Llama 3.1 405B for both inference and fine-tuning, supporting a 128K context window from day one, while enabling real-time inference with up to 3x lower end-to-end latency and 1.4xhigher throughput than existing open source offerings. It allows for fine-tuning on the massive model using just a single GPU node — eliminating costs and complexity for developers and users — all within Cortex AI.
…
Google and Veritas Technologies have announced the availability of Veritas’ Alta Data Protection offering on the Google Cloud Marketplace. They claim Veritas Alta Data Protection reduces cloud computing costs by as much as 40 percent, and cloud snapshot storage costs by as much as 90 percent. It is self-managed and designed to efficiently secure and protect mission-critical data in the cloud, at scale. Features enable enterprises to:
Write data from backup to on-prem to Google
Seamlessly copy applications to and from Google
Run applications (including PaaS workload, Cloud SQL, MySQL, PostreSQL ad Google Cloud Storage) on the Google Cloud Engine
Leverage Veritas technologies to back up customer data to any Google Cloud Storage type
Purchase solutions through Cloud Marketplace
…
Western Digital has added an 8TB variant to its M.2 SN850X gaming SSD, which uses a PCIe gen 4 link. It previously maxed out at 4TB. More info can be found in an AnandTech report. The 8TB SSD costs $899.99 (MSRP) with a heat spreader.
Sk hynix revenues for Q2 soared as demand for high-bandwidth GPU memory went up in lockstep with AI application use.
The Korean NAND and DRAM fabrication outfit reported revenue of ₩16.4 trillion ($11.8 billion) in the quarter ending June 30, up 125 percent annually and 32 percent sequentially, with a profit of ₩4.1 trillion ($2.97 billion) contrasting with the ₩3 trillion ($2.17 billion) loss a year ago.
The company makes both DRAM and NAND, with its DRAM manufacturing lines making either HBM or lower-price ordinary DRAM. It said HBM sales went up by more than 80 percent compared to the previous quarter and more than 250 percent compared to the same period last year.
SK hynix earned its highest quarterly profit since 2018 in its Q2 2024
Enterprise SSD sales rose 50 percent sequentially while mobile NAND sales were less impressive. We don’t know the split between DRAM and NAND revenues, but understand DRAM revenues are significantly higher. Wedbush analyst Matt Bryson told subscribers that SK hynix’ “NAND bit growth decreased in the low single digit percentage range Q/Q while ASPs were up in the mid to high teen percentage range.”
Generally, Bryson said: “Hynix generally talked to strength in enterprise applications more than offsetting modest recovery in handsets and weaker than expected conditions in PCs.”
A statement from SK hynix CFO Kim Woohyun about the results said: “The company will further solidify the position as a leader in AI memory products by focusing on developing the best process technology and high-performance products based on a stable financial structure.”
SK hynix revenue recovery is accelerating
In the NAND area, it says it will “lead the market in the second half with 60 TB products, expecting eSSD sales to be more than quadrupled compared to last year.” Results for its Solidigm subsidiary were not revealed. In May, an SK hynix statement said: “We are preparing an eSSD product that combines SK hynix’s unique NAND technology with Solidigm’s eSSD solution capabilities.”
The company aims to capitalize on HBM demand by mass-producing 12-layer HBM3e products, which are being sampled by customers now. It will launch 32 Gb DDR5 DRAM for servers and MCRDIMM (Multiplexer Combined Ranks Dual In-line Memory Module) for high-performance computing in the next two quarters.
It is increasing its HBM production capacity with its new Cheongju M15X fab, due to start mass production in the second half of 2025. It will also start construction of the first fab of the Yongin semiconductor cluster in March 2025, and complete it in May 2027.
SK hynix is being chased by Samsung and Micron in the HBM area, with Nvidia certifying Samsung’s HBM3e chips for use in its H20 GPUs sold in China. Nvidia has not yet certified Samsung’s HBM3e chips more generally, which are currently supplied by SK hynix.
Micron has said it expects “to generate several hundred million dollars of revenue from HBM in FY2024, and multiple billions in revenue from HBM in FY2025.”
The next and faster HBM4 generation is coming and all three suppliers – Sk hynix, Samsung, and Micron – will want to participate in the anticipated HBM4 AI server gold rush in the 2026-2027 period. If this gold rush takes place, with SSD demand rising as well, then SK hynix revenues could grow to the ₩50 trillion level and beyond. We could be looking at a DRAM supercycle, and not just a normal and traditional recovery in the memory boom and bust cycle – if the surge in server AI workloads continues (a big “if”).
A financially stronger SK hynix could then play an influential role in consolidating the NAND supplier market, with dominant Samsung facing four weaker players – Kioxia, its joint-venture partner Western Digital, SK hynix with subsidiary Solidigm, and Micron. SK hynix, which part-owns Kioxia, has already blocked a Kioxia-Western Digital merger.
Commissioned: The importance of data has never been more salient in this golden age of AI services. Whether you’re running large language models for generative AI systems or predictive modeling simulations for more traditional AI, these systems require access to high-quality data.
Seventy-six percent of organizations are counting on GenAI to prove significant if not transformative for their businesses, according to Dell research.
Organizations teem with sales summaries, marketing materials, human resources files and obscene amounts of operational data, which course through the data center and all the way to the edge of the network.
Yet readily accessing this data to create value is easier said than done. Most organizations lack a coherent data management strategy, storing data in ways that aren’t easy to access, let alone manage. For most businesses, anywhere and everywhere is just where the data ended up.
Think about how many times employees have tried and failed to find files on their PCs. Now multiply that experiences thousands of times daily across an enterprise. Finding information can often feel like looking for a virtual needle in a data haystack.
You probably tried to centralize it and streamline it to feed analytics systems, but without structure or governance, the monster has grown unwieldy. And don’t look now – with the advent of GenAI and other evolving AI applications, your organization craves access to even more data.
Accessible data in the AI age
Maybe you’ve been tasked with activating AI for several business units, with partners in marketing and sales collateral to product development and supply chain operations looking to try out dozens or even hundreds of use cases.
Given the years of data neglect, affording these colleagues access to the freshest data is a great challenge. How do you move forward when these tools require data that must be cleaned, prepped and staged? As it stands, IT typically spends a lot of time on the heavy lifting the comes with requests for datasets, including managing data pipes, feeds, formats and protocols. The struggle of tackling block, file and other storage types is real.
What IT doesn’t tackle may get left for others to wrangle – the data analysts, engineers and scientists who need high-quality data to plug into AI models. Asking the folks who work with this data to take on even more work threatens to overwhelm and capsize the AI initiatives you may be putting in place.
But what if IT could abstract a lot of that effort, and make the data usable more rapidly to those who need it, whether they’re running LLMs or AI simulations in HPC clusters?
To the lakehouse
Organizations have turned to the usual suspects, including data warehouses and lakes, for this critical task. But with AI technologies consuming and generating a variety of structured and unstructured data sources, such systems may benefit from a different approach: A data lakehouse.
The data lakehouse approach shares some things in common with its data lake predecessor. Both accept diverse – structured and unstructured – data. Both use extract, transform and load (ETL) to ingest data and transform it.
However, too many organizations simply let raw data flow into their lakes without structure, such as cataloguing and tagging, which can lead to data quality issues – the dreaded data swamp.
Conversely, the data lakehouse abstracts the complexity of managing storage systems and surfaces the right data where, when and how it’s needed. As the data lakehouse stores data in an open format and structures it on-the-fly when queried, data engineers and analysts can use SQL queries and tools to derive business insights from structured and unstructured data.
Organizations have unlocked previously siloed data to make personalized recommendations to customers. Others have tapped lakehouses to optimize their supply chains, reducing inventory shortfalls.
Democratizing data insights
While a data lakehouse can help organizations achieve their business outcomes, it shouldn’t be mistaken for a lamp. You can’t plug it in, switch it on and walk away. That’s where a trusted partner comes in. Dell offers the Dell Data Lakehouse, which affords engineers self-service access to query their data and achieve outcomes they desire. The solution leverages compute, storage and software in a single platform that supports open file and table formats and integrates with the ecosystem of AI and ML tools.
Your data is your differentiator and the Dell Data Lakehouse respects that by baking in governance to help you maintain control of your data and adhere to data sovereignty requirements.
The Dell Data Lakehouse is part of the Dell AI Factory, a modular approach to running your data on premises and at the edge using AI-enabled infrastructure with support from an open ecosystem of partners. The Dell AI Factory also includes professional services and use cases to help organizations accelerate their AI journeys.
How is your organization making finding the needle in the haystack easier?
