Home Blog Page 148

NEO Semiconductor claims 3D X-DRAM has 8x density of rivals’ DRAM

Startup NEO Semiconductor suggests its 3D X-DRAM technology could produce a 128Gbit DRAM chip with 230 layers – eight times more than current DRAM density.

Update. Andy Hsu comments on floating body technology added, 5 May 2023.

NEO is focused on semiconductor memory and has developed X-NAND technology, NAND dies with multiple parallel planes to speed data access – 1,600MBps sequential write bandwidth in gen 1 and 3,200MBps in gen 2. It says there is an industry-wide effort under way to bring 3D cells to DRAM and NEO now has 3D DRAM technology. Like 3D NAND, it could drive up DRAM density significantly.

Andy Hsu, NEO Semiconductor
Andy Hsu

Andy Hsu, co-founder and CEO, bullishly said: “3D X-DRAM will be the absolute future growth driver for the semiconductor industry.” Hsu was with an unnamed semiconductor startup for 16 years after graduating with a Masters from the Rensselaer Polytechnic Institute in 1995. He founded NEO Semiconductor in August 2012 and has 12 patents to his name. NEO itself has 23 patents.

He added: “Today I can say with confidence that NEO is becoming a clear leader in the 3D DRAM market. Our invention, compared to the other solutions in the market today, is very simple and less expensive to manufacture and scale. The industry can expect to achieve 8X density and capacity improvements per decade with our 3D X-DRAM.”

NEO says 3D X-DRAM is necessary to address the increase in demand for high-performance and high-capacity memory semiconductors driven by the next wave of AI applications such as ChatGPT.

A graphic shows the 3D X-DRAM concept: 

NEO Semiconductor 3D X-DRAM

The concept has a 3D NAND-like DRAM cell array structure based on capacitor-less floating body cell (FBC) technology. This FBC technology stores data as electric charges using one transistor and zero capacitors. NEO says it can be manufactured using current 3D NAND-like processes and only needs one mask to define the bit line holes and form the cell structure inside the holes. This provides a high-speed, high-density, low-cost, and high-yield fabrication solution, which is a significant claim.

NEO Semiconductor FBC technology
NEO graphic

Jay Kramer, president of Network Storage Advisors, a marketing consulting service, said: “Evolving from 2D to 3D architectures has introduced compelling and extremely valuable benefits to NAND flash, so achieving a similar evolution for DRAM is highly desirable industry-wide. NEO Semiconductor’s innovative 3D X-DRAM allows the memory industry to leverage current technologies, nodes and processes for enhancing DRAM products with NAND-like 3D architectures.”

NEO suggests 3D X-DRAM could scale past a 1Tb chip in the 2030-2035 period:

NEO suggests 3D X-DRAM could scale past a 1 terabit chip in the 2030-2035 period

We understand NEO will be looking to license its technology to DRAM fabrication companies such as Micron, Samsung, and SK hynix. NAND fabbers like Kioxia and Western Digital could also be licensing possibilities.

Relevant 3D X-DRAM patent applications were published with the United States Patent Application on April 6, we’re told. US patent 20230106561 by Fu Chang Hsu, 30 Sep 2022, is entitled 3D Memory Cells and Array Architectures. Its summary states: “a novel 3D array structure using floating-body cells to implement DRAM is disclosed. The array structure is formed using a deep trench process similar to 3D NAND flash memory. Therefore, ultra-high-density DRAM can be realized.” 

Bootnote

Neo Semiconductor has disclosed no funding information and LInkedIn shows just two employees; co-founders Andy “Fu Chang” Hsu and VP Engineering Ray Tsay. Both worked at a semiconductor startup; no name revealed but possible Aplus Flash Technology, involved in NOR flash, EEPROM and NVSRAM for 16 years and 11 years respectively before founding Neo Semiconductor. Andy Hsu is said to hold more than 100 patents. LinkedIn lists a third Neo Semiconductor employee; Hong Cui, but lists him or her as an engineer at Nero Semiconductor.

Update

Andy Hsu commented on a Linkedin post saying: “The early 2D floating body cell from 2000 – 2010 had some issues, especially when the cell size shrunk, the floating body became smaller. Unlike DRAM that can make the capacitor taller to keep the same capacitor size, 2D floating body cell has no way to increase the capacitor size. Recently, many research has been done by using new mechanisms such as heterostructure or polysilicon grain boundaries for storage and showed excellent data retention. Plus, 3D floating body cells’ size (diameter >100 nm) is much larger than 2D floating body cells, that provides enough capacitance for charge storage. The announced cell structure is just the basic form. The real structure may contain other features. We will disclose more details in FMS 2023.” 

The bigger HDDs come, the harder they fall. Well, kind of

Cloud and backup storage supplier Backblaze publishes quarterly hard disk drive failure rate statistics and has just started doing so for its SSD boot drives as well. Not much has changed in the annualized failure rates (AFR) with the latest HDD failure stats. As lifetime drive hours increase, the HDD population grows: there are 236,893 now, and an average AFR of 1.54 percent.

The latest Backblaze HDD AFR report has the usual table of HDD failure numbers by drive model and failure rate charts by capacity. As drives gain more capacity over time, we could say this chart shows a recovery in drive reliability after a 10TB failure rate peak. The newest 16TB drives are now the most reliable in Backblaze’s HDD buying history.

Backblaze also has a failure rate chart by manufacturer:

It’s no surprise that Seagate has the lowest reliability rating; Backblaze reported this was the case in late 2021 and, as these are annualized numbers, it will take years for the failing Seagate drives to disappear from the stats.

Average failed disk age

This time around, the Backblaze team examined the average age of failed drives, a topic which B&F reported on here. Data recovery business Secure Data Recovery found the average failed hard drive was just two years and 10 months old. Was the same true for Backblaze’s failed HDD population?

That population totalled 17,155 drives, dating from April 2013 up to the end of March this year. Their average age when they failed was two years and six months; not much different from the Secure Data Recovery number. The calculation was rerun to exclude drives which were still in operation, and therefore likely to fail at some point.

Backblaze table

This revealed “35 drive models consisting of 3,379 drives that have a failed average age of two years and seven months.”

You can access Backblaze’s HDD test data page and run your own calculations on the data. If you use or publish the data please cite Backblaze as the source. Backblaze notes that “you are solely responsible for how you use the data” and you may not “sell this data to anyone; it is free.”

Gartner: Begun the cloud file services wars have

A report put out at the end of February titled “Modernize Your File Storage and Data Services for the Hybrid Cloud Future” by Gartner research VP Julia Palmer pulls a lot into focus about the cloud file services marketing war – which B&F has written about in the past.

Julia Palmer, Gartner
Julia Palmer

Palmer starts from more or less the same points: unstructured (file) data is growing and organizations are using the scalability of the public cloud for several things, bursting compute capacity, bursting storage, archival storage, remote worker access and so forth.  

There are three groups of suppliers, she says:

Gartner file services modernization options
  • Next-generation file platforms: on-premises filers adding hybrid cloud capability and new software-only file services suppliers
  • Hybrid cloud file data services providing access, movement, life cycle management and orchestration
  • Hybrid cloud file platforms providing public cloud-based distributed file services

A table lists representative suppliers in each category:

Gartner top vendors for hybrid cloud file services

Next-generation file platforms have characteristics such as “being designed to integrate and leverage public cloud IaaS.” Vendors like NetApp and Qumulo have become software-defined and ported their filer operating systems to the public cloud, providing a file system and services abstraction layer covering the on-premises and public cloud worlds. VAST Data is heading the same way. Weka is there already.

At the other end of this spectrum – hybrid cloud file data services are in the middle – are the hybrid cloud file platforms. We have covered CTERA, Nasuni and Panzura often, also LucidLink and to an extent Vcinity, but not really included Zadara in this category. Nor have we looked at the public cloud suppliers themselves or Huawei or Oracle as being in this category.

The mid-spectrum players provide file data services, such as migration – Atempo, Data Dobi, Data Dynamics, Peer Software and WANdisco, for example. Komprise also has a migration offering but is better known for its file lifecycle management capabilities, as is Nodeum. Hammerspace is pioneering its data orchestration technology.

Palmer really nails it with the a tripartite viewpoint, which is a neat way to group vendors and then to compare them in the three categories and map their capabilities to an organization’s needs and direction.

It leaves us asking if the mid-spectrum hybrid cloud file data services players face having their capabilities provided as features by the next-generation file platforms on the one hand and the hybrid cloud file platforms on the other.

The Gartner paper doesn’t look at object storage, which we think may be a limiting factor, especially as combined file and object offerings are becoming common. We can view S3 as AWS’s hybrid cloud object platform and also look at Pure Storage, VAST Data and others as offering next-generation file+object platforms, while Cloudian and Scality layer file access on their object storage products, which already have public cloud support.

Commvault sees fewer large deals in flat but guidance-exceeding quarter

Commvault’s earnings were hit in the last quarter by a degree of large deal buyer reluctance, as revenues flattened.

Update: Financial analysts and earnings call-related information added 3 May, 2023.

Sanjay Mirchandani.

Revenues in the fourth FY2023 quarter, ended March 3, were $203.5 million – 1 percent lower than a year ago, but exceeding Commvault’s guidance of $197 million. There was a loss of $43.5 million, dramatically down on the year-ago profit of $8.0 million. Full FY2023 revenues were $784.6 million, 2 percent higher than a year ago, with a loss of $35.8 million – pretty much all due to the loss-making fourth quarter.

Sanjay Mirchandani, Commvault president and CEO, accentuated the positive, saying: “Commvault closed out the year strong, highlighted by Metallic eclipsing the $100 million ARR mark, 15 percent year over year total ARR growth, and strong cash flow. We enter the new fiscal year with momentum and confidence that Commvault customers are future proofed for the road ahead.”

Two up and two down quarters in Commvault’s FY2023.

The loss is partly attributable to an operating expenditure increase. Total Q4 opex rose from $162.5 million a year ago to $206.3 million, and within that there was a non-cash headquarters impairment of $53.5 million. Commvault is selling its 278,000 square foot Tinton Falls HQ to Ashling Development LLC for $40 million, leasing back just 10 percent of it. Remote work is the new normal for everyone, it seems. 

Financial summary

  • Operating cash flow: $67.8 million compared to year-ago $87.1 million
  • Free cash flow: $66.8 million.
  • Diluted loss/share: $0.98 vs $0.17 diluted earnings/share a year ago
  • Total cash: $287.8 million vs $267.5 million a year ago

Software and products brought in $90.2 million, 10 percent lower year on year. There were 187 large deals – transactions with greater than $100,000 of software and products revenue – in the quarter, down from 206 in the prior quarter and 226 a year ago. These large deals represented 72 percent of Commvault’s software and products revenue, so hiccups there directly affect the bottom line.

Services revenue rose 7 percent annually to $113.2 million.

Total recurring revenue in the quarter was flat year on year at $173.9 million, while annual recurring revenue (ARR) was up strongly with a 15 percent jump to $668.4 million. Subscription ARR rose 38 percent year on year to $477 million. The subscription business is 71 percent of Commvault’s ARR and subscription revenue – up 9 percent year on year to $95 million – is 46 percent of Commvault’s total revenue.

Commvault said deferred revenue growth related to Metallic as-a-service offerings continues to be a driver of cash flow. About 70 percent of Metallic customers are new to Commvault and there is a 125 percent net dollar retention rate; customers are spending more.

Its board has increased its share repurchase program funding to $250 million. Commvault notes that free cash flow supports share repurchases and it wants to use around 75 percent of FY2024 free cash flow for repurchasing shares.

Update

Wells Fargo analyst Aaron Rakers notes that Commvault’s perpetual license revenue at $17.561 million was -25 percent y/y and -11 percent q/q, contrasting strongly with the subscription revenue rises. Commvault disclosed that it has ~7,300 subscription customers; about half of its customer base. This compares to 6,600 and 5,000 in the prior and year ago quarters respectively.

William Blair Analyst Jason Ader noted: “While subscription revenue boosted growth, the weaker macro environment continued to challenge large deals. Management noted that while macro did not deteriorate from the prior quarter, budget scrutiny and elongated deal cycles persisted,” primarily in the enterprise market.

Ader is cautious about Commvault’s prospects because of competitive pressures; “As noted in our recent VAR tracker note, the data protection market remains highly fragmented and competitive, with newer vendors like Cohesity, Rubrik, Druva, Veeam, HYCU, and OwnBackup all seeing good traction. While VARs continued to see Commvault as the most feature-rich and robust data protection platform on the market, we worry that the company remains at risk of donating market share to newer, innovative, more brand-forward vendors.” 

Commvault’s revenue guidance for the next quarter – Q1 fy2024 – is for revenues between $195 million and $199 million; $197 million at the mid-point. This compares with $198 million a year ago (Q1 fy2023). It provided a full fy2024 revenue guide of $805 million to $815 million compared to fy 2023’s $784.6 million, a 3.2 percent rise at the mid-point.

Rakers told subscribers: “We are Overweight CVLT shares, and believe the company’s positioning as a data management play for hybrid cloud, improving competitive positioning with HyperScale, Metallic, and Activate, as well as increasing subscription renewal cycle opportunity should be viewed as positives. We believe Commvault can return to a sustainable mid single-digit revenue growth story.”

Storage news ticker – May 2

Open source data synchronizer Airbyte has a new API, complementing its existing cloud UI, that makes it possible to automate data movement and manage large deployments. Other suppliers can embed Airbyte Cloud’s data movement capabilities directly into their product offerings. Integrating Airbyte Cloud with orchestration tools like Airflow, Dagster, or Prefect makes it easier to manage and monitor data pipelines and ensure data is flowing smoothly between sources and destinations. A Terraform SDK is coming soon to define configurations and automate deployments as part of CI/CD pipelines. This will enable Airbyte resources and configurations to be version-controlled, reviewed, and tested.

Analyst house KuppingerCole says Arcserve is one of six market champions in its 2023 Leadership Compass for Cloud Backup for Ransomware Protection report. In alphabetical order the champions are Arcserve, Cohesity, Commvault, OpenText (MicroFocus), Veeam and Veritas. Read a blog and download the Arcserve-focused report extract here.

Storage provider rankings

Backup supplier CrashPlan has a new Partner Program giving channel partners dedicated resources  to make onboarding quick and painless, enhance sales opportunity identification, and shorten time-to-revenue. With incumbency protection and CrashPlan’s long lifetime value, it will continue to reliably produce revenue for partners in the program for years to come.

Data lakehouse supplier Dremio has achieved Microsoft Top Tier Partner status, Microsoft’s highest level of partner designation for Azure IP co-sell partners, driving strategic collaboration, solution differentiation and dynamic go-to-market momentum.

Data orchestrator Hammerspace said in a quarterly briefing that its business progress is accelerating. It showed a slide listing 50 users:

Customers of storage company Hammerspace

It says its channel partners are bringing in more opportunities and its pipeline is growing. 

Infinidat has produced a datasheet describing its InfuzeOS’s capabilities. InfuzeOS runs on the InfiniBox, InfiniBox SSA and InfiniGuard systems. Download it here.

Infinidat is supporting Tech Tackles Cancer through the “Battle of the Tech Rockstars” charity competition in London on May 24 at the O’Meara Nightclub and in Boston on September 21 at The Sinclair (Cambridge). These fundraising events are part of a broad effort within the tech community to advance cancer research and help pediatric cancer patients and their families. Infinidat is actively involved as a Gold Sponsor of both Tech Tackles Cancer London and Tech Tackles Cancer Boston. Infinidat also has a seat on the volunteer-led team that oversees and plans the organization’s activities.

Semiconductor analyst Mark Webb of MKW Ventures asks “When Will SSD Cost be Lower than HDD Cost?” He says he updated all the numbers he had previously shown at FMS 2022 on April 26 and the answer is still 2034, possibly, as he has previously calculated! He reckons that HDD technology has a few tricks left to pull and NAND scaling will slow over time [see Pure note below].

Peer Software is partnering with Wasabi Technologies to provide a combination of built-in  failover/failback and disaster recovery protection that reduces or eliminates snapshot or backup window limitations to keep critical data current and easily accessible on Wasabi cloud storage. Peer provides PeerGFS, a distributed file system that integrates existing storage systems across multi-site, on-premises and cloud storage. It has local access to data, high availability and redundancy across synchronized storage systems, and continuous data protection, and now supports Wasabi.

Proact has become a Rubrik Authorized Support Partner.

Shawn Rosemarin, VP, Product Management at Pure Storage, believes there will be no hard disks sold five years from now. He says NAND cost dynamics, density, energy efficiency, resiliency and innovative storage software are the game changers here and Pure Storage is leading this charge. Flash density will be a major factor in accelerating the move from HDD to Flash. We’re arranging a briefing to find out why he thinks this is true.

Analyst house TrendForce says SK hynix, Samsung and Micron had 50, 40 and 10 percent HBM market share respectively on 2022. As more customers adopt HBM3, SK hynix, the only mass producer of HBM3 chips, will get a 53 percent HBM market share in 2023 with Samsung and Micron having 38 and 9 percent, respectively. These two are expected to start HBM3 mass production towards the end of this year or in early 2024.

UK law firm Charles Russell Speechlys has selected a StorMagic SvSAN, running on two HPE DL380 servers in a hyperconverged configuration, for nine of its global offices. It says it found SvSAN was easy to deploy, showed better disk latency and offered the easiest storage usage visualization compared to Starwind and Linux Ceph. Speechlys previously used VMware vSAN to manage its business-level applications, including domain controller, Microsoft Exchange databases, and SQL.

Nvidia AI supercomputer shows its Lustre in Oracle cloud

Nvidia is running its AI supercomputer on Oracle’s cloud infrastructure with its Lustre file system relying on NVMe block access SSDs.

An Nvidia blog details how its DGX Cloud uses Oracle Cloud Infrastructure (OCI) to provide compute, networking, and storage infrastructure to OCI users. DGX Cloud is a multi-node AI-training-as-a-service for enterprises through cloud service providers like OCI.

The team says: “DGX Cloud eliminates the need to procure and install a supercomputer for enterprises that are already operating in the cloud – just open a browser to get started.”

With Nvidia DGX Cloud on OCI, “Nvidia pairs Oracle’s bare-metal infrastructure with the Nvidia NVMesh software. This enables file storage that is scalable on demand for use on DGX Cloud.” Nvidia acquired NVMesh technology by buying Excelero in 2022. The software takes block data from SSDs and presents it to remote systems as a pool of block storage, like a SAN (we’ll get to Lustre in a moment).

OCI bare metal E4 DenseIO compute instances, also known as shapes, are the building blocks for this high-performance storage. They consist of:

  • 128 AMD EPYC Milan processor cores
  • 2 TB of system memory
  • 54.4 TB of NVMe storage across a total of 8 NVMe devices (SSDs)
  • 2 x 50 Gbps NICs for Ethernet networking

The two 50Gbps physical NICs on the E4 DenseIO shapes enable high availability. The bare metal instance means no resources are lost to virtualization.

NVMesh takes the raw E4 shape NVMe storage and uses it to build a high-performance data volume. The shapes are combined into pairs with the NVMesh software providing high-availability across the pair. In-built data protection in other words. Encryption is also included.

These shape pairs are then used as the storage underpinnings for a Lustre file system, for both data and metadata storage. 

OCI running Nvidia DGX Cloud

Lustre capacity scales out on-demand dynamically by adding more shape pairs, which also means more metadata capacity is added as well. This ensures metadata capacity limitations don’t cause a processing bottleneck.

The users see Lustre as an Nvidia Base Command Platform (BCP) data set and workspace storage facility. BCP provides a management and control interface to DGX Cloud, acting as its operating system, and providing AI training software as a service. It works with both the DGX Cloud and with an on-premises or co-lo with a deployed DGX SuperPOD.  You can access a datasheet to find out more.

Nvidia says testing showed that its DGX Cloud on OCI had storage performance matching on-premises Nvidia Base Command Platform environments. DGX Cloud is also available on Azure with AWS and GCP availability coming.

Kinetica: AI is a ‘killer app’ for data analytics

Kinetica has integrated its streaming real-time analytics data warehouse with ChatGPT so customers can ask conversational queries of their proprietary data and receive answers in seconds. Amit Vij, Kinetica co-founder and president, claimed: “Generative AI revolution is a killer app for data analytics.”

ChatGPT is an application built on GPT-3.5 and GPT-4, which are Large Language Models, created by OpenAI which can respond to complex natural language queries and search public data sets, with comprehensive replies which are often right but may also be wrong. Kinetica’s analytics database, which supports both GPUs and CPUs, contains both time-series and spatial data, and they can be analyzed in real time. The database can now use ChatGPT as a front-end interface that converts natural language to Structured Query Language (SQL) and then runs the SQL queries.

Nima Negahban, Kinetica co-founder and CEO, said: “While ChatGPT integration with analytic databases will become table stakes for vendors in 2023, the real value will come from rapid insights to complex ad-hoc questions. Enterprise users will soon expect the same lightning-fast response times to random text-based questions of their data as they currently do for questions against data in the public domain with ChatGPT.”

With the current dash for ChatGPT integration by other analytics data warehouse/lake companies such as Databricks, Pinecone, SingleStore and Zilliz, we’d bet Negahban’s table stakes point is correct.

Kinetica says existing conventional analytic databases require extensive, before-the-fact data engineering, indexing and tuning to enable fast queries, which means the question must be known in advance. If the questions are not known in advance, a query may take hours to run or not complete at all. Using ChatGPT as a conversational front end does away with the need for pre-engineering data and also with the need to write complex SQL queries or navigate through complex user interfaces.  

Users can ask questions using natural language. ChatGPT can, Kinetica claims, understand the user’s intent and the logical data model and generate queries based on their questions. The user can then ask follow-up questions or provide additional context. Users get immediate answers to their questions without waiting for long-running queries or data pipelines to be built.

Vij said: “Kinetica plus ChatGPT makes complex, ad-hoc queries truly interactive, avoiding the ‘interactus interruptus’ of other integrations between large language models and databases.”

ChatGPT-type chatbots can provide not just wrong but imaginary answers. The datalakers integrating chatbot front ends which generate SQL queries against proprietary data sets will hope this bounding will stop the chatbot dreaming up false information.

Kinetica background

Kinetica was founded in 2016 and has raised $77.4 million across four funding events, the last a $14.4 million venture round in 2020.

It says its technology uses native vectorization to outperform other cloud analytic databases. In a vectorized query engine, data is stored in fixed-size blocks called vectors, and query operations are performed on these vectors in parallel, rather than on individual data elements. This querying of multiple data elements simultaneously results in faster query execution and improved performance.

The software supports a large number of spatial and temporal join types, graph operations, hundreds of analytic SQL functions, and enables visualization of billions of data points on a map. Customers include Liberty Mutual, TD Bank, the NBA, Lockheed Martin, USPS, T-Mobile, FAA, Ford, Point72, Verizon and Citi. Its streaming data warehouse can ingest, analyze, and visualize data sets with trillions of rows to model possible outcomes and assess risk for the US Air Force by tracking airborne objects for example.

In February Kinetica announced 90 percent annual recurring revenue growth over the past 12 months, a net dollar retention rate of 163 percent, and a doubling of its customer base. 

ChatGPT hype snags database startup $100M

Vector embeddings and ChatGPT are the key to database startup Pinecone unlocking a $100 million funding round.

Pinecone says it provides long-term memory for AI, meaning a vector database that stores numeric descriptors – vector embeddings – of the parameters describing an item such as an object, an activity, an image, video, audio file etc. The database can be searched for items that match a starting vector to detect likely recommendations to suggest to a purchaser, potentially fraudulent transactions, facial detection and so forth. Such databases are needed for large language models such as ChatGPT in the machine learning arena.

Edo Liberty, Pinecone
Edo Liberty

Edo Liberty, founder and CEO of Pinecone, said: “The new wave of AI-powered applications create, store, and search through a large number of vector embeddings, which are AI representations of data. We created Pinecone and the vector database category as a whole to let all AI developers easily work with a scalable and cost efficient database for this workload.”

Pinecone was set up in 2019 and its funding progress has accelerated, with a $10 million seed round in 2021, an A-round worth $28 million in 2022, and now a $100 million B-round. This was led by led by Andreessen Horowitz, with participation from ICONIQ Growth and previous investors Menlo Ventures and Wing Venture Capital.

The funding is in response to Pinecone’s progress in 2022; it claimed to have experienced an explosion in paying customers across all industries and customer sizes, mentioning Gong with its revenue intelligence offering, and Zapier, with is workflow automation tools.

There has been a sudden surge in ChatGPT-related activity in the storage area, with widespread adoption of the ML tool needing data storage facilities and being applied as a tool to query stored information.

  • SingleStore sees it as a database querying possibility.
  • Databricks sees it as a new way to query analytics datasets.
  • Panmnesia wants to run recommendation models faster.
  • SK hynix wants to supply HBM chips to servers running chatbots.
  • Zilliz is developing a cloud vector database.
  • Nuclia is developing language search models.
  • Cohesity will provide data structures to OpenAI.

Peter Levine, General Partner at Andreessen Horowitz, said: “The rise of AI is driving a fundamental shift in the way companies approach data management. In a very short amount of time, Pinecone has become a standard and critical component of the modern AI stack.”

B&F expects all storage data management, datalake, data warehouse and AIOps suppliers to be exploring chatbot technology.

Bob Wiederhold, Pinecone president, encapsulated the fervor around the ChatGPT area: “With this funding, we will capitalize on our hypergrowth amidst the AI revolution transforming every line of business and even creating new ones, from AI search to AI chatbots to AI agents and beyond.”

Gartner storage hype cycle

Chatbot technology is experiencing hypergrowth, and has found a place on Gartner’s hype curve where it is rocketing up to the peak of hyper-inflated expectations. Then it will tumble down to the trough of disillusionment, assuming the status of forgotten technologies like phrenology and holographic storage. But the hope is that it will emerge and climb the slope of enlightenment. The race is on to provide the platform technologies that will be needed when the chatbot phenomenon passes that phase and becomes widely deployed in the plateau of productivity.

UniFabriX ‘doubles’ rack server density with CXL

UniFabriX says its CXL memory expansion technology can double the number of servers in a CSP’s datacenter racks.

The Israeli startup has developed a Smart Memory Node with 32TB of DDR5 DRAM in a 2RU chassis. The pitch is that attached servers need less local DRAM, sharing the Smart Memory Node’s memory. That way they don’t have local stranded DRAM that’s not being used efficiently.

Ronen Hyatt, UnifabriX
Ronen Hyatt

CEO and co-founder Ronen Hyatt told an IT Press Tour: “For Cloud Service Providers we can 2x servers on a rack, up from 12 to 24. We take some memory from existing 12 servers and put it in Smart Memory Node. CSPs can then improve their server/DRAM granularity. We do it when the CSP builds new server racks. You don’t physically move DRAM DIMMS from the existing servers; you populate the rack with servers with smaller DRAM capacity than before.”

The reason is that by using the Smart Memory Node each server needs half as much electrical power as before, and the benefit is that the CSPs get twice as much revenue from a rack; more servers means more compute instances available to the CSP’s customers.

Unifabrix Smart Memory Node's SK hynix DDR5 DIMM
UnifabriX Smart Memory Node’s SK hynix DDR5 DIMM

The UniFabriX node uses standard DDR5 DIMMs on a main printed circuit board with FPGAs running the controlling software which provisions and deprovisions memory. Hyatt said the DIMMs are denser than a CXL memory expander card such as one from Samsung. Thermal considerations were also important in the DIMM choice, he added..

UniFabriX had to develop special cables to link the Smart Memory Node to the servers in a rack. The node has 10 x PCIe gen 5 ports in its rear face and each port can be bifurcated. It developed a 16-lane PCIe gen 5 cable with each lane carrying 32Gbps full duplex.

Danny Volkind, UniFabriX
CTO and co-founder Danny Volkind, the cable guy, holding up the PCIe gen 5 cable

UniFabriX’s software provides CXL 1.1. (memory expansion) and 2.0 (switching and memory pooling) functions and most of CXL 3.0 but not the multi-level switching and some security features. Hyatt said: “We can abstract 90 percent of CXL 3.0 features and make them usable in CXL 1.1 and 2.0 environments.”

The node provides memory checkpointing, useful during long application runs as a way of recovering from any failures. Hyatt said that UniFabriX provides software primitives and a mediation layer for hyperscalers to compose Smart Memory with servers, providing memory as a service.

The company has paying customers – the HPC market was mentioned – and is looking for OEM deals. There has also been interest from a composability company, but UniFabriX did not want to follow up that opportunity, preferring to focus on the CSP and HSP server memory expansion market.

It believes the CXL memory market could have a $20 billion total addressable market (TAM) by 2030, with the specific memory pooling TAM being in the $14 billion to $17 billion range. If UniFabriX can capture a decent percentage of that, its backers will be very happy with CXL memory expansion and pooling driving their own revenues.

Single sales rep inflated WANdisco numbers

WANdisco has confirmed that one senior sales rep entered $14.9 million in fictitious orders in 2022, dwarfing the $9.7 million brought in by the rest of the sales team.

The company sells replication-based software to move data from edge and central databases to the public cloud under commit-to-consume contracts. Recorded sales bookings for 2022 of $127 million had grown so fast that co-founder, chairman and CEO Dave Richards thought his business was unstoppable. He told us days before the fiasco was uncovered in March that in the live data movement market WANdisco has no competition: “We have 100 percent market share.”

Days later the company’s share trading on AIM was suspended when its 2022 results were found to be unauditable due to fake sales, and it hired an accountant to go over its books.

Richards and his CFO, both duped, subsequently resigned, a new chairman, Ken Lever, and new CFO, Ijoma Maluza, were put in place. The search for a new permanent CEO is under way.

Forensic accounting firm FRP Advisory confirmed today that the fake sales were all due to a single senior sales employee. Lever said: “We are pleased to receive these findings, which confirm the limits of the impact of the identified irregularities in line with our announcement of April 3.”

That statement early this month said sales bookings of $127 million, initally reported for 2022, were false and should have been $11.4 million. Revenue was actually $9.7 million and not $24 million. Now that FRP’s conclusions are being sent to the auditors, 2022 accounts will be forthcoming.

A WANdisco statement is admirably clear: “All of the purchase orders associated with the senior sales employee in question are illegitimate… all other purchase orders… are legitimate.”

In summary: “It remains the case that the evidence identified supports the initial view that the irregularities are as a result of the actions of one senior sales employee only.” 

The board wants share trading for WANdisco stock to resume and is working to that end, with Lever saying: “The Board remains squarely focused on workstreams to lift the suspension of our shares as soon as is practicable and position WANdisco for long-term growth and success.”

The FCA investigation into the company is ongoing.

Komprise service auto tiers files to Azure blobs

Komprise has an Intelligent Tiering for Azure service in the Azure marketplace to scan on-premises NAS stores and transfer cold files to Azure blob stores, saving customers cash.

The Komprise Intelligent Data Management (KIDM) offering enables customers to scan data across all on-prem and public cloud NAS and Object storage silos and analyze, migrate, tier, archive, replicate and generally manage it through policies. Komprise launched an offering, based on its Elastic Data Migration software in KIDM, with Azure to migrate on-prem and non-Azure cloud files and S3 object stores to the Azure cloud in February last year. Targets included Azure Blob Storage, Azure Files, or Azure NetApp Files. Now it has a simpler offering – just moving files from on-prem NAS and non-Azure and Azure file stores to Azure Blob buckets.

Krishna Subramanian.

Krishna Subramanian, president and COO of Komprise, issued a statement: “More than 100 enterprises are already using Komprise through the Microsoft Azure Storage Migration Program because of its simplicity and convenience. The new Komprise Intelligent Tiering for Azure extends this ease of use by allowing customers to use services such as Microsoft Purview, Microsoft Defender for Storage, Azure Synapse Analytics and Azure AI with data copied or tiered from on-premises.”

That is to say, pump the files up to Azure Blob storage and then they can be accessed  by services such as Microsoft Purview and Defender for Storage, Azure Synapse Analytics and Azure AI.

KIDM connects to any file storage including NetApp, Dell PowerScale, Dell Isilon, Dell Unity/VNX, Windows Server, Nutanix, Pure, Azure Files, Azure NetApp Files, Qumulo and any NFS, SMB or dual-mode, mixed-mode, It identifies cold data, and based on the policies you set, it transparently tiers cold files to Azure Blob. It provides analytics on how much data you have, how fast it is growing, who is using it and how much it costs. 

Komprise Transparent Move Technology.

Users can then set policies on what data should be tiered and when and to which tier of Azure Blob. It then tiers data with Komprise’s Transparent Move Technology to the appropriate Azure tiers. 

A video describes KIDM:

Komprise Intelligent Tiering for Azure video.

The pricing is $0.008/GB/month based on annual Azure subscription of 250TB+. That’s an annual payment of $25,000. If the Azure subscription is for 100TB then the annual payment is $15,000, 500TB costs $50,000 and 1PB is priced at $100,000.

According to Subramanian: “Azure customers can use their existing Azure contracts and utilize their Azure Consumption Commitments through this specially priced Komprise offer in the Azure Marketplace.”

Jurgen Willis, VP Azure Specialized Workloads and Storage, said: “Every organization in the current environment is looking to do more with less while reducing cost. … Since Komprise tiers data to Microsoft Azure in native readable format and provides data workflows, customers can cut costs and leverage the full power of Azure services to address AI, big data, security and compliance use cases.”

CECT tops SPC-1 for price with Optane SSDs

SPC-1, the performance benchmark dominated by Chinese suppliers, has claimed a new price/performance champion – CECT using 240 Optane SSDs.

The Storage Performance Council’s SPC-1 benchmark measures storage array business-critical workload IO request performance, recording SPC-1 input/output operations per second (IOPS), price per IOPS, response time, capacity and price per gigabyte. In recent years it was abandoned by Western manufacturers and became dominated by Chinese storage vendors such as Huawei and Inspur.

CECT (China Electronics Cloud Technology Company) has submitted its first SPC-1 benchmark, recording 10,000,690 SPC IOPS with a system costing just $1,658,233. The result is seemingly astonishing as systems above and below its score are much more expensive.

CECT SPC-1 benchmark result
B&F table of top ten SPC-1 results with CECT result highlighted. The right hand column shows total system costs

Inspur’s AS5600G recorded 7,520,358 IOPS with a $2.9 million all-flash system in 2020. The Fujitsu ETERNUS DX8900 S4 hybrid SSD/HDD array came in with a 10,001.522 IOPS score and a system costing $6,442,522.88 back in 2019.

Even more surprising is that the CECT configuration included 240 x 256GB Optane PMEM 200 SSDs and 180 x 1.6TB Intel DC P4610 NVMe SSDs in 30 CeaStor 18116E storage nodes; 8 x Optane/node and 6 x NVMe/node for a total of 349,440GB capacity. The Optane drives were used to store metadata with the NVMe SSDs used for data.

CECT in SPC-1 results
High-end SPC-1 results in in a 2D space defined by a $/KIOPS vertical axis and SPC-1 IOPS horizontal axis. The lower the dot and the further to the right the better. CECT has established a new price/performance level

This is a simple and clear result showing the price/performance value of Optane SSDs in a mission-critical storage IO benchmark. It is perhaps a pity that it should emerge now after Intel has withdrawn the product and is just selling drives from its inventory. B&F wonders whether the availability of such a clear-cut and dramatic result five years ago would have altered the general perception of the usefulness of Optane SSD.

The CECT SPC-1 executive summary report and full disclosure report contain all the details.