Micron had another poor quarter, with revenues 57 percent down annually, but it met its guidance, and still thinks the memory downcycle has bottomed out, although China’s security investigation could slow its recovery.
In the fourth quarter of the memory market recession, revenues were $3.75 billion in the third fiscal quarter ended June1, with a loss of 1.9 billion vs a profit if $2.6 billion a year ago. DRAM revenues were $2.7 billion, down 57.4 percent while NAND brought in $1 billion, down 56 percent.
CEO and President Sanjay Mehrotra said: “Micron delivered fiscal third quarter revenue, gross margin, and EPS all above the midpoint of the guidance range. We believe that the memory industry has passed its trough in revenue, and we expect margins to improve as industry supply-demand balance is gradually restored.”
The revenue curves since Q3 fy2022 do start to look like the left-hand side of a U-shaped recession
He admitted that: “The recent Cyberspace Administration of China (CAC) decision is a significant headwind that is impacting our outlook and slowing our recovery,” referring to China’s decision Micron’s products represent a security risk, thought to be a tit-for-tat response to US blacklisting of Huawei. Up to 12.5 percent of Micron’s world-wide revenues could go away if China bans its products from being used by Chinese companies.
Financial summary
Gross margin: 17.8 percent of revenue vs 46.7 percent a year ago
The compute and network business unit (BU) was most affected with revenues down 64 percent to $1.4 billion, while the embedded BU, with its growing automotive component, least affected with a 36 percent revenue fall to $912 million. Mobile BU revenues dropped 58 percent to 819 million and the storage BU brought in $627 million, a 53 percent fall.
Micron’s strategy is to cut its costs, lower production to help supply get back in balance with demand, keep up its technology status vs its competitors, and wait it out until recovery starts in calendar 2024. Its storage and embedded BUs are experiencing quarter-on-quarter slight revenue rises which could be an early indicator of a positive change, if they’re sustained next quarter.
Bright spots are the automotive sector where cars, SUVs, pickups and delivery vans are all getting smarter on the driver assistance and infotainment front, meaning more memory is needed.
But the biggest boost could come from AI where generative AI’s sudden boom could mean datacenter servers need much more DRAM and NAND, with SSD sales into storage growing as well. Mehrotra said: “AI servers have six to eight times the DRAM content of a regular server and three times the NAND content. In fact, some customers are deploying AI compute capability with substantially higher memory content. … We are in a strong position to serve AI demand for fast storage as these data-intensive applications proliferate.”
He added this in the earnings call [paywall]: “We consider 2024 to be a big banner year for AI, for memory and storage. And Micron will be well positioned to capture this with our strong portfolio of products.”
Micron will be hoping this AI boost will be long-lived and not short-term. [We asked the Perplexity ChatGPT-like chatbot abut this and it said: “It seems that the generative AI boom is likely to continue for the foreseeable future.”]
The China situation is a pain-in-the-proverbial for Micron. If it loses its China-based business it aims to keep its current market share. How? Mehrotra said: “Keep in mind that our share in DRAM is approximately 23 percent and our share in NAND is approximately 12 percent. So obviously, we have opportunities to gain share with other customers. And this is what we are focused on. It will take some time, and the CAC decision … is hurting our business. It is slowing our recovery. It can result in quarter-to-quarter variations as well. But over longer term, our target is to maintain our share.”
The outlook is for Q4 revenues to be $3.9 billion +/- $200 million, a 41.3 percent Y/Y drop at the midpoint. That’s less of a fall than in the previous two quarters. The full fy2023 revenues would then be $15.43 billion, a near enough 50 percent drop plunge on fy2022 and its lowest result since fy2016. Downcycles in the memory industry can be harsh.
The two leading cloud data warehouse and lakehouse suppliers, Snowflake and Databricks, are racing to get the most comprehensive generative AI functions working on their customers’ data sets.
Both offer data platform solutions across the three main cloud providers: AWS, Microsoft Azure, and GCP. Both are highly funded, Databricks with $3.6 billion raised and post-IPO Snowflake with a $2.5 billion annual revenue run rate. Snowflake has made a series of AI-related announcements at its Snowflake summit the day after Databricks said it was buying large language model (LLM) startup MosaicML with technology used in generative AI.
LLMs are used to make chatbots that can “understand” natural language queries and requests, analyze source data sets and respond with query answers, software code, generated images – like a superintelligent hotel concierge offering a white glove service to guests.
Snowflake was founded in 2012 by three data warehouse experts who saw limitations in existing on-premises data warehouse products and set out build a proprietary cloud-native, cloud-resident, scalable and fully managed data warehouse. The technology uses a shared-nothing massively parallel processing (MPP) SQL-based query cluster that accesses shared-disk data storage.
It has added external table support to broaden its data sources outside its own data warehouse wall.
Databricks was started up a year later by the original creators of the Apache Spark in-memory big data processing platform. It wanted to simplify large-scale data processing and help enterprises get insights from troves of structured, semi-structured and unstructured data. It has espoused the lakehouse idea, combining traditional data warehousing and data lake workloads in a single unified platform which is offered via a platform-as-a-service (PaaS) model with customers able to customize the compute side.
Databricks’ software works on raw data in the lakehouse, without it having to undergo an extract, transform and load (ETL) process, as is the case with Snowflake. Of course, Databricks’ processes have to select and subset the data first.
Snowflake has its own data storage format while Databricks has its open source Delta Lake storage facility.
As well as SQL queries, Databricks supports real-time stream processing, machine learning, and graph processing. Databricks has MLlib and TensorFlow; built-in libraries for machine learning and deep learning. It has also developed a facility for building and deploying LLMs and has its open source Dolly chatbot. It is now buying MosaicML for $1.3 billion to help customers build and deploy AI models on their own data.
Snowflake is running parallel to this and has just announced:
A partnership with Nvidia so that customers can use Nvidia’s NeMo framework for developers to build, customize, and deploy generative AI models using their organization’s data, with billions of parameters.
The private preview Document AI LLM using Applica’s generative AI technology to let users analyze documents.
Extended Iceberg tables to push out the data set boundaries for Snowflake.
A private preview of Snowpark Container Services in which developers can run AI and ML models directly within Snowflake’s Data cloud. Customers get access to third-party software including LLMs, ML APIs for model development, Notebooks, and MLOps tools.
SnowPark is Snowflake’s facility for enabling non-SQL access to its data.
The net effect of this generative AI race between Databricks and Snowflake is that every other data lake, lakehouse and data warehouse supplier may be under pressure to follow in their footsteps or risk being left behind in what has quickly become a table stakes game.
Lakehouse supplier Dremio has already added Text-to-SQL, Autonomous Semantic Layer, and Vector Lakehouse functionality to its product. Open source vector database supplier Zilliz may start to field partnership and potential acquisition queries from various analytics suppliers. That’s because its technology stores the vector embeddings data needed by LLMs. Transactional and real-time database supplier SingleStore has demoed ChatGPT working against data it stores.
We are witnessing a generative AI frenzy as vendors try to ensure they don’t get left behind – and both Databricks and Snowflake just ratcheted the knob to a higher setting.
Aerospike, focused on real-time data, is now a member of the HPE Complete program which allows customers to acquire Aerospike software directly from HPE and its authorized partners. The Aerospike Real-time Data Platform, has been tested and validated with the HPE Alletra 4110 data storage server.
…
Since the beta release of the Airbyte connector builder two months before the official launch, more than 100 connectors have been built and deployed to production by users to support critical data movement workloads for long-tail connectors. Those are in addition to the 350 available data connectors. Airbyte started up in January 2020 with the aim of building open-source connectors from data sources to data lakes and warehouses.
…
Data intelligence vendor Alation announced the Open Data Quality Framework to bring best-of-breed data quality and data observability capabilities to the Snowflake Data Cloud. This allows users to capture data health, including DQ metrics and warnings for consumers, within their workflow and at the points of consumption across numerous data sources. Launch partners include Acceldata, Anomalo, Bigeye, Datactics, Experian, FirstEigen, Lightup, and Soda. With the Open Data Quality Framework, Alation customers can strengthen data governance for Snowflake by making data quality information visible, it says.
…
Spreadsheets remain the most popular data analytics tool in use, with 78 million users worldwide using the tool every day, but they are also some of the most error-prone tools, as 90 percent of spreadsheets contain errors due to incorrect or out-of-date data. Alation says its Connected Sheets for Snowflake allows users to instantly find, understand, and trust the data they need without leaving their spreadsheet. They can access Snowflake source data directly from the spreadsheets in which they work without the need to understand SQL or rely on central data teams.
…
There is a public preview of Azure NetApp Files double encryption at rest. It provides multiple independent encryption layers, protecting against attacks to any single encryption layer. This feature is currently available in West Europe, East US 2, East Asia regions and will roll out to other regions as the preview progresses.
…
Italian startup Cubbit, a geo-distributed cloud storage enabler, has an advisory board with the first members being Alec Ross, Mikko Suonenlahti, and Nicolas Ott. It launched the Next Generation Cloud Pioneers programme dedicated to Italian companies 18 months ago. More than 130 companies have joined the network, including Aeroporto Marconi di Bologna, Amadori, Bonfiglioli, CNP Vita (Unicredit Group), Granarolo, Marcheno Municipality, and SCM Group. Cubbit is preparing to expand its offer across Europe.
Cubbit DS3 is a multi-tenant S3-compatible object store optimised for edge and multi-cloud use cases. Files are encrypted, fragmented, redundantly stored, and distributed across geo-distributed networks called Swarms. The Swarm is built on recycled and dedicated computing and storage resources, helping to optimise the storage infrastructure of companies and enterprises at a fraction of the cost of traditional cloud providers while guaranteeing complete data sovereignty and avoiding lock-in. At least that is the pitch. Cubbit is present in 70+ countries and its technology has been deployed with over 5,000 active nodes.
…
Lakehouse shipper Databricks has added a Universal Format to the open-source Delta lake which allows data to be read from Delta as if it were Apache Iceberg or Apache Hudi. UniForm offers automatic support for Iceberg and Hudi within Delta Lake 3.0. The Apache Iceberg format tables, are used in big data and enable SQL querying. Query engines such as Spark, Trino, Flink, Presto, Hive, Impala, StarRocks, and others can work on the tables simultaneously. The tables are managed by metadata tracking and snapshotting changes.
Apache Hudi (Hadoop Upserts Deletes and Incrementals)is an open-source framework for building transactional data lakes with processes for ingesting, managing, and querying large volumes of data. Uber developed it in 2016 and Hudi became an Apache Software Foundation top-level project in 2019. It manages the storage of large analytical datasets on DFS (Cloud stores, HDFS or any Hadoop FileSystem compatible storage) and is a transactional data lake platform that brings database and data warehouse capabilities to the data lake.
…
DataCore Perifery.
DataCore says its Perifery Transporter is now shipping. The purpose-built physical media appliance, we’re told, provides local nearline storage and intelligently initiates long-term archiving by transferring content from remote and on-set locations to on-premises and cloud facilities. The content is fully encrypted and protected, even during transport, providing end-to-end data security. Automated preprocessing capabilities on the edge appliance save significant time and effort, eliminating the need for expensive, complex servers and other hardware, DataCore says.
…
Singapore-based DapuStor, a provider of high-end enterprise and data center SSD storage products, and Taknet Systems, a distributor of IT products and services in the APAC region, have announced a strategic partnership to offer enterprise SSD storage systems in the APAC market.
…
Multi-cloud data security supplier Fortanix has released Fortanix Confidential Data Search, which it describes as a high-performance product that supports highly scalable searches in encrypted databases with sensitive data, without compromising data security or privacy regulations. It is said to be thousands of times faster than current technologies and supports off-the-shelf databases to accelerate adoption.
…
NetApp has updated its BlueXP management control plane for on-prem and public cloud NetApp storage arrays and instances. BlueXP backup and recovery has a single control plane that simplifies customized backup strategies on a workload-by-workload basis. BlueXP backup and recovery capabilities have support for app-consistent database deployments in major clouds using either NetApp software-defined and hyperscaler-native storage offerings such as Oracle databases on Amazon FSx for NetApp ONTAP.
NetApp storage and data services can be deployed in highly secure, compliance-sensitive environments, including government clouds and “dark sites” with full isolation from internet connectivity via BlueXP private and restricted modes of deployment, NetApp says. Cloud Insights Federal Edition is now available for FedRAMP deployments. Cloud Volumes ONTAP (CVO) is available in the AWS Marketplace for the U.S. Intelligence Community (IC).
…
Quantum has announced the qualification of its backup storage portfolio with Veeam Backup & Replication (VBR) V12. Quantum offers a portfolio of Veeam Ready storage products including ActiveScale object storage with immutable object locking and versioning. Quantum integrated Veeam’s Smart Object Storage API (SOSAPI) into ActiveScale. Quantum’s DXi physical and virtual backup appliances integrate with the Veeam Data Mover Service (VDMS). Its Scalar Tape provides physically air-gapped storage to protect against ransomware and cyberthreats.
Quantum’s Myriad all-flash file and object storage platform is designed for rapid recovery of mission-critical data and expected to provide Veeam customers with an additional solution to minimise recovery time objective (RTO) and recovery point objective (RPO). Quantum intends to apply for the Veeam Ready program once it is generally available later this year.
…
RisingWave Labs, a startup founded in 2021, announced GA of RisingWave Cloud, its fully managed SQL stream processing platform, built on a cloud-native architecture. RisingWave Cloud consumes streaming data, performs incremental computations when new data comes in, and updates results dynamically at a fraction of the cost compared to Apache Flink.
RisingWave is built with Rust, and is optimized to support high-throughput and low-latency stream processing in the cloud. It guarantees data consistency and service reliability using checkpointing and shared nothing architecture. It has its own native storage for persisting data and serving user-initiated queries, and offers dozens of connectors to other open standards and computing frameworks, allowing users to freely connect to external systems.
…
SanBlaze announced availability of its SBExpress-DT5, a turnkey PCIe 5.0 NVMe SSD validation test system. It works in conjunction with the company’s Certified by SANBlaze Test Suite, bringing enterprise class NVMe validation to the developer’s desktop. The system is self-contained, quiet, and portable, good for a work-from-home environment or traveling to customer sites, the vendor says.
…
Seagate has formally announced its FireCuda 540 M.2 gaming SSD after many of its details were leaked last week. Maximum capacity is 2TB, not 4TB as leaked info suggested. It has a 10,000 MBps sequential write and read speed. There is a 1.8 million MTBF rating, a five-year warranty and three years of Seagate’s data recovery services are included.
…
Research house TrendForce says High Bandwidth Memory (HBM) is emerging as the preferred solution for overcoming memory transfer speed restrictions due to the bandwidth limitations of DDR SDRAM in high-speed computation. HBM is recognized for its transmission efficiency and plays a pivotal role in allowing core computational components to operate at their maximum capacity. Top-tier AI server GPUs have set a new industry standard by primarily using HBM. TrendForce forecasts that global demand for HBM will experience almost 60 percent growth annually in 2023, reaching 290 million GB, with a further 30 percent growth in 2024.
The Virtuozzo Hybrid Cloud offers managed service providers (MSPs) a set of cloud services it claims they can sell at at lower cost than those of the three cloud giants, and is pitching a 20 percent margin.
Switzerland-based Virtuozzo provides open source-based hyperconverged cloud technology which it claims enables MSPs to reduce costs for customers by as much as 25 percent. The Virtuozzo Hybrid Cloud brings together core cloud services in a single easy-to-use web interface for MSPs, we’re told.
Joe Morgan, Virtuozzo VP of Cloud, said: “No one needs today’s crazy public cloud prices, or the complexity of trying to piece together your own cloud service.”
Morgan claimed that when he was at an MSP, “I went through these exact same challenges. Nobody could give me a ready-to-sell solution for the cloud services my customers needed. I had to build everything myself, or lock my customers into one of the big hyperscale public cloud ecosystems. Virtuozzo Hybrid Cloud was built to overcome all of those problems. One cloud, designed for MSPs; all of the core cloud services an MSP’s customer needs; one easy-to-use interface; and pricing that actually makes sense for today’s economy.”
The Virtuozzo Hybrid Cloud (VHC) features:
Infrastructure, Platform, Kubernetes and Storage-as-a-service,
Flexible pricing: Pay As You Go, and/or commitment tiers to reduce cost/increase margin further,
Self-service portal for MSPs, and for end customers if desired,
Certified cloud infrastructure managed by Virtuozzo partners, with 24×7 support and 99.9 percent SLAs,
Instant activation, free online training.
VHC is designed for the channel, hosted by Virtuozzo and its network of Cloud Service Provider partners, and available now at locations across the Americas, Europe and Asia-Pacific. Distributor Climb Channel Solutions is leading its roll out in North America and the UK.
Startup Lightbits is offering full SAN-in-the-Cloud data services built on top of raw, ephemeral and fast NVMe instances for both the AWS and Azure clouds.
Lightbits is the fourth storage company I have come across that is taking advantage of public cloud ephemeral storage instances to provide block storage that is both faster than and more affordable than the cloud provider’s own block storage instances – EBS in AWS’s case. The other three vendors are Silk, Dell with PowerFlex and Volumez. Interestingly all four orgs have Israeli backgrounds.
Kam Eshghi.
Lightbits’ strategy was presented in a briefing by co-founder and CSO Kam Eshghi and chief system architect Abel Gordon.
Back in 2019, Lightbits, which was founded in 2015, had developed its NVMe/TCP-using SuperSSD Ethernet-attached all-flash appliance and hardware acceleration card. Its LightOS software could also run on commodity hardware and supply block storage as fast as, for example, all-flash arrays from Pure Storage and NetApp but at considerably lower cost.
In late 2022 it started talking publicly about porting its Lightbits software to the public cloud, as the Lightbits Cloud Data Platform, and demoed this at AWS re:Invent. The software provided block storage by using, it said, Intel-based storage-optimized Elastic Compute Cloud (EC2) instances. Normally SAN storage software ported to Amazon’s public cloud will use Elastic Block Store (EBS) instances and will provide a suite of data services on top of EC2, including compression, thin provisioning, snapshots, etc. The EC2 storage is raw and not persistent, going away when the instance is shut down.
Lightbits on AWS is now available, both as a bring-your-own-license (BYOL) offering and in AWS’s marketplace. It provides auto-scaling, thin provisioning, unlimited backup and restores, snapshots and clones and it can provide block storage to any EC2 instance type. This is block storage with a high IOPS rating and consistently low latency.
It is also, Eshgi said, much easier to set up and use than EBS storage, which has to be provisioned one volume at a time. Lightbits will support multi-availability zones for its AWS storage and AWS does not have that.
Azure
Lightbits is also available on Azure, in BYOL form but with a managed application marketplace offering coming later in the year. Gordon reckoned setting up Azure native block storage “is great for many use-cases but complex and costly for storage-intensive applications.” Lightbits setup is relatively easy and Gordon showed a performance slide of Lightbits’ Azure scalability and latency:
The left-hand chart is for reads and the green, blue and dark gray lines show the IOPS numbers for three, six, and nine servers respectively – which rise as the queue depth increases from 8 to 128. The nine-server IOPS line reaches almost three million IOPS.
The three bars (blue, orange and gray) show the 99th percentile tail latency for nine, six, and three servers respectively. They are all quite consistent, with only the rightmost 128-deep queue entry showing much variation.
The right-hand chart is for writes and shows higher latencies and lower IOPS – as would be expected.
Lightbits performance in Azure when supplying an Oracle database scales up to almost a million IOPS whereas Azure’s own Ultra Disk tops out at 80,000 IOPS.
An Oracle performance evaluation with Lightbits storage shows it reaching almost to one million IOPS when only reading, and progressively going down to over 250,000 IOPS when only writing. It is very much faster than Azure’s own Ultra Disk instances with their 80,000 IOPS when only reading.
These charts make Lightbts in the public cloud an attractive proposition. With Oracle for example, fewer Oracle licenses could be needed at a particular performance level. Lightbits also charges by cluster whereas an alternative Azure EBS 102 facility charges by provisioned IOPS and capacity. Oracle is limited to specific instance types and multiple Oracle instances can be needed to reach a target TPS number. Lightbits claims a single one of its volumes on Azure can support high Oracle TPS.
Customers with Lightbits systems on-premises can transfer their licenses to the public cloud Lightbits instances if they wish.
Workloads of interest to Lightbits in the cloud are databases and AI – specifically data pre-processing, model training, real-time inference, and database optimization.
Looking at the Lightbits competition we understand that Volumez’s offering is fundamentally a control plane, and doesn’t provide a full Lightbits-style storage stack. Silk is similarly a high-performance offering but with limited data services.
We should position Lightbits in AWS and Azure as providing a full storage stack with data services, high IOPS and low latency, whereas legacy SAN vendors porting their software to the cloud on EBS-style instances also provide data services – but not with the performance (IOPS and latency) of Lightbits.
Get a solution brief document discussing Lightbits in AWS here and a Lightbits on Azure solution brief here.
Bootnote.
The LightOS software name was retired in 2021 in favour of calling it just Lightbits software.
Hitachi Vantara has entered an OEM deal with CTERA for file and data management services with the Hitachi Content Platform storing the data.
CTERA provides edge-to-core-to-cloud global file system and management services that can be based in the public cloud or use an on-premises object store for its data. Accessing users in datacenters or remote edge locations have caching edge filer facilities to speed file IO. The Hitachi Content Platform (HCP) has all-flash and disk-based nodes and supports S3. Hitachi V likes OEM and similar deals for HCP. In July 2020, it signed up to OEM WekaIO’s fast file system software, integrated with HCP. It signed a deal in January with Model9 to feed mainframe data to its HCP and VSP 5000 storage systems.
Oded Nagel
CTERA CEO Oded Nagel said: “In working with Hitachi Vantara, we are delivering best-in-class file services and comprehensive data management solutions. We are delighted to forge this partnership enabling businesses to seamlessly manage their data across the entire lifecycle, from edge-to-cloud, with enhanced security and scalability.”
Hitachi Vantara’s SVP for Product Management and Enablement, Dan McConnell, added: “We are excited to collaborate with CTERA to continue to build out our multi-cloud data services. This collaboration empowers our customers with unparalleled control, security, and scalability in managing their distributed data across multi-cloud environments.”
Dan McConnell
The collaborative CTERA-Hitachi Vantara product will be marketed globally as Hitachi Content Platform Anywhere Enterprise (HCP Anywhere Enterprise). There will be a fully integrated migration path for customers using legacy HCP Gateway, Hitachi Data Ingestor or other discrete network-attached storage (NAS) systems. Effectively, it appears, CTERA replaces them, covering primary and secondary edge-to-core-to-cloud file services and machine-generated data workflows.
In May, CTERA was involved with Hitachi’s announcement of its Hitachi Data Ingestor (HDI) reaching end-of-life. CTERA introduced CTERA Migrate for HDI, a turnkey migration path that replaced HDI and preserved the existing storage repository investment.
HCP Anywhere Enterprise provides network-attached storage (NAS) capabilities for SMB and NFS, archival storage, enterprise file sync ‘n’ share, AI-based ransomware detection, malicious user blocking, rapid rollback and immutable storage (WORM).
It has scalability to petabytes of data with central multi-tenant administration and the HCP object-storage backend. It has defense-grade protection and 100 percent behind-firewall deployment for security-conscious customers. It is also integrated with Hitachi’s Lumada Data Catalog. With this customers can meet regulatory compliance requirements by cataloging, categorizing and setting data management policy for distributed edge data.
Hitachi has an existing HCP Anywhere product, launched in 2013, which provides file-sync-and-share capabilities, endpoint data protection, and enterprise data mobilization. We think this may be absorbed into HCP Anywhere Enterprise.
Hitachi Vantara, by using CTERA, gets its own defense against inroads into its customer base from Nasuni and Panzura. It also bolsters the HCP product’s role as a central repository for mainframe extracted data (Mode9), parallel file system data (WEKA) and now cloud-style file services data (CTERA). For its part CTERA gets a further leg up in credibility and, hopefully, a nice reliable revenue increase. It’s an excellent deal for both companies.
Host-based SSD controller array startup Pliops has a new CEO – Ido Bukspan coming from Nvidia.
Pliops, founded in Israel in 2017, has built its XDP (eXtreme Data Processor) to offload and accelerate low-level storage stack processing from a host x86 server. The XDP has added code to run RocksDB storage IO faster and also to function as a RAID controller. It’s sold to hyperscaler and near-hyperscaler customers with thousands of servers for whom extra performance means more cores can run application code.
Co-founder and former CEO Uri Beitler said: “I am very proud of Pliops’ achievements to date and am excited to work closely with Ido, who has an impressive track record. I have no doubt that Ido will successfully lead Pliops forward and help implement our product roadmap and integrations with our customers.”
Ido Bukspan (left) and Uri Beitler
Beitler becomes chief strategy and business development officer. He’ll develop strategic partnerships and create new business opportunities.
Bukspan was SVP of Chip Design at Nvidia, joining in 2019 when Nvidia bought his then-employer Mellanox for $6.9 billion. He spent more almost 20 years at Mellanox.
There are hundreds of employees in Nvidia’s Chip Design operation and we’re told Bukspan led it to impressive results. Nvidia said it was expanding its chip operations in Israel in January 2022. It aimed to develop new processors and recruit hundreds of engineers. Bukspan has walked away from this, the world’s foremost GPU and AI hardware supplier, to join the much smaller Pliops, with just over 100 employees.
The startup’s chairman, co-founder Aryeh Mergi, said: “In his position as CEO, Uri built and led Pliops from the day it was established to many outstanding achievements. Now, in bringing on Ido, we are adding someone with substantial experience in our target markets, and I have no doubt that the impressive career Ido built at Nvidia and his leadership capabilities will help Pliops accelerate its current positive momentum to increase growth and expansion into additional markets.”
Pliops has been associated with running its software on Nvidia’s BlueField DPUs. Perhaps Bukspan will help bring this to fruition. Pliops is also partnering with IaaS provider phoenixNAP to offer petabyte-scale Redis clusters to users with in-memory-class performance. We can look forward to more of such partnerships as Beitler gets busy.
Lakehouse developer Databricks is buying generative AI startup MosaicML for $1.3 billion so its customers can build and deploy AI models on their own data.
Large language models (LLMs) have ushered in AI that can understand queries, analyze multiple data sources and respond with natural language answers or even produce programming language. However, they can also produce wrong or imaginary answers, and need significant GPU resources to run. MosaicML helps customers run on minimal systems, and have their models trained with their own and not generally available public data.
Databricks CEO Ali Ghodsi said: “Every organization should be able to benefit from the AI revolution with more control over how their data is used. Databricks and MosaicML have an incredible opportunity to democratize AI and make the Lakehouse the best place to build generative AI and LLMs.”
In April Databricks revealed its updated open source Dolly LLM to make its AI facilities available for business applications without needing massive GPU resources or costly API use. The chatbot can be used to generate queries that run against Databtricks’ lakehouse.
Naveen Rao and Hanlin Tang
MosaicML was founded in 2021 by CEO Naveen Rao, a former VP and general manager of Intel’s AI Products Group, and CTO Hanling Tang, previously the senior director of Intel’s AI Labs. It has pulled in $64 million in funding. MosaicML’s open source LLMs are based on its MPT-7B architecture, built with 7 billion parameters and a 64,000 token context window.
There have been over 3.3 million downloads of MPT-7B and the recent release of MPT-30B. The latter is significantly more powerful than MPT-7B and outperforms the original GPT-3. MosaicML says the size of MPT-30B was specifically chosen to make it easy to deploy on a single GPU – either 1xA100-80GB in 16-bit precision or 1xA100-40GB in 8-bit precision. MosaicML says other comparable LLMs such as Falcon-40B have larger parameter counts and cannot be served on a single datacenter GPU; this necessitates 2+ GPUs, which increases the minimum inference system cost.
The US firm has offices in San Francisco, New York, Palo Alto and San Diego. Customers include AI2 (Allen Institute for AI), Generally Intelligent, Hippocratic AI, Replit and Scatter Labs.
Databricks says that bringing in MosaicML’s tech will offer its customers a simple and fast way to retain control, security and ownership over their data without incurring high costs.
MosaicML’s optimization provides 2-7x faster model training compared to standard approaches and is linearly scalable. It claims multibillion-parameter models can be trained in hours, not days.
The entire MosaicML team is expected to join Databricks after the transaction closes. MosaicML’s platform will be supported, scaled and integrated over time. Customers will get a unified platform on which they can build, own and secure their generative AI models, training them with their own data.
Rao said: “We started MosaicML to solve the hard engineering and research problems necessary to make large scale training more accessible to everyone. With the recent generative AI wave, this mission has taken center stage. Together with Databricks, we will tip the scales in the favor of many – and we’ll do it as kindred spirits: researchers turned entrepreneurs sharing a similar mission.”
The proposed acquisition is subject to customary closing conditions, including any required regulatory clearances. Other generative AI startups may now be getting approaches from Databricks’ competitors.
Bootnote
Databricks is itself a startup, having been founded in 2013 and raising $3.6 billion across multiple rounds. The acquisition cost includes some retention packages for MosaicML employees.
Data protector NAKIVO has added bare-metal recovery, a malware scan capability plus direct recovery from tape in v10.9 of its Backup & Replication software.
The company provides backup, ransomware protection and disaster recovery for virtual, physical, cloud and SaaS environments. It claims to have had 28 consecutive quarters of double-digit growth, and says it has a network of over 7,500 partners and 25,000 paid customers worldwide, including Honda, Cisco, Coca-Cola and Siemens. In September it said it had 23,000 customers, up from 16,000 at the beginning of 2021, and 7,000 partners, up from 5,000 at the start of 2021. Customers include direct end users and MSPs.
Bruce Talley
NAKIVO is an established backup player and, like the others, has to constantly and incrementally improve its offering as backup data sources proliferate and recovery needs expand. Co-founder and CEO Bruce Talley reflected this, saying: “Today’s highly competitive digital environment makes high availability a new norm. We keep improving our solution to provide everything our customers need to meet their objectives while maintaining the required level of security.”
Recovery to a bare-metal, new-environment server is about restoring an entire server or workstation from a backup to identical hardware without rebuilding the operating systems or reconfiguring application settings. This extends the existing physical-to-virtual machine recovery function.
It’s now thought necessary to check backups for malware to avoid restoring a system with in-built malware.
Commvault announced its Threat Scan feature earlier this month, which looks for corrupted or suspicious datasets, and can locate and quarantine malware and threats from backup content.
Veritas v10, released in February last year, provides automatic malware scanning during backups and prior to restores. It has AI-driven anomaly detection, which can automatically initiate malware scanning.
NAKIVO’s Backup Malware scan feature enables customers to integrate external anti-malware software and scan backups for malicious code before using them for recovery. It is an add-on to the existing anti-ransomware features. Immutable backups should prevent this happening, unless the backed-up data contains malware.
The company provides immutable backups for Amazon S3, generic S3-compatible storage (Beta), Wasabi, Azure Blob Storage, or Backblaze B2 Cloud Storage and also local folders. Object Lock or version-level immutability support should be enabled for the bucket or blob container used to store backups. This type of immutability cannot be shortened or lifted, not even by the root user.
With the local folder type of Backup Repository, immutable recovery points cannot be overwritten, deleted, or changed by anyone except the root user before the specified period expires. But when deployed as part of a VMware vSphere, Nutanix AHV virtual appliance, or a pre-configured AMI in Amazon EC2, users can make recovery points stored in this type of repository immutable, and the immutability cannot be lifted or changed by anyone, not even the root user.
v10.9 enables faster recovery from tape of full VMs, EC2 instances, and physical machines to VMware VMs, without using a staging repository first. It also adds support for VMware vSphere 8.0 U1 and Debian v10-11.
NAKIVO Backup & Replication v10.9 is now available for download and a 15-day free trial is available.
Google has announced GA of BigLake support for Apache Iceberg, which has been in preview since October 2022. BigLake is a storage facility that unifies data warehouses and lakes and allows BigQuery and open source frameworks like Spark to access data with fine-grained access control. It provides accelerated query performance across multi-cloud storage and open formats such as Apache Iceberg. BigLake Metastore provides shared metadata for Iceberg tables across BigQuery and open source engines, eliminating the need to maintain multiple table definitions. You can provide BigQuery datasets and table properties when creating new Iceberg tables in Spark, and those tables become automatically available for the BigQuery user to query. More information here.
… Decentralized storage provider Filebase tells us that it now has over 35,000-plus users on the platform. The uploaded file count has surpassed 1.2 billion files (live in-production for four-plus years). It has launched GA support for its Dedicated IPFS Gateway product and also recently launched its Filebase for Startups program, which now has about 50 entries, to provide Enterprise Managed IPFS Resources. Here is a NFT Metadata Storage Provider Comparison in which Filebase shines. …
iStorage has announced its USB 3.2 datAshur PRO+C, with the Type-C interface, the world’s first and only flash drive pending the new FIPS 140-3 Level 3 validation scheme. It employs PIN protection and hardware encryption to safeguard data to military standards. Capacities range from 32GB to 512GB. Built into the drive is a rechargeable battery, allowing the user to enter an 8 to 15-digit PIN via the onboard keypad before connecting the device to a USB port. The datAshur PRO+C cannot be accessed without the user’s unique PIN.
If the user PIN is entered incorrectly 10 consecutive times, the user PIN will be deleted. All data will remain on the device but can only be accessed by entering the admin PIN. If the admin PIN is entered incorrectly 10 times, the PIN, encryption key and data will be lost forever, and the device will revert to factory default settings.
…
The Piql-based digital storage vault in Svalbard is being used by the National Library and Archives (NLA) of the UAE, and Piql’s partner Melara Middle East. The vendor says the National Library and Archives of the UAE will be the first archival center in the Arab world and the Middle East to contract with Piql, obtaining its secure, immutable and permanent data storage technology. The vault will include original data conversion and preservation services securely for long periods of time – up to a claimed 2,000 years, says the vendor – including digital preservation services for the UAE, the GCC countries, the Middle East and North Africa.
…
System-on-chip builder SMART Modular’s DC4800 NVMe PCIe gen 4 SSD product in U.2 and EDSFF E.1S formats have been accepted by the Open Compute Project as an OCP Inspired product and will be featured on the OCP website in the Marketplace section. They have capacities of up to 7.68TB.
…
Stravito is a Swedish SaaS startup (2017) which provides a central place to store and analyze market research data. Its customers include McDonald’s, Comcast, Burberry and Danone. It has announced proprietary generative AI capabilities to improve the search experience and provide users with quicker insights. Users can ask full questions using everyday short natural language and get answers synthesized from various proprietary sources. They can also dive deeper with recommended questions – all with the security of a closed system that generates accurate answers exclusively from a client’s owned research materials, and not beyond. Users can now ask Stravito any question, of any complexity, and quickly receive complete answers that aggregate insights from various sources.
…
Taiwan-based TeamGroup has launched a set of industrial wide temperature (a temperature range of -40-85°C) storage products called ULTRA. There is DIMM memory from 8GB to 32GB, available in two types: DDR4/DDR5 U-DIMM and SO-DIMM. They comply with JEDEC frequency standards and can reach transmission rates as high as 5,600MT/s. An ULTRA SSD comes in the PCIe Gen 3×4 2280 form factor with a capacity range of 128GB to 1TB. The maximum read and write speed is 2,100MB/s and 1,700MB/s, respectively. The products are available for sampling. PCIe interfaces are replacing SATA III in the automotive embedded and other harsh temperature environments.
…
China’s TerraMaster is announcing its Centralized Backup and Duple Backup. Centralized Backup uses a TerraMaster NAS as the store for backups of virtual machines, servers, PCs and Macs. It supports multiple types of device backup and recovery, allowing backups to be made of the data from dozens or even hundreds of PCs, servers, or virtual machines with only one TNAS. Duple Backup is a disaster recovery tool for TNAS devices. Important folders or iSCSI LUNs in TNAS can be backed up to multiple destinations, including a remote TNAS device, file server, or cloud disk, using an intuitive user menu. It also supports multiple backup strategies such as incremental backup and multi-version backup.
…
TrendForce reports that explosive growth in generative AI applications like chatbots has spurred significant expansion in AI server development in 2023. This will drive up demand for HBM during 2023-2024. Presently, Nvidia’s A100 and H100 chips each boast up to 80 GB of HBM2e and HBM3. In its latest integrated CPU and GPU, the Grace Hopper Superchip, Nvidia expanded a single chip’s HBM capacity by 20 percent, hitting a mark of 96GB. AMD’s MI300 also uses HBM3, with the MI300A capacity remaining at 128GB like its predecessor, while the more advanced MI300X has ramped up to 192 GB, marking a 50 percent increase. TrendForce predicts that AI accelerator chips that primarily utilize HBM will have a total HBM capacity of 290 million GB in 2023 – a nearly 60 percent growth rate. This momentum is projected to sustain at a rate of 30 percent or more into 2024.
…
There’s some good news for WANdisco, which has been recapitalizing and struggling to regain its stock market listing after a senior sales rep devastated the company’s reported 2022 earnings with possibly fraudulent sales reports. It announced a new two-year agreement valued at $113,125 with Accenture to use WANdisco Data Migrator to support the data modernization program for a leading Australian bank. The company expects to recognize revenue from Q2 2023.
…
Western Digital, following critical security vulnerability CVE-2022-36327, has blocked access to its cloud services by My Cloud devices unless users update their firmware to My Cloud OS version 5.26.202. It fixes this bug as well as three other medium-severity issues. This affects Western Digital’s My Cloud Home, My Cloud Home Duo, SanDisk ibi, and My Cloud OS 5 devices. My Cloud OS 5 users can still access their data on these devices locally, but My Cloud Home, My Cloud Home Duo, and SanDisk ibi devices will not be able to access their data until they update to the latest firmware release.
The first Seagate PCIe gen 5 SSD, FireCuda 540, is incoming and we know because Amazon and another etailer leaked the details before Seagate had announced the drive.
Tom’s Hardware first noticed the offending webpages at Amazon UK and B&H, which were soon were closed down. Cue red faces in Seagate’s online sales department. We have asked the company to comment.
Seagate’s FireCuda products – such as its FireCuda 530 and 520 and 510 – are gaming SSDs in the M.2 gumstick format. The higher the product number, the newer and faster the drive. A table shows the progression, and has the 540’s details, such as we know, added:
The controller is assumed to be Phison’s PS5026 E26 product as Phison has announced a demo PCIe gen 5 SSD using it, and Seagate has a strong history of using Phison controllers for its FireCuda drives.
The 540 is half as fast again in terms of IOPS as the 530, with random read/write IOPS listed at up to 1,500,000. The sequential read/write bandwidth is given as up to 10GBps, a 37 percent increase on the 530’s maximum sequential read number.
An Apacer PCIe gen 5 gaming SSD uses the same Phison controller to achieve 13GBps sequential reads and 12GBps sequential writes. So too does Gigabyte’s Aorus Gen5 10000 SSD, supporting PCIe 5 and with 200-plus layer 3D NAND in 1, 2 and 4TB M.2 2280 format. It delivers 1.3 million/1.16 million random read/write IOPS and 12.5/10.1GBps sequential read/write bandwidth. The FireCuda looks a tad light on the sequential bandwidth front.
The 2TB 540 is rated at 2,000 TB written endurance during its five-year warranty period, with the 4TB model reaching 3,949 TB written. The etailer webpages said the drive has a graphene heat-spreader instead of a heat sink. The US price for the 1TB version is $189.99 and the 2TB 540 is $319.99.
MinIO doubled down on its open source, Kubernetes-native object storage at an IT Press Tour briefing.
The company said its object storage software was frequently downloaded (1.2 billion Docker pulls) and widely deployed. Its thesis starts with the point that the cloud is taking over virtually every IT environment. Kubernetes powers the cloud operating model, providing one abstraction across any computing infrastructure – public, private or edge.
CMO Jonathan Symonds claimed: “We are the world’s fastest object store and no one has challenged us.”
He justifies the claim by citing GET speeds of 325GiBps and PUT speeds of 177 GiBps on standard hardware.
The primary storage format or protocol in Kubernetes environments is object, not block or file. The cloud runs on RESTful APIs. POSIX is legacy at this point as are the storage classes that depend on it.
MinIO says properly architected systems offer performance at scale – a requirement for AI/ML workloads. MinIO object storage is built for this, with erasure coding and S3 Select support. It asserts that, despite decades of evolution, file and block storage, lacking these things is not good enough for modern cloud environments.
Symonds said: “We don’t see SAN and NAS as being natural in the cloud world.”
In a nutshell, its position is that EC2 and file in the cloud cannot scale as easily as object. iSCSI as a service and TCP are both too chatty for cloud, it claims, and NFS v4 and pNFS don’t scale. S3 came much later, dropped the legacy baggage, and became primary. RAID doesn’t scale, MinIO feels.
Basically, MinIO is convinced its object software will wash over everything else in the storage world, including on-premises file, block and object arrays, and their semi-detached, CSI-based cloud incarnations. Why? Because it is fast and Kubernetes-native, and therefore suited for storage provision in the cloud.
MInIO slide images
Being Kubernetes-native matters because the entire storage software runs as a container inside of Kubernetes, with complete support for all Kubernetes functionality, it says.
Symonds said file and block are in the cloud for legacy apps, and file and block have to go through the kernel and so have CSI drivers. MinIO’s object store runs as a container and doesn’t need CSI.
MinIO claims S3 is the default API for object storage and MinIO is the leader in compatibility. It claims to have been the first to market with V4 and one of the few to support S3 Select.
Its operator provisions multiple MinIO clusters on top of Kubernetes. It brings stateful services to the world of Kubernetes, which doesn’t understand the concept of state, and makes having multiple tenants with multiple versions straightforward. Operators enable seamless integration with the broader K8s ecosystem, giving access to service discovery, container orchestration, and monitoring systems.
The operator works with Amazon EKS, AKS, GKE, OpenShift, RAFAY, the Rancher Kubernetes Engine and VMware Tanzu. V5.0 brings in zero trust. It’s workload-centric and workload identity can be validated by MinIO Operator with help of K8s. V5 has rotating credentials control. All access granted to object store is temporary and there is no need to worry about rotating credentials ending up with the wrong user/workload.
There is dynamic local persistent volume (PV) provisioning as opposed to network PV provisioning.
Scale
MinIO engineer Dileeshvar Radhakrishnan said unstructured data will represent 90 percent of all of data by 2025. MinIO storage can be used as a data lake and is suited to storing data for AI/ML processing. AI predominantly uses object storage and MinIO said it was ready for any ChatGPT-led acceleration.
The petabyte is the new terabyte, as it were, and we are moving into exabyte territory. Data lakes use object storage, not SAN, NAS or block storage, because they don’t scale.
But PB/EB-scale data is often distributed and must be seen as a single element instead of being viewed as siloed. Public cloud economics don’t work at EB scale; it is simply too costly. Data has to be on-premises or in a colo, and stored with MinIO software. This uses Kafka. It has an Iceberg driver which takes data and stores it in MinIO ISO tables.
In Radhakrishnan’s view: “Basically we are the poster child for the modern data lake.”
We asked the MinIO presenters about about enterprise customers of VAST Data, Pure Storage, etc. with commitment to Nvidia and its GPUDirect protocol. MinIO, which doesn’t support GPUDirect, thinks these customers have made the wrong choice, but they won’t listen to MinIO.
It will leave these customers of legacy vendors like Dell, NetApp, Pure, VAST, etc, on their own. They will stay with their choices and MinIO will not address them. Instead it will let this smaller SAN/NAS market wither away over time, while it grows in the the much larger object market. This will consolidate, with Kubernetes acting as the convergence point.
Defending open source
MinIO has been vocal in acting against suppliers such as Nutanix and Weka.
Symonds said: ”We are deeply committed to open source.”
He said this about the actions against Nutanix and Weka: “We see this as defending open source.” MinIO wants them to play by the open source rule book, meaning its licensing rules.
Bootnote
Operators are software extensions to Kubernetes that make use of custom resources to manage applications and their components. Operators follow Kubernetes principles, notably the control loop. They are clients of the Kubernetes API that act as controllers for a custom resource.
