Blog

What makes flash memory really fast?

Flash is seen as the beacon of hope for high-performance storage systems.But the most powerful flash arrays can only play to their true strengths with SCM (Storage Class Memory).

Flash aka Flash EEPROM is an electrically erasable read-only memory (EEPROM for short, Electrically Erasable Programmable Read-Only Memory), i.e. a non-volatile computer storage medium that can be rewritten, a so-called NVM (Non-Volatile Memory). Pro Face Touch Screen

What makes flash memory really fast?

Flash stores information (bits) as electrical charges on a floating gate or in a charge-trapping storage element of a so-called MISFETS (Metal Isolator Solid State Field Effect Transistor) - permanently and without consumption of maintenance energy.The storage of data that has already been saved does not require any additional energy supply - unlike in the case of volatile (working) memory.

One of the peculiarities of flash is that it requires an erase operation to rewrite previously used memory cells.Compared to read access, this process is very energy-intensive.

A floating gate transistor consists of two gates: a floating gate and a control gate.The floating gate is isolated from the rest of the transistor structure and is usually made of polysilicon.The control gate is a very "ordinary" transistor gate.Writing to the memory cell is accomplished by applying a pulse to the control gate, causing electrons to tunnel into (or out of) the floating gate.The presence (or absence) of charges modifies the transistor's threshold voltage, a shift called the memory window.The information is thus encoded in the threshold voltage of the floating gate transistor, and it is read out by measuring the leakage current.The charges stored in the insulated gate remain unchanged over long periods of time, which is what gives the memory its non-volatile character.

In a charge trap, the junction is an insulator (usually silicon nitride) that reduces electrostatic interference between adjacent cells.This charge trap cell is the basis of most 3D NAND structures today.

In order to write to flash, stationary electrical charges must be changed in complete electrical isolation from other parts.This change in the charge state is possible due to the quantum mechanical tunnel effect (the so-called Fowler-Nordheim tunnel effect): It allows electrons to pass through the high potential barrier of the non-conductor of the insulating layer.

Memory cells are the focus of many current research efforts worldwide.The Fowler-Nordheim tunneling effect in silicon-based NVMs forces a compromise between the properties of access speed, durability and resilience.GaN-on-Si technology could overcome this "trilemma", as researchers from HKUST University in Hong Kong have shown.

Flash technology is evolving rapidly at all levels of memory architecture.The type of flash module is not primarily determined by the manufacturing technology, but by the way the flash memory cells are connected.Two alternative architectures currently dominate the market: NAND and NOR.

The type of connection of memory cells - NAND or NOR - determines the key data of the resulting flash modules with regard to the achievable capacities, access times, production costs and service life, among other things.

The memory cell circuitry in NOR flash is similar to the structure of a NOR gate in a semiconductor.Each memory cell has one end connected to the source line and the other end connected directly to a bit line.This parallel connection allows random access to data, thereby minimizing read access latency (compared to NAND flash), but increases manufacturing costs.At the same time, doing without common components reduces the storage density and thus the capacity.

NAND flash has multiple memory cells (usually eight) connected in series, reminiscent of a NAND gate.To maximize storage density, an entire group of memory cells shares a common data line.For this reason, read and write operations on individual memory cells cannot be random, but must instead take place in blocks - similar to hard disks.NAND memory therefore shows its strengths primarily in write processes.Benefits include higher storage capacities and lower storage costs;the access times recede into the background.

By clicking on "Subscribe to newsletter" I agree to the processing and use of my data in accordance with the declaration of consent (please open for details) and accept the terms of use.I can find more information in our privacy policy.

It goes without saying that we handle your personal data responsibly.If we collect personal data from you, we process it in compliance with the applicable data protection regulations.You can find detailed information in our data protection declaration.

I agree that Vogel IT-Medien GmbH, Max-Josef-Metzger-Straße 21, 86157 Augsburg, including all companies affiliated with it within the meaning of Sections 15 et seq. AktG (hereinafter: Vogel Communications Group) my E e-mail address for sending editorial newsletters.Lists of the respective associated companies can be accessed here.

The content of the newsletter extends to the products and services of all the companies mentioned above, including, for example, trade journals and specialist books, events and trade fairs as well as event-related products and services, print and digital media offers and services such as other (editorial) newsletters, competitions, lead campaigns, Market research in the online and offline area, subject-specific web portals and e-learning offers.If my personal telephone number was also collected, it may be used for submitting offers for the aforementioned products and services from the aforementioned companies and for market research.

If I call up protected content on the Vogel Communications Group portals, including its affiliated companies within the meaning of §§ 15 ff. AktG, I have to register with additional data for access to this content.In return for this free access to editorial content, my data may be used in accordance with this consent for the purposes stated here.

I am aware that I can revoke this consent at any time for the future.My revocation does not affect the legality of the processing carried out on the basis of my consent up to the time of revocation.In order to declare my revocation, I can use the contact form available at https://support.vogel.de as one option.If I no longer wish to receive individual newsletters to which I have subscribed, I can also click on the unsubscribe link at the end of a newsletter.I can find more information about my right of withdrawal and how to exercise it, as well as the consequences of my withdrawal, in the data protection declaration, section Editorial newsletters.

The number of bits that a NAND cell can hold determines the type of NAND memory and its technical characteristics.Common types include SLC, MLC, TLC, and QLC.They differ in cost, capacity, and lifetime (that is, the number of program/erase cycles a flash cell can go through).

The most popular architecture for arranging memory cells in NAND modules is TLC flash (triple-level cell).This architecture uses eight voltage levels to store three bits per cell.TLC scores particularly well with write-heavy workloads.

Read-intensive workloads can be addressed more efficiently with QLC (quad-level cell) flash.This architecture uses 16 voltage levels to store four bits per cell, but requires advanced error correction and other advanced features such as wear leveling and overprovisioning.

The flash market was dominated for a while by NAND memory cells, originally based on floating gate transistors in a planar array configuration.A notable innovation was the replacement of the floating gate cell with a charge trap, allowing for a simplified process flow.

One promising approach to increasing bit storage density is 3D NAND. The idea is to stack memory cells vertically to achieve higher density per unit area to increase capacity and keep costs down.

3D NAND is generally characterized by higher production costs than 2D NAND (English planar NAND), but can boast a higher storage density.Vendors like Micron have now arrived at 176 layers of 3D NAND.SK Hynix has been shipping its 176-layer, 512-gigabit TLC (triple-level cell) flash to SSD controller manufacturers under the codename "4D NAND" since the end of 2020.

An SSD (Solid State Drive) is a component that organizes memory modules for the use of a common system interface.An SSD is basically a drive that uses semiconductor modules to store data and does not have any moving mechanical parts.Flash is the most common storage medium on which modern SSDs are based.

A memory cell's latency is a measure of how long the CPU waits for the data it previously requested;latency varies with variables such as application type or use case, and processor and memory architecture.To reduce latency when accessing flash memory, leading flash array vendors have given their systems a turbo boost: non-volatile memory that is faster than flash, namely SCM (Storage Class Memory).

In general, DRAM boasts a latency in the 15 to 100 nanosecond range, while NAND has a latency of 80 to 120 microseconds.SCM closes the gap between the two readings.

Storage class memory is persistent data-storing-capable memory that approaches the operating speed of DRAM but can retain data even in the event of a power failure.

Storage class memory (SCM) typically combines NAND flash modules to ensure data persistence with dynamic random access memory (DRAM) as a turbo booster and power source.SCM treats NVM like DRAM: it adds it to a system's memory.

SCM includes storage technologies such as 3D XPoint from Intel and Micron Technology, RRAM/ReRAM (Resistive RAM) from Fujitsu and Crossbar, MRAM (Magnetoresistive RAM) and STT-MRAM (Spin-transfer torque MRAM) from Everspin Technologies, and NRAM (Nanotube RAM). ) from Nantero.Some of these technologies, most notably Intel Optane, have found their way into high-performance all-flash arrays from leading vendors, but many others are still in their infancy.

Some architectures of SCM (in particular, as well as those of non-volatile latency-optimized memory in general) can now even rival DRAM in terms of latency.This category includes MRAM (Magnetoresistive Random Access Memory).Other technologies achieve a latency that exceeds that of DRAM, but still beats flash-based SSDs by far.These include technologies such as PCM (phase change memory) and ReRAM (resistive random access memory).

The first flash-based storage devices used SATA (Serial Advanced Technology Attachment) based on AHCI (Advanced Host Controller Interface) or SAS (Serial-Attached SCSI aka Small Computer System Interface).These standards were tailored to meet the needs of HDDs and were developed with consideration for characteristics such as rotation delay, head seek times, and other mechanical drive specifics.If you want to take full advantage of flash storage, you have to shed the baggage of this legacy.

Flash storage is catching on

Instead of traditional data transfer protocols such as SAS and SATA, NVM Express (Non-Volatile Memory Express or NVMe, alternative designation: NVM Host Controller Interface Specification) is appearing in conjunction with the high-performance PCIe 4.0 transmission interface.The current version of NVMe is labeled 2.0b (Revision January 6, 2022).

The NVMe specification defines how host software and hardware communicate with non-volatile storage over a PCI Express (PCIe, short for Peripheral Component Interconnector Express) bus.

NVMe reduces CPU overhead and streamlines operations, lowering latency and increasing IOPS (input/output operations per second) and throughput.

Both SAS and SATA (ACHI) can only handle a single queue at a time.While SAS can handle just 254 commands and SATA (AHCI) 32 commands per queue and are each limited to a single queue, NVMe can handle queue depths of up to 65K.Those are 65K queues with up to 65K instructions each.In addition, there is also the lower latency.

The benefits of NVMe over legacy assets can be summed up as follows:

Incidentally, NVMe can perform over 1 million I/O operations per second (IOPS).

As a result of these many improvements, NVMe SSDs offer write speeds of up to 7,000 MB/s.So they are up to 12 times faster than SATA SSDs and up to 70 times faster than HDDs.

NVMe SSDs (Solid State Drives) come in different sizes;one form factor is better for booting, the other for maximizing performance, capacity, or system density.NVMe is the current industry standard for PCIe SSDs in all form factors: U.2 (old designation: SFF-8639), M.2, AIC and EDSFF (E1.S and E1.L).

With EDSFF (Enterprise & Data Center SSD Form Factor), a new form factor for enterprise-class storage devices has emerged, which is intended to overcome the limitations of the existing interfaces in data centers.The special highlights include the compact design and the hot-plug capabilities of EDSFF SSDs.

EDSFF currently comes in one of two versions: E1.S and E1.L.The E1.S uses a x4 interface and is designed for vertical installation in a 1U housing.The E1.L uses a maximum of one x8 interface and fits vertically in a 1U chassis;it is about three times as long as an E1.S.

The M.2 form factor (formerly known as the Next Generation Form Factor or NGFF) is a more modern type of internally mounted expansion card that replaces the previous mSATA (mini-SATA) standard.In contrast to standard HDDs and SSDs, M.2 drives are not connected to the motherboard via a cable, but plugged into a special M.2 slot directly on the board.

NVMe SSDs use the PCIe interface.However, the NVMe standard is not limited to the mere connection of Direct Attached Storage (DAS for short), but can also control storage devices via the network.Here we are talking about the so-called NVMe over Fabrics (short: NVMe oF).

The NVMe oF protocol specification extends the spectacular performance of NVMe from the storage array controllers to the network fabric, taking advantage of RDMA (Remote Direct Memory Access), Fiber Channel (FC-NVMe, NVMe over FC), or TCP.

Instead of the PCIe bus used by NVMe, NVMe oF takes advantage of an alternative data transport protocol over fabrics as transport mapping.Fabrics are based on the concept of message transfer between a sender and a receiver with no shared memory between the endpoints.The NVMe commands and responses travel across the fabric encapsulated as messages.

NVMe over Fabrics (NVMe oF) transforms traditional data center network infrastructure into scalable storage fabrics that combine the performance of NVMe-based DAS devices with the efficiency of shared storage infrastructure.

Typical RDMA implementations include RDMA over Converged Ethernet (RoCE), InfiniBand, iWARP, Omni-Path, and the Virtual Interface Architecture.Among the RDMA protocols, RoCE—a new breed of SAN—offers unique advantages.

The acronym RoCE (pronounced as rocky) stands for RDMA (Remote Direct Memory Access) over Converged Ethernet (i.e. Ethernet with the so-called Data Center Bridging Enhancements, DCB).RoCE is a standard protocol overseen by the InfiniBand Trade Association.Its task is to enable direct access to the working memory of a remote system.To do this, it encapsulates InfiniBand transport packets via Ethernet and transmits them using UDP.Rather than wasting valuable CPU cycles coordinating network communications, RDMA routes the needed data directly between different memory areas of the distributed application.

*The author duo Anna Kobylinska and Filipe Pereira Martins work for McKinley Denali Inc. (USA).

Demanding applications use their storage to the limit.Today's data-centric workloads demand uncompromising peak performance: record-breaking access latencies, weighty data throughput, and unprecedented concurrency.Flash storage runs like hot cakes.Our new eBook offers you an update.The topics at a glance: Continually high demand for flash storage Market overview All flash arrays Flash storage technology Market overview Scale-out systems

What is NAND flash (NAND flash memory)?

Storage innovations: Kalray, Micron and HYCU

NVMe flash is increasingly becoming storage mainstream

Terms and Conditions Cookie Manager Help Customer Center Media Data Data Protection Imprint & Contact Authors

Copyright © 2023 Vogel Communications Group

What makes flash memory really fast?

Servo Motor This website is a trademark of Vogel Communications Group.You can find an overview of all products and services at www.vogel.de