Electronics Guide

Subcategories

The Memory-Storage Gap

The tiers of a modern system differ by orders of magnitude in both latency and cost per bit. DRAM answers a load in tens of nanoseconds but stores each bit as charge on a tiny capacitor that leaks, so every cell must be refreshed on the order of every 32 to 64 milliseconds and all contents vanish when power is removed. NAND flash reads a page in tens of microseconds, programs in hundreds of microseconds, and erases a whole block in milliseconds. A hard disk drive must move a head and wait for the platter, which costs several milliseconds. Between DRAM and flash lies a factor of roughly one thousand; between DRAM and disk, a factor closer to one hundred thousand.

That discontinuity shapes software as much as hardware. Operating systems, databases, and file systems all exist in part to stage data across the gap: page caches, buffer pools, write-ahead logs, and checkpointing routines are machinery for hiding the cost of a slow, block-oriented persistent tier. The overhead is not merely latency. Every transition between tiers costs a system call or a driver round trip, a copy, and a translation between the byte-oriented view software prefers and the block-oriented view storage devices present.

Storage-class memory, also called persistent memory, narrows the gap by combining near-DRAM speed with non-volatility, so data survives a power loss without an explicit save. Because it is byte-addressable rather than block-oriented, software reaches it through ordinary load and store instructions instead of a block driver. Intel Optane persistent memory modules, the most widely deployed example, delivered random read latencies of roughly three hundred nanoseconds, several times slower than DRAM but two orders of magnitude faster than NAND. That combination enables architectures that were previously impractical, from databases that recover state almost immediately after a power failure to in-memory systems that keep their entire working set durable.

Emerging Non-Volatile Memory Devices

The candidate media for storage-class memory share a common strategy. Each stores a bit in a physical state of the material itself rather than as trapped charge, which removes the wear mechanism and the high programming voltages that limit flash. Each also lends itself to a crossbar array, in which cells sit at the intersections of perpendicular word and bit lines, paired with a selector device that suppresses sneak currents through unselected cells. The four families below differ mainly in how they encode the bit, and therefore in their speed, endurance, and density.

Phase-Change Memory

Phase-change memory stores a bit in the crystalline or amorphous state of a chalcogenide alloy, typically a germanium-antimony-tellurium composition. The two phases differ in electrical resistance by orders of magnitude, and a current pulse switches between them by melting and quenching or by annealing the cell. Phase-change memory is the most commercially mature of the four for high-density use: it supplied the memory element in 3D XPoint, and STMicroelectronics ships embedded phase-change memory in FD-SOI microcontrollers, including 18-nanometer STM32 parts introduced in 2024. Its principal weaknesses are the energy required to melt the material and resistance drift in the amorphous state, which complicates multi-level cell operation. The underlying materials science, including chalcogenide alloy design and ovonic threshold switching, is treated in phase-change materials, and the device and array detail in persistent memory technologies.

Resistive RAM

Resistive RAM, or ReRAM, forms and ruptures a conductive filament through a thin metal-oxide layer such as hafnium oxide or tantalum oxide, switching the cell between high- and low-resistance states. The structure is simple, the switching energy is low, and the materials integrate readily into standard CMOS back-end processing, which makes ReRAM attractive as an embedded replacement for NOR flash at nodes where flash no longer scales. GlobalFoundries and TSMC both offer embedded ReRAM at 22 nanometers, and Weebit Nano has worked with foundry partners to qualify the technology for embedded system-on-chip use. Cycle-to-cycle and device-to-device variability in filament formation remains the main obstacle to high-density standalone products.

Magnetic RAM

Magnetic RAM stores a bit in the relative orientation of two ferromagnetic layers separated by a thin tunnel barrier. The resulting magnetic tunnel junction presents different resistances when the layers are parallel or antiparallel. Spin-transfer-torque MRAM, the dominant commercial variant, switches the free layer by passing a spin-polarized current through the junction. MRAM offers the highest write endurance of the emerging memories and needs no erase step, but its resistance contrast is modest, which constrains read margin and array size. It is the emerging memory furthest along in foundry deployment: TSMC and GlobalFoundries both produce embedded MRAM at 22 nanometers, and Everspin ships discrete toggle and spin-transfer-torque parts. Spin-orbit-torque devices, which separate the read and write paths, are an active research direction aimed at faster and more reliable switching.

Ferroelectric RAM

Ferroelectric RAM stores a bit in the polarization direction of a ferroelectric capacitor. It writes at very low energy and tolerates enormous cycle counts, which has long made it a favorite for metering, automotive event recorders, and battery-free RFID. Conventional one-transistor, one-capacitor cells suffer a destructive read that requires a write-back, and the perovskite materials used in early parts scaled poorly. The discovery of ferroelectricity in doped hafnium oxide changed that outlook, because the material is already familiar to CMOS gate-stack processing, and it has revived interest in both scaled FeRAM arrays and ferroelectric field-effect transistors that read non-destructively.

No single family wins on every axis. MRAM and FeRAM lead on endurance, phase-change memory and ReRAM on density and cost per bit, and all four remain more expensive per bit than DRAM at comparable capacity. That is why most of these technologies have found their first durable commercial home embedded on logic dies, where they displace flash, rather than as standalone main-memory replacements. For the broader device landscape, see non-volatile memory.

From Proprietary Media to Open Interfaces

Early storage-class memory was defined largely by a single product family. Intel Optane, built on the 3D XPoint memory that Intel and Micron developed jointly, shipped both as persistent-memory DIMMs in DDR4 slots and as high-endurance NVMe solid-state drives. The modules ran in two modes: Memory Mode, which presented a large, cheaper memory tier with DRAM acting as a cache and no persistence exposed to software, and App Direct Mode, which exposed byte-addressable persistence directly to applications.

The business did not survive its economics. Micron ended 3D XPoint development in 2021 and sold its Lehi, Utah, fabrication plant. Intel announced the wind-down of the Optane business in 2022, taking a charge of roughly $559 million against Optane inventory, and set the end of 2024 as the last date for orders of the Persistent Memory 200 series, with shipments concluding through 2025. A proprietary medium had to carry the full cost of its own fab, its own DIMM ecosystem, and its own software enablement, against a DRAM industry with vastly greater volume. Optane proved that a practical tier between DRAM and flash was achievable; it did not prove that one company could sustain it alone.

The industry's center of gravity has since shifted to CXL, an open interconnect that runs on the PCI Express physical layer and adds cache-coherent memory semantics. Rather than betting on one novel medium, CXL treats memory as a resource that can be attached, expanded, pooled, and shared across a fabric, so a server can draw on shared capacity instead of stranding DRAM inside individual machines. The protocol comprises three sub-protocols: CXL.io for discovery, configuration, and bulk transfer; CXL.cache, which lets a device cache host memory coherently; and CXL.mem, which lets the host address memory attached to a device. A memory expander that implements CXL.io and CXL.mem is the simplest and most widely deployed form.

The specification has advanced steadily. CXL 1.0 and 1.1 arrived in 2019 with single-host device attachment. CXL 2.0, released in 2020, added switching, memory pooling across hosts, and link-level encryption. CXL 3.0, released in 2022, doubled the signaling rate to 64 GT/s on the PCI Express 6.0 physical layer and introduced true cross-host memory sharing, multi-level switching, and peer-to-peer transfers. CXL 3.1 followed in 2023 with fabric-manager interfaces, port-based routing enhancements, and the Trusted Security Protocol for confidential computing; CXL 3.2, released in December 2024, refined memory-device monitoring, management, and security. CXL 4.0, released in November 2025, doubles the rate again to 128 GT/s over the PCI Express 7.0 physical layer, adds bundled ports that present several physical links as one logical connection, supports a native two-lane width for greater fan-out, permits up to four retimers for longer reach, and improves memory error reporting and sparing. Every release remains backward compatible with its predecessors.

The trade-off is latency. Memory reached over CXL is coherent but not local; access costs roughly what a remote NUMA node costs, not what a directly attached DIMM costs. CXL therefore works best as a capacity and bandwidth tier beneath local DRAM rather than as a replacement for it, which places the burden on hardware and system software to route each page to the tier that suits its access pattern.

Archival Storage at Extreme Density

At the cold end of the hierarchy the design problem inverts. Latency scarcely matters; density, durability, and cost per bit over decades dominate. Magnetic tape still sets the cost floor for cold archives, but tape cartridges must be migrated to new formats every several years as drives and media reach end of life, and the world generates archival data faster than areal density improves.

DNA data storage answers with volumetric density no magnetic or optical medium approaches. Data are encoded into nucleotide sequences, written by chemical synthesis, and read by sequencing, with random access supplied by polymerase chain reaction primers that amplify a selected pool. Theoretical densities reach hundreds of petabytes per gram, and DNA kept cold, dry, and dark remains readable for centuries. The obstacles are write cost and write latency: synthesis is slow and expensive compared with any electronic medium, and reads take hours rather than microseconds. That profile suits deep archives that are written once and read rarely, not working storage.

Holographic storage records data through the volume of a photosensitive medium rather than on its surface, interfering a data-bearing beam with a reference beam to write an interference pattern and reconstructing a whole page of bits at once on readout. Multiplexing schemes that vary the reference beam angle, the medium position, or its rotation superimpose many pages in the same volume, and page-parallel readout offers high transfer rates. Commercial ventures in the 2000s failed to reach volume production, defeated by media shrinkage during recording, photopolymer stability, and the cost of precision optomechanics, but research continues on the promise of high-capacity, long-lived optical archives. The underlying optics, including recording materials, spatial light modulators, and multiplexing geometries, are covered in holographic recording and display.

Impact on Computing

Persistent and disaggregated memory force a rewrite of assumptions that decades of software took for granted. Once memory survives power loss, correctness depends on knowing precisely when a store has reached durable media, and processor caches sit in the way. On x86 platforms, applications use cache-line writeback and optimized flush instructions together with store fences to order their writes, and platforms that implement extended asynchronous DRAM refresh extend the power-fail-protected domain to include the CPU caches, removing the need for explicit flushes. The SNIA NVM Programming Model standardizes the resulting abstractions, and direct-access file system modes let applications map persistent memory and address it without a page cache in the path.

The failure modes are new as well. A partially completed update that survives a crash is worse than one that does not, so persistent data structures need failure-atomic update protocols, and allocators must avoid leaking memory that outlives the process that allocated it. Databases can shorten or eliminate recovery replay, but only if their durability protocols are rebuilt around byte-granular persistence rather than block writes.

Hardware evolves in step. Memory controllers must accommodate devices whose latency, bandwidth, endurance, and persistence differ from DRAM while preserving compatibility with existing software, and they increasingly carry wear-leveling and error-management logic once confined to storage controllers. Operating systems have grown tiered-memory subsystems that track page access frequency and promote or demote pages between local DRAM and slower attached memory. In parallel, high-bandwidth memory stacks DRAM dies over a wide interface to feed accelerators, and memory-centric computing pushes computation into or beside the memory array to avoid moving data at all. The common thread is that memory has become an architectural design space rather than a fixed commodity beneath the processor.

Challenges and Trade-offs

Cost per bit remains the decisive constraint. DRAM benefits from decades of manufacturing scale, and any challenger must beat it on capacity per dollar, not merely on a technical figure of merit; Optane failed on precisely this axis. Emerging memories also carry endurance and retention limits that vary with temperature and cycling, so controllers must manage wear rather than assume it away.

Persistence introduces security and privacy obligations that volatile memory did not. Data that survives power loss survives decommissioning, theft, and resale, so encryption at rest, verified erasure, and key management become requirements rather than options. Pooled and shared memory adds isolation problems, since several hosts drawing on one memory device need strong guarantees that they cannot read one another's data; the Trusted Security Protocol in recent CXL revisions exists to address exactly that.

Finally, the software ecosystem lags the hardware. Applications that gain the most from byte-addressable persistence are the ones that must be restructured most deeply, and the retreat of Optane left developers wary of investing in programming models tied to a specific medium. Interface-level standardization through CXL reduces that risk by decoupling the software contract from the underlying device.

Outlook

The most likely near-term outcome is not a single technology that replaces DRAM, but a hierarchy with more rungs and better software control over them. Emerging non-volatile memories are consolidating around embedded roles on logic dies, where they replace flash at nodes that flash cannot reach. CXL is establishing the interface through which additional tiers, whatever medium they use, attach to the host. Archival research on DNA and holographic media targets a different problem entirely, namely storing exabytes for decades at a cost that tape migration cannot sustain. Progress in this field will be measured less by any single device breakthrough than by how well systems learn to place each byte on the tier that fits it.