Persistent memory is a type of computer memory that keeps its data even when the power goes off, while still operating at speeds close to traditional RAM. It sits in a unique position between the two main tiers of a computer’s data storage: DRAM (the fast, volatile working memory that loses everything during a shutdown) and SSDs or hard drives (the slow, non-volatile storage that retains data permanently). Persistent memory combines the best traits of both, delivering read/write speeds roughly 25 times slower than DRAM but an order of magnitude faster than the fastest solid-state drives.
How Persistent Memory Differs From RAM and SSDs
Standard DRAM has a read/write latency of about 14 nanoseconds, which makes it ideal for active computations. The tradeoff is that it’s volatile: cut the power and every bit of data vanishes. SSDs retain data without power but operate in the microsecond range, thousands of times slower than DRAM. Persistent memory lands in between, with latency around 350 nanoseconds. That’s slow compared to DRAM, but measured against SSDs, it’s extraordinarily fast. Microsoft’s testing of persistent memory on Windows Server showed file system latency consistently under 40 microseconds at over 13.7 million input/output operations per second.
This middle ground matters because many workloads, particularly databases and transaction processing systems, need both speed and data safety. Traditionally, software had to write critical data out to an SSD or disk to protect against power loss, which created a bottleneck. Persistent memory lets applications keep large datasets in something that behaves like RAM but survives a reboot.
The Hardware Behind It
The most prominent persistent memory hardware was Intel’s Optane Persistent Memory, which used a technology called 3D XPoint (pronounced “cross point”). Unlike traditional flash storage, which stores data by trapping electrical charges, 3D XPoint stored bits by changing the electrical resistance of special materials. The storage cells used a chalcogenide compound (germanium-antimony-tellurium) accessed through tiny threshold switches, all arranged in a stackable cross-grid architecture. This design allowed the memory to be addressed at the individual bit level rather than in large blocks, which is one reason it could operate so much faster than flash.
What made Optane Persistent Memory physically distinctive is that the modules plugged into standard DIMM slots on server motherboards, the same slots normally reserved for RAM. These modules, called Data Center Persistent Memory Modules (DCPMMs), came in capacities far larger than typical DRAM sticks. A single server could be outfitted with terabytes of persistent memory, something that would be prohibitively expensive with DRAM alone.
NVDIMM Types
Persistent memory modules that fit into DIMM slots fall under the umbrella term NVDIMM (non-volatile DIMM), and there are several approaches to making them work. The first type, called NVDIMM-N, pairs standard DRAM with a backup NAND flash chip and a supercapacitor on the same module. During normal operation, you get full DRAM speed. When power fails, the supercapacitor provides enough energy for a small controller to copy all the DRAM contents into the flash chip. When power returns, the data moves back into DRAM automatically.
A second type, NVDIMM-F, is essentially a solid-state drive redesigned to communicate over the memory bus instead of a storage interface. It’s persistent by nature but doesn’t match DRAM speeds. Intel’s Optane DCPMMs represented a third approach: a fundamentally different memory material that’s inherently non-volatile and fast enough to sit on the memory bus without needing a DRAM-plus-flash workaround.
Two Ways to Use It
Persistent memory modules can operate in two distinct modes, and the choice between them determines whether the data actually persists.
In Memory Mode, the persistent memory acts as a large pool of cheaper, higher-capacity RAM, with the server’s regular DRAM serving as a fast cache in front of it. This lets you dramatically expand a system’s total memory capacity at lower cost. The catch: Memory Mode makes the persistent memory volatile. If power fails, everything stored in it is lost, just like regular RAM. The application never talks to the persistent memory directly; it only sees the DRAM cache layer.
In App Direct Mode, the persistent memory functions as extremely fast non-volatile storage. Applications can read and write to it with awareness that the data will survive a power failure. This is the mode that unlocks the core promise of the technology. Software can access the persistent memory in several ways, each with different speed tradeoffs. Direct memory access, where the application maps the persistent memory into its own address space and reads or writes with simple load and store instructions, is the fastest. File system access, where the application uses standard file operations, is slower but requires fewer code changes. A middle option, raw device access, talks directly to the driver and skips the file system overhead.
Intel published a Persistent Memory Development Kit with libraries that handle the complexities of working with this new memory tier, so developers didn’t have to build everything from scratch. One popular pattern is persistent in-memory key-value stores, where an application can look up individual keys and values directly from persistent memory without reading entire storage blocks or allocating temporary RAM buffers.
Real-World Performance Gains
The most compelling use case for persistent memory has been in-memory databases. SAP HANA, one of the largest enterprise database platforms, showed that running on persistent memory delivered 99.7% of the performance of a system using DRAM alone when handling identical multi-million-record processing workloads. The real wins came in capacity and recovery time. Persistent memory allowed memory density to increase in an 8:1 ratio compared to DRAM, expanding a server’s usable memory from 1.5 terabytes to 4.5 terabytes per processor, with at least 20% cost savings over equivalent DRAM configurations.
Startup time told an even more dramatic story. A 6-terabyte SAP HANA database that previously took 50 minutes to load into DRAM from storage after a restart could be ready in 4 minutes with persistent memory, because the data was already sitting in a memory-speed tier that didn’t need reloading. For businesses running mission-critical databases, cutting recovery time from nearly an hour to a few minutes is a significant operational improvement.
Current Status and What Comes Next
Intel discontinued its entire Optane Persistent Memory product line. The 100 and 200 Series are listed as discontinued products, and in January 2023, Intel canceled the 300 Series before it ever shipped, announcing it would not develop future Optane products. Existing 100 and 200 Series modules continue to receive warranty support, but no new orders or production runs are planned.
The technology concept hasn’t disappeared, though. Intel has signaled that Compute Express Link (CXL) will be the future standard for tiered-memory solutions. CXL is an open interconnect standard that allows different types of memory, including non-volatile memory, to be attached to processors through a high-speed link. Rather than requiring special DIMM-slot modules from a single vendor, CXL opens the door for multiple manufacturers to build memory devices that a server can use as an additional memory tier. The practical effect for the industry is a shift from one proprietary persistent memory product to a broader ecosystem where various memory technologies can plug into a common interface and serve similar roles.
For organizations that adopted Optane Persistent Memory, the transition means evaluating CXL-based alternatives as they mature. For everyone else, the core idea remains relevant: the traditional boundary between “fast but volatile” and “slow but persistent” is dissolving, and future server architectures will increasingly treat memory as a spectrum of tiers rather than two rigid categories.