Computers traditionally treated working memory and permanent storage as two very different resources. DRAM was fast enough to keep information close to the processor, but its contents disappeared when electrical power was removed. Storage retained information without power, but accessing it took considerably longer.
That separation influenced almost everything about computer architecture. Programs and files lived on persistent storage and had to be moved into system memory before the processor could work with them efficiently.
A different class of non-volatile memory was beginning to challenge that familiar arrangement. 3D XPoint was designed to retain information without power while providing substantially lower latency than NAND flash.
The Gap Between DRAM and NAND Was Enormous
DRAM and NAND flash were both semiconductor memory technologies, but they served very different jobs. DRAM provided the rapidly accessible working space used by running software. NAND provided persistent capacity for solid-state drives and other storage devices.
The difference was not merely whether information survived a shutdown. The technologies behaved differently when reading, writing, erasing, and addressing information.
Very low latency made it suitable for active system memory, but stored information depended on continuous power and periodic refreshing.
Information remained after power was removed, making NAND practical for persistent storage, but access latency was much greater than system memory.
Computer systems compensated for the difference by creating a hierarchy. Frequently needed information was kept close to the processor in caches and DRAM, while much larger quantities could remain on slower persistent storage.
A New Memory Technology Targeted the Space Between Them
3D XPoint was developed jointly by Intel and Micron as a new form of non-volatile memory. Its intended position was unusual because it did not simply attempt to become another version of conventional NAND flash.
The technology promised much lower latency and substantially greater endurance than NAND while retaining information when power disappeared. At the same time, it was intended to provide considerably greater density than conventional DRAM.
The important idea was not that storage had suddenly become RAM. A new layer was appearing between two technologies that had previously been separated by a very large performance gap.
That middle position created possibilities that were difficult to achieve economically with DRAM and difficult to achieve quickly with NAND.
The Cells Were Arranged at Crossing Points
The name 3D XPoint described part of the physical organization of the technology. Memory cells were positioned where perpendicular conductors crossed one another.
By selecting the appropriate word line and bit line, an individual cell could be addressed. The structure did not require a transistor at every memory cell in the way some other memory arrangements did.
An arrangement in which memory cells are located at intersections between perpendicular conductors. Selecting the corresponding lines provides access to a particular cell.
The arrays could also be stacked. The initial implementation used two memory layers, allowing greater density without depending entirely on shrinking features horizontally across the silicon.
Persistent Data Did Not Have to Mean NAND Flash
For ordinary computer users, solid-state storage had become closely associated with NAND. SSDs, USB flash drives, and memory cards all helped make flash storage familiar.
Non-volatile memory, however, is a broader concept. It simply describes memory capable of retaining information when power is removed. NAND is one way to accomplish that, but it is not the definition itself.
3D XPoint demonstrated why that distinction mattered. It was non-volatile, yet its internal organization and operating characteristics were different from the NAND flash used in conventional SSDs.
A technology can preserve information without electrical power while using a storage mechanism and access architecture different from NAND flash.
Small Accesses Could Become Much More Important
NAND flash has restrictions that influence how SSD controllers manage information. Data can be read and programmed in pages, while erasing occurs across much larger blocks. Updating small pieces of information therefore involves more internal management than simply changing one isolated byte in place.
SSDs hide much of this complexity through controller logic, mapping tables, garbage collection, caching, and other techniques. Those mechanisms make NAND extremely useful, but they also reflect the characteristics of the underlying medium.
3D XPoint was designed around individually addressable cells and fast state changes. That offered a different path for workloads dominated by small, latency-sensitive accesses.
Keeping Data Closer to the Processor Could Change the Waiting Time
Processor performance can be wasted whenever execution has to wait for information to arrive. Modern computers use several layers of increasingly large but generally slower storage to reduce that problem.
Processor caches occupy the fastest end of the hierarchy. DRAM provides much greater working capacity. Persistent storage supplies still greater capacity but introduces another major increase in access time.
A persistent medium with latency significantly below NAND could occupy a new position in that chain. Information that was too large or too persistent for conventional memory could potentially remain closer to the performance characteristics expected by the processor.
Processor cache
Very small quantities of frequently needed information remain extremely close to the processor cores.
System memory
DRAM provides the larger working area used by the operating system and active applications.
Emerging persistent memory
A technology such as 3D XPoint could occupy territory between volatile memory and traditional solid-state storage.
NAND storage
SSDs provide much larger persistent capacity at a lower cost per unit of storage, but with greater access latency.
Changing the Medium Did Not Eliminate the Interface
A fast memory technology alone could not determine how quickly software received its data. The controller, protocol, operating system, driver stack, and physical interface could all add latency between an application and the underlying medium.
This distinction was particularly important when a new memory technology was packaged as a storage device. Even if the medium itself could respond extremely quickly, accessing it through a conventional storage path could prevent software from seeing its full native behavior.
The response time of the physical memory is only one part of the complete path. Controllers, interfaces, protocols, drivers, and software can add additional delay before data reaches the processor.
That was one reason the boundary between memory and storage was becoming so interesting. The hardware could improve enough that older software and interface assumptions themselves became noticeable parts of the delay.
Persistence Could Change More Than Startup Speed
It was tempting to think of faster non-volatile memory simply as a way to make an SSD quicker. The larger possibility involved changing where information needed to live and how frequently it needed to be copied between layers of the system.
Traditional software assumes that important persistent information belongs on storage while active working information belongs in volatile memory. Applications load information from one into the other and later write persistent changes back.
If a sufficiently fast persistent medium could sit much closer to the processor, some of those assumptions could eventually be reconsidered. Large data sets might remain available without requiring the same movement between slow persistent storage and limited volatile memory.
The significance of a faster persistent medium extended beyond transfer rates. It raised the possibility of changing the traditional distance between information that was stored and information that was ready to be processed.
The Old Boundary Was Starting to Become Less Distinct
DRAM still had advantages that made it essential as system memory, while NAND remained valuable for providing enormous persistent capacity at practical cost. A new technology did not instantly replace either one.
What changed was the assumption that every useful memory technology had to fit neatly into one of those two roles. 3D XPoint demonstrated that persistent storage could be designed with characteristics much closer to memory, creating another level in the hierarchy rather than merely replacing an existing level.
That made the technology important even before its eventual role in commercial computers was fully established. The familiar line between memory and storage was no longer defined only by whether information disappeared when the power went off.
