Intel Optane will use new 3D XPoint memory
Thu 20 Aug 2015
In the opening keynote speech of the recent Intel Developer Forum, Intel Optane was announced, combining Intel and Micron’s new 3D XPoint memory “with the company’s advanced system memory controller, interface hardware and software IP”.
Starting from 2016, Optane SSDs (Solid State Drives) will span a wide range of products, from pocketbooks to data centres. Optane will also equip a new line of DIMMs (Dual In-line Memory Modules) for Intel’s next-gen data centres. PCWorld reports that, “The DIMMs will work with future Intel Xeon server processors.”
Intel demoed an Optane SSD (running 3D XPoint technology) versus one of their fastest current models, the DC P3700 (running NAND technology). The Optane peformed approximately 5 to 7 times as many IOPS (Input/Output Operations Per Second) than the P3700, “according to the queue depth of the workload under test”, TechReport says.
While Intel claim that 3D XPoint can theoretically be up to 1000 times faster than NAND flash (based on latency), real-world operations are a different matter, with many different factors affecting functional speed. The number of IOPS is generally more relevant to large-scale uses (such as data centres) than the average consumer, who would likely demand less of their machine. As DigitalTrends points out, “Random access times and sustained read/write bandwidth are more important to most PCs.”
Also, this is an early prototype, so functional speed would be expected to improve as Intel develop Optane further.
Based on Intel’s ‘1000 times’ estimate [PDF], “In the time it takes an HDD to sprint the length of a basketball court, NAND could finish a marathon, and 3D XPoint technology could nearly circle the globe.”
In any case, Optane is really, really fast, whether or not it yet qualifies as ludicrous speed. It’s faster than Flash; will it one day be faster than The Flash?
One implication of this increased speed is in the video games industry, where it could eliminate the need for load screens, as well as wait times when players move from one environment to another.
PCWorld adds, “It’s also being positioned for running huge in-memory databases, so companies can do real-time analytics on much larger sets of data.”
Onto the 3D XPoint technology itself, Intel classes it as the first new category of memory in over 25 years. The nine categories so far are: RAM (Random Access Memory) in 1947, PROM (Programmable Read-Only Memory) in 1956, SRAM (Static Random Access Memory) in 1961, DRAM (Dynamic Random Access Memory) in 1966, EPROM (Eraseable Programmable Read-Only Memory) in 1971, Bubble Memory in 1979, NOR (from the logical NOT OR) Flash Memory in 1984, NAND (from the logical NOT AND) Flash Memory in 1989, and now 3D XPoint Memory in 2015.
There’s a film industry maxim which suggests you can either get a film made quickly and cheaply but not well, or quickly and well but not cheaply, or well and cheaply but not quickly, but not all three. A similar thing could previously have been said of computer memory: DRAM is fast but expensive and volatile (loses its information when the power is turned off), and while NAND is inexpensive and non-volatile, it’s slow compared to DRAM (1500 times), though fast compared to hard drives (up to 1300 times).
Columns and selectors in 3D XPoint
Also, as opposed to many other components that are made of silicon, DRAM’s density is not currently doubling every two years.
In contrast, 3D XPoint is the sought-after combination of fast, non-volatile, and inexpensive. It’s made by chopping tiny layers of materials into individual columns of one selector above one memory cell.
The columns are connected by “an innovative cross-point structure consisting of individual wires”. This means that the system can address individual memory cells by selecting two wires perpendicular to each other, approximately similar to a ‘latitude and longitude’ system (or ‘rows and columns’).
Density is achieved by stacking such memory grids in 3D, and by eliminating transistors, making 3D XPoint up to 10 times denser than DRAM.
Whereas DRAM needs transistors which take up space, one for each memory cell, in order to access and modify cells, 3D XPoint gets around this by varying the voltage sent through the ‘latitude and longitude’ selecting wires.
Intel calls their design, “Simple. Stackable. Transistor-less.”
Compared to NAND’s latency, measured in 10s of micro-seconds, 3D XPoint’s latency is measured in 10s of nano-seconds, making the latter a much higher performance type of memory.
3D XPoint also has a higher level of endurance, with the number of write cycles it goes through not significantly affecting its lifespan.
This “means it can simultaneously serve as a single, high-speed, high-capacity pool of system and storage memory.”