Has Intel Invented a Universal Memory Tech? – IEEE Spectrum
Today’s computers shuttle data around a byzantine system of several different kinds of short- and long-term memory. No wonder, then, that engineers have long dreamed of one memory technology to rule them all, a universal memory that would simplify computing and streamline the path of data.
In March, Intel announced that it will sell to data centers a new kind of solid-state drive, called Optane, that it says could lead to this kind of simplification. Optane drives are nonvolatile, like flash memory, which means that they should use relatively little standby power and that they’re fast, like DRAM. “It really starts to marry the worlds of memory and storage together,” company CEO Brian Krzanich says in a promotional video, over the swells of heroic music. The technology “comes close to being the holy grail of memory,” says Intel executive vice president William Holt in the same video.
Whether 3D XPoint, the mystery technology inside Optane, can live up to this promise is likely to depend on the performance it delivers as well as Intel’s ability to scale up manufacturing using new materials and build out the right market. The 375-gigabyte Optane drive on offer now costs US $1,520, about three times the price of an equivalent solid-state drive.
This first product will enable data centers to do more with a smaller number of servers, says James Myers, who works on nonvolatile memory architecture at Intel. Myers gives an example of servers running a MySQL database, which, among other things, apps use to store instant messages. An equivalent flash drive can perform 1,400 such transactions per second; the Optane drive can perform over 16,000.
The Optane drive was announced with bombast, but the company is coy about the technology behind it. Myers says “3D” refers to the fact that the memory cells are stacked; “XPoint” alludes to the way the memory elements are arranged. While flash memory elements must be read and written in groups, XPoint elements—situated at the crossing point of interconnects—can be addressed individually. Myers says that this architecture, and something inherent to the storage materials themselves, makes 3D XPoint faster than flash memory.
Intel, which initially developed 3D XPoint in conjunction with Micron, won’t say what the technology really is, but this doesn’t seem to bother researchers or analysts. “Everyone seems to think it’s phase-change memory,” says semiconductor analyst Jim Handy. “I don’t care.” What matters to him—and, Intel hopes, its customers—is the performance.
The complexity of today’s memory hierarchy—a combination that often includes magnetic disks and flash for storage and DRAM and static RAM for memory—is a necessary evil. Each technology has its own strengths, so they must be combined. Data are shuttled around from speedy but expensive SRAM caches—which are close to the processor and embedded within it—to slower, less expensive (but still pricey) DRAM. Finally, data are stored in slow but reliable flash or hard-disk drives, or both. Even if it’s not possible to do it all in one memory technology, using only one for working memory close to the processor and one for longer-term storage would help simplify things. Intel says that XPoint memory could provide a speedier alternative to flash memory and magnetic hard disks. The company has also suggested it could supplement or supplant DRAM.
“DRAM is unique in its ability to waste power, so anything you can do to get rid of it is great,” says Handy. For example, Google is thought to store the index of the entire Internet on several power-hungry, quick-access DRAM servers. If the company could switch this over to 3D XPoint—which Intel claims has 10 times the density of DRAM—Google could use fewer servers and thus save power and money, according to Steven Swanson, a computer scientist at the University of California, San Diego.
Intel is providing software that will enable computers to operate Optane as memory as opposed to storage, but it will be slow. Optane drive latencies max out at 7 or 8 microseconds—way faster than flash, which takes hundreds of microseconds, but not touching DRAM’s low hundreds of nanoseconds. The Optane drives are fettered by the interface they use: They connect to the storage interface, not the memory interface.
In the short term, Intel’s drives are not likely to replace any existing memory technologies but will instead supplement them, says Swanson, who built a research drive based on the company’s 3D XPoint technology in 2011. Swanson expects the path to memory and storage simplification to be complex because computing systems will have to be redesigned to route data in new ways.
Swanson and Handy believe Intel started with storage to help smooth out some of the risk in launching a new memory technology. Making Optane a memory requires new circuit board designs and cooperation from programmers. To get those, Intel needs to show that there is a market for 3D XPoint and demonstrate its reliability. Intel says that a product using a memory interface will be out in 2018. Even using the interface, 3DXPoint still won’t be quite as fast as DRAM, but Intel promises that it will be denser and less expensive.
The success of this new memory, then, will hinge on data centers taking it up in a less than ideal initial form while the company works on scaling up production. Even though the flaws in today’s memory and storage hierarchy are universally acknowledged, trying to change it is a risky move. “Almost all memory companies have one or two potential competitors to this technology, and they’re all waiting to see what happens before they jump into [a] big investment,” says Swanson.
Memory enthusiasts disagree about whether a true universal memory is even physically possible. It is perhaps most useful as a goal to guide the computer industry forward. “The concept of the universal memory is attractive because the idea is to simplify,” says Wei Lu, a computer scientist at the University of Michigan and chief scientist at Crossbar, a resistive RAM startup company. “We have a big-data problem, and today’s computers are fundamentally not good at this.”