3D XPOINT: INTEL’S SECRET WEAPON
Jeremy Laird asks whether the tech behind Optane is just a supercharged flash alternative, or the gateway to 100GHz optical computing?
Is the tech behind Optane is just a supercharged flash alternative, or the gateway to 100GHz optical computing?
Is Intel’s new 3D XPoint tech the biggest paradigm shift in computing since the introduction of silicon chips? Is it merely the basis of faster SSDs? Or is this story nothing more than a combination of vaporware and conspiracy theories? 3D XPoint is one of the most mysterious technologies to come out of Intel in years. The first products derived from it have finally started to appear, yet many don’t fully understand what it is or how it works. In fact, to all but a few, it’s probably not even clear what it’s intended to do.
That’s partly down to a lack of clarity on Intel’s (and its partner, Micron’s) behalf. The messaging has been mixed, both in terms of the intended applications for 3D XPoint, and the insight Intel has provided into its technical underpinnings. It’s the latter, in particular, that has fueled speculation. Nature abhors a vacuum, and into the space created by equivocation has flowed extensive conjecture.
Is 3D XPoint the basis for a whole new approach to chips, with chalcogenide and phase-change materials replacing silicon and the transistor? Does it open the door to fully integrated on-chip optical computing? Will it smash the circa-4GHz ceiling on CPU speeds, achieving frequencies in hundreds of gigahertz?
Meanwhile, rumours abound of corporate manoeuvering and even skulduggery. Is the patchy technical insight an attempt to protect valuable and vulnerable intellectual property? Could Intel be playing down the true value of 3D XPoint, while it scoops up the relevant patents, ensuring a lucrative lockdown on the next age of computing? Or could there be an intentional cover-up involving infringement of third-party patents?
Heady stuff. Getting to the truth, however, isn’t easy. All we know is that the first 3D XPoint products will take the form of non-volatile solid-state storage, sold under the new Intel Optane brand. But even here, there’s scope for confusion. The first Optane drives aren’t intended to replace system memory or hard drives. Instead, they’re somewhere in between — a kind of fast cache designed to act as a buffer in front of the relatively slow performance of an SSD containing NAND flash memory. With that, let us begin...
When 3D XPoint was first mooted, it was exciting simply as a new memory technology, let alone as the foundation for a whole new approach to computer chips. The industry has spent decades trying to cook up a new non-volatile storage technology with which to improve upon flash memory. Finally, here it is.
The sales pitch was stellar. In early presentations, Intel said 3D XPoint was up to 1,000 times faster and 1,000 times more durable than existing NAND flash memory. Even by the standards of the chip industry, for which exponential improvements over time are the norm, that kind of one-off step change is exotic.
The commercial reality isn’t proving quite so spectacular. The first retail products based on 3D XPoint are just about to hit the market under Intel’s Optane branding, and expectations have been scaled back dramatically. Specific performance data hasn’t been released. But Intel is talking about Optane drives being up to 10 times faster than a conventional SSD. Still a big jump, but perhaps rather more plausible.
Let’s not get ahead of ourselves. First, let’s lay out what we know about 3D XPoint. Currently, there are two prevailing memory technologies, if we’re talking integrated circuits rather than magnetic memory solutions, such as a hard disk. RAM or system memory is super fast and effectively lasts forever, but requires a constant electrical current. NAND or flash memory is non-volatile, so maintains data integrity without a power supply, but at the expense of speed and durability.
There are a few other details that separate the two, such as memory cell size and power consumption, both of which are relatively poor with RAM. But the holy grail for the last 20 to 30 years has been to develop a memory technology that combines the best of both: speed and durability with non-volatility. Several technologies and even whole companies promising just such a solution have come and gone over the past few decades. But 3D XPoint is arguably the first that really looks like it’s going to stick.
It’s actually the work — or at least the property — of two of the biggest players in computer chips in general: Intel and Micron.
The former is best known today as a maker of PC processors, but Intel’s genesis was actually as a maker of memory. Its first product was in fact DRAM, or dynamic random access memory, which remains the basis of system memory in PCs and other computing devices today. Micron, meanwhile, is well known as a maker of all kinds of memory, from sticks of DDR to solid-state drives, many of which sell under its Crucial retail branding.
Although 3D XPoint may eventually end up replacing both DRAM and NAND, for now, it’s the latter that begs the most obvious comparison, not least because it, too, is a non-volatile memory type. But there are huge differences thereafter. For instance, NAND stores data by trapping electrons in an insulated floating gate. Thus, each memory cell is a sort of capacitor, with the charged or discharged state representing a bit of data.
3D XPoint, by contrast, is a resistance-based technology that uses bulk property change to the cell material to alter its resistance level, and so record a memory state. When you drill right down to cell level, 3D XPoint also has a remarkably simple structure. There’s the memory cell itself, a switch or selector element, then the connecting wires. And that’s really it. The cleverness comes in the materials and how it’s all arranged.
Currently, Intel and Micron haven’t revealed all the details, leaving space for speculation as to the full implications of the technology. But what is clear is that the bitline and wordline connecting wires are arranged above and below the cells, and laid perpendicular to one another, in a grid pattern, hence the ‘XPoint’ or crosspoint moniker.
There are some immediate benefits that come out of all that. For starters, the simplicity of the structure, including the notable lack of a transistor, makes for excellent cell density. In other words, you can pack in more data, especially compared to DRAM. What’s more, the ‘crosspoint’ wiring means the data is addressable at individual bit level. This has all kinds of upsides compared to NAND flash, which must be read and written in large blocks.
The latter is a reasonable approach when reading or writing large chunks of data, but it quickly becomes inefficient with random access workloads involving high-frequency requests to read and write little snippets of data. The net result is that you end up reading and writing large amounts of redundant data. By enabling bit-level access to data, 3D XPoint has NAND flash dramatically beaten for speed and latency.
As we mentioned earlier, 3D XPoint is also claimed to offer up to 1,000 times better endurance than NAND. That makes for something in the order of a few million write-erase cycles, or hundreds of petabytes of writes overall, and hundreds of terabytes of writes per day. So, for most users, an SSD using 3D XPoint technology could be considered to last forever. It will never wear out.
Overall, the broad picture for the first implementation of 3D XPoint in Optane drives puts its capabilities roughly between DRAM and NAND, but arguably closer to the former. It’s far faster than NAND, but not as fast as DRAM. Likewise, it’s much denser than DRAM in terms of data storage, but not quite as dense as NAND.
If that’s what we know for sure today, how will 3D XPoint deliver in the future, and could it really be the beginning of a new age of computing? In the short to medium term, it seems fairly certain that Optane drives will transition from cache drives into general storage. The first Optane products have been announced at 16GB and 32GB capacities, using a production node roughly equivalent to 20nm transistors (without any transistors, direct comparisons of feature size become quite subjective), reflecting the relatively high cost of the very first 3D XPoint chips. Prices haven’t been announced, but as Intel and Micron get to grips with production, we expect to see the cost come down fairly rapidly. Within a couple of years at most, Optane drives measuring hundreds of gigabytes and more should be available at prices normal consumers can afford.
It’s everything beyond that where things get complicated. For instance,
co-CEO Guy Blalock of the IMFT Intel-Micron joint venture responsible for 3D XPoint has said, “3D XPoint specifically offers a path to provide massive amounts of memory storage at a latency rate sufficient to place it in close proximity to the CPU.” Intriguing, but what does it actually mean? It’s not exactly saying that 3D XPoint is poised to replace DRAM. But it’s also placing it in an unambiguously different position in the system hierarchy compared to existing solid-state drives. Overall, it points to a new class of memory that’s difficult to currently define.
Even more significantly, however, could 3D XPoint be the basis of a new age of post-silicon computing? At this point, a warning is required. Much of what follows is highly speculative, though the scope for speculation is at least in part due to Intel and Micron’s failure to provide clarity on some of the technical details.
Much of the speculation revolves around the materials and operational specifics of 3D XPoint. Some earlier speculation implied it depended on phase-change physics to store data, an approach that has been under evaluation for non-volatile memory for some time. Phase-change memory is built around a material known as chalcogenide glass, which sports some pretty funky properties. In simple terms, chalcogenide can be switched between amorphous and crystalline solid states or phases, courtesy of applying an electrical current. What’s more, a number of distinct intermediate states can also be achieved. Amorphous versus crystalline gives you 0 and 1 for the storage of a single bit, obviously. The intermediate states allow for multiple bits of memory per cell.
Multiple companies, including IBM, Samsung, Numonyx and Intel, have been working on phase-change memory, or PRAM, for some time. However, Intel and Micron have repeatedly denied that 3D XPoint depends upon phase-change physics. For instance, when UK tech site The Register put the question to Micron regarding 3D XPoint’s use of phase-change materials, it was told, “3D XPoint technology is a new class of non-volatile memory invented by Intel and Micron that relies on resistance change of the bulk material to achieve non-volatility. Unlike phase-change memory, 3D XPoint technology uses a unique cross point architecture, enabling it to scale in ways that phase-change memory has not been able to accomplish.” That’s a pretty unambiguous denial that 3D XPoint is a phase-change memory technology. So why are we even discussing it?
How about this quote in reference to 3D XPoint, again from Guy Blalock of the IMFT? “Chalcogenide material and an Ovonyx switch are magic parts of this technology, with the original work starting back in the 1960s.” So 3D XPoint is a chalcogenide device and builds on its development of that material in a computing context since the 1960s, which has been primarily centred on its phase-change properties. All very confusing.
This is where the conspiracy theories come in. Are Intel and Micron obscuring 3D XPoint’s phase-change properties with a view to sidestepping patents they don’t own? Could it be a double bluff, in which they imply that it’s phase-change by indicating that the core tech is chalcogenide, because they’re either in the process of acquiring patents for something else, or know it can’t be protected and want to give themselves as big a head start over the competition as possible? Who knows?