3D XPoint vs. NAND flash: Why there's room for both
While there are key differences between 3D XPoint and NAND flash -- in terms of performance, cost and uses -- the two technologies will likely coexist for years.
In their comparisons of 3D XPoint vs. NAND flash, storage admins shouldn't feel pressure to commit to one over the other. In fact, both technologies have a place in the enterprise.
Let's explore where 3D XPoint memory fits into the memory/storage hierarchy, how it compares to NAND and potential uses.
3D XPoint vs. NAND: Cost and performance
3D XPoint, the basis for Intel's Optane products, is not a threat to NAND flash. A good understanding of system architecture, however, is required to understand why. XPoint's cost/performance tradeoff in systems dictates its use in the enterprise.
Although a pyramid is commonly used to illustrate the memory/storage hierarchy, that shape doesn't adequately illustrate the reason that XPoint -- or any memory or storage technology -- can or can't fit there.
The graphic below, which is explained in detail in this blog post from my company Objective Analysis, better demonstrates these points.
Memory and storage technologies can benefit a computer's cost-to-performance ratio as long as they follow the series of orbs in this chart. Speed is on the vertical axis and cost is on the horizontal axis. A new layer needs to be faster than the next-cheaper technology, and cheaper than the next-faster technology. So, if XPoint is slower than dynamic RAM, it won't sell unless it's cheaper than DRAM. If it's more costly than NAND flash, it needs to be faster than NAND flash.
Intel understands this very well. The company consistently sets its Optane products' prices to fit between DRAM prices and NAND flash prices, and it typically sells them for about half the price of DRAM. This has been tough on Intel because, with 3D XPoint memory being a new technology, the cost to manufacture it is higher than the price at which it sells. Intel sees Optane as such a strategic part of its server platform that the company, by our estimates, has underwritten losses of more than $5 billion to ramp 3D XPoint into volume production. Eventually, 3D XPoint should cost significantly less than DRAM to produce, but it hasn't reached that point yet.
3D XPoint memory uses a production process that is more costly per gigabyte than the NAND flash manufacturing process. This makes it unlikely that 3D XPoint will ever cost as little as NAND -- yet, some XPoint users and nonusers argue that XPoint's technical strengths warrant a higher price. Unfortunately, the market doesn't work that way. When designers choose which technology they are going to use, the deciding factor is whether the system can meet the performance objectives at the lowest cost. Invariably, memory and storage technologies with technical disadvantages nearly always win out over their better-performing, but more costly, counterparts.
3D XPoint uses
In 3D XPoint vs. NAND flash comparisons, the former can never compete with the latter on cost. So, who will use 3D XPoint? This is where Intel has made a clever, if costly, move.
Intel isn't trying to get designers to replace their NAND flash, HDDs and magnetic tape with 3D XPoint. The company is trying to harness a technology that it has been working on for 50 years to gain a competitive advantage over its rivals.
Wait, is 3D XPoint technology really that old?
It certainly is. 3D XPoint is an updated rendition of phase-change memory (PCM). PCM, also known as Ovonics, was first introduced in a September 1970 article co-authored by Gordon Moore, of Moore's Law fame, and Ron Neale, who still contributes posts to my blog. The article can be found here.
Intel's XPoint-based Optane products bring persistence closer to the processor. Data can be persisted a few orders of magnitude faster on either an Optane DIMM or an Optane SSD than is possible on a NAND flash SSD. Applications that require persistence can take advantage of this to improve their throughput, although the software must adapt to the new environment.
The applications that need high-speed persists are mainly financial databases. Programmers who write financial database code often worry how their code will perform if there's a power outage at the absolute worst moment. They then write code to accommodate this worst-case scenario, and that code invariably requires an elaborate series of persists to protect against the possibility that a power glitch will cause a financial mishap. The faster these persists can be made, the faster the system will run.
These systems were the first to convert to SSDs, once SSDs became a cost-effective way to achieve faster persistence than an HDD could. With Optane DIMMs, all of these persists become faster than they are on a NAND SSD.
Optane DIMMs use proprietary commands, layered on top of a standard DDR4 memory channel interface, to communicate. The only processors that can communicate with an Optane DIMM are Intel Xeon server processor chips. This means users who want to get the fastest persists need to adopt Intel processors. That's the key to Intel's Optane strategy: provide a compelling speed advantage and make sure it's only available to Intel users.
Note that nothing in this strategy involves displacing NAND flash. NAND SSDs are complementary to Optane DIMMs -- they fit into adjacent places in the memory/storage hierarchy. NAND flash and 3D XPoint should peacefully coexist for decades to come.
