Share this post

August 25, 2015

Intel 3D XPoint – and what it means to software define storage

Priyadarshi Prasad - Atlantis


If we could merge the benefits of Memory (DRAM) and Storage (HDD/SSD), what would that look like? At the recently concluded Intel Developer Forum (IDF), Intel talked a lot about their new technology, named 3D XPoint (pronounced ‘crosspoint’) and how it could deliver the holy grail of storage performance and persistence but what is it?


3D XPoint

Before we discuss this new technology let’s look at the characteristics of Memory and Storage:

  Price Performance Persistence Density Endurance
Memory (DRAM) Expensive Fast Non-persistent Low Virtually Unlimited
Storage (SSD) Cheaper Slower Persistent High Limited
Storage  (HDD) Cheapest Dog Slow Persistent Medium Virtually Unlimited
Looking at the above table, the holy grail of memory-storage technology would be something that is:
  • Cheap
  • Fast
  • Persistent
  • Dense
  • With virtually unlimited endurance
 This is essentially what Intel is promising its new 3D XPoint technology to be. Very exciting indeed. Unlike NAND, it is not made of transistors, and doesn’t work on the principle of trapping electrons in cells to achieve its storage states. It appears more like a resistive technology that use difference in resistance as the storage states – this is helpful in improving endurance over the long term. 3D XPoint is also byte-addressable enabling granular R/W operations and that can explain its performance improvements over NAND. It is unclear though what material exactly do they use to manufacture 3D XPoint. GA is not too far out– the expected availability is likely to be in 2016. From an enterprise adoption standpoint, it will be interesting to see when server behemoths like HP and Dell start shipping 3D XPoint to their customers.

Is this the last nail in the coffin of traditional storage?

What does this mean for traditional storage? Quite simply, 3D XPoint will accelerate the extinction of traditional monolithic storage designs. Furthermore, it will relegate the all-flash array (AFA) segment to a niche. Why?
Software defined storage (SDS) is by no means perfect today. Compared to the old traditional (think VMAX) systems, SDS solutions offer relatively limited capacity scaling. And before you say it, yes you can scale out SDS systems to very large clusters but remember that every node you add to your SDS solution would likely increase the amount of network traffic, potentially creating a bottleneck in communication and impacting latencies. However, good SDS solutions easily offer impressive scalabilities thanks to their data reduction technologies such as deduplication and compression offering very large effective capacities without wreaking a havoc on your budget, space, power and cooling envelop. And it gets better with 3D XPoint…
With 3D XPoint, you can get 10x better density than DRAM. And this is with 2X process geometry. When it matures and goes to 1X and 1Y geometries, the densities will continue to improve even more, easily meeting 3D NAND geometries. All of a sudden, I can think of supporting 16TB 3D XPoint devices without the accompanying NAND related issues of write performance and endurance. With four such devices per server in a cluster of eight servers, we would be able to support half a petabyte of raw capacity. From my experience in enterprise storage, there are not many storage systems deployed with more than half a PB of raw capacity. Once you take deduplication/compression into account, you are talking closer to 2.5PB of effective capacity in a single system!

Goodbye AFA, hello HCA

The old traditional storage systems are on their way out any way, it’s not a matter of if but when, but what about the new shiny flash arrays – how does 3D XPoint impact them? Here is a question – if one can provide a memory-like technology within servers, that is obviously fast (faster than flash), doesn’t have endurance issues like flash and offers similar density and price points as flash, why in the world would one look at an all-flash array. A possible answer could be that AFAs would adopt 3D XPoint as well. Sure they will but look at this Intel video:
They briefly allude to a fundamental memory architecture element – the best performance is realized when data is closest to compute, either right on the motherboard or as direct-attached storage within servers. AFAs are traditional storage design in that they still treat compute and storage as silos. SDS with 3D XPoint brings a faster media closer to the compute. Fast, simple and cheap generally wins because developers quickly figure out how to handle failures and make things resilient – that’s what software is for.
What do you think? Join the conversation at @AtlantisSDS, @Priyadarshi_Pd.
Current rating: 4.5 (6 ratings)