3D Xpoint: Is It a Game-Changer?

26 May 2016 • 4 minute read

You have probably at least heard of 3D Xpoint. This is a memory technology jointly developed between Intel and Micron. I attended one of the keynotes at the Open Server Summit by Rob Peglar of Micron. He went for one of the less catchy titles of the day, Speed Up Your Applications With 3D Xpoint Technology. One issue is that there has been a limited amount of information about the technology, and it all comes from Intel and Micron themselves. If you watch the video below, Rick Coulson even says that they are deliberately not giving a lot of details. "We'll get it into production and then everyone can try and reverse engineer it."

So what is 3D Xpoint technology? (I'm going to call it 3DX from now on.) One way to look at it is that it is a better flash. 1000X faster than flash, 1000X more endurance (bits can be written a thousand more times before problems start to occur), and 10X the density. As I said above, it is denser than DRAM but only slower by a factor of four.

The big opportunity for 3DX is to change the memory hierarchy in a way that results in increased performance and decreased cost. The reason this is important is that there are two fast growing areas of memory use, the latency-critical, high-throughput "in memory" use for big data crunching. And the other end of the scale, latency-insensitive "cold storage" for preserving data that is not intensively used. By contrast, the blue stripe across the middle in the above picture is not growing, the traditional bread and butter of the corporate datacenter supporting relatively small (compared to internet scale) databases.

Let's start with some numbers:

On-chip cache access 1-10ns
DRAM access 100ns
Flash on PCI Express (PCIe) 10us
Flash on SATA (solid state disk) 100us
HDD on SATA (rotating disk) 10ms

That is quite a range. Where does 3DX fit in?

DRAM-like memory sitting on PCIe 500ns (4X DRAM)
SSD-like bulk-storage devices 2us

The standard memory hierarchy for many years was DRAM in the processor (and caches) and a hard-disk. If a page miss occurred and it had to be brought in from disk, the difference in latency was 10ms instead of 100ns, a difference of about 100,000 times.

The first thing that has been happening in datacenters is to add flash (and perhaps battery backed non-volatile DRAM) at various points in the memory hierarchy. This has reduced the difference between getting a value from DRAM vesus having to go and get it from the storage subsystem to a factor of 80.

Adding 3DX can bring that difference down to 4X. This doesn't just mean that things run faster, it is a qualitatively much smaller number and means that we can think about radically changing the memory hierarchy. In particular, removing HDDs from the servers and putting them elsewhere in specialized storage areas.

This offers the opportunity to increase the performance of the memory hierarchy for the same investment. It is already well known in the datacenter world that if you have SSDs on your system, you don't need so much DRAM compared to when using rotating media. 3DX offers the opportunity to reduce DRAM dramatically with no decrease in performance. Rick Coulson, who is an Intel Senior Fellow in non-volatile memory solutions, made a seminar presentation at Stanford in March this year, which is well worth watching if you want to know more than I can include in this blog post. One experiment that he pointed out was where they built systems with 3DX and then cut the DRAM back to what, these days, counts as almost nothing:

We did a little demo a while back where we took a PC and restricted it to 256 megabytes of memory. And we ran a side-by-side demo with an 8 gig system. You couldn't tell the difference. And the little meter on the paging was 200,000 page misses a second. And that was okay if you're fast enough.

3DX, with non-volatility, low cost per bit, and high performance offers the potential for even more radical transformation since it offers the possibility of relatively large persistent memory, therefore making it possible to make major changes in system balance between cores, memory and storage. In turn, this offers the possibility in datacenter design for better power, cooling, and (physical) space.

Here is a video of the Rick Coulson seminar at Stanford referred to above:

Next: Breakfast Bytes: Post #150

Previous: Andrew Kahng on Industry-Academia Cooperation