Several months ago, Intel and Micron demonstrated 3D XPoint (Cross Point) for the first time. It was advertised as a quantum leap for storage technology.
3D XPoint may very well be one of the more exciting upcoming technologies for storage and solid state drives around. Although we are seeing the end of Moore’s Law and a near breakdown in single threaded performance improvement, there are still quite a few exciting technologies that may dramatically boost the performance of computers as we know them. GPUs continue to progress, as does storage technology.
So far Intel and Micron have declined to give the technical details. They claim that it is “different” from Resistive RAM (RRAM) but won’t say how. Patent searches suggest though that it is based on chalcogenide materials for selector and storage. Intel and Micron have been very tight lipped though so there is no way to know for sure until we get more details. Obviously they are keeping it quiet for competitive advantage for now. I suppose this adds to the “mystique” of 3D XPoint for now.
Just how well does it perform?
Anandtech wrote an outstanding article about 3D XPoint and I highly recommend that if you are interested, that you give it a read. Well, look at the chart below from Anandtech.
|Endurance (P/E Cycles)||10^15||10^7||10^3|
|Read Latency||Nanoseconds||10s of Nanoseconds||~100 Microseconds|
Well, from what we can see, it is clearly much faster in read latency than NAND – by about 3-4 orders of magnitude, and about 4 orders of magnitude more durable in terms of P/E cycles. Yet, it is still slower than DRAM by one order of magnitude in read latency and DRAM effectively lasts forever, while XPoint clearly has a finite endurance, although for practical purposes, I suspect that the P/E cycles will not be a concern on 3D XPoint. Actually, they may not even be a concern on NAND – in practice independent tests have shown that the actual life of SSDs is as much as 20 times more than the rated cycles.
We don’t know what the write latency is like, but I suspect that it will be quite a bit better than NAND no matter what.
Anandtech’s writers believe that the price of 3D XPoint will be higher than that of NAND, which is used in current SSDs, but lower than that of DRAM. So it won’t replace either technology – DRAM for performance reasons, and NAND for cost reasons, but is meant to complement both. The big driver of adoption or barrier then, may be cost.
Why does any of this matter for consumers?
This would be worthy of an article itself, but in 2014, NVMe was introduced as a new standard that was introduced because the old standard, AHCI, could not keep up with upcoming SSDs. SSDs had outperformed the AHCI standard, having been made for hard drives. NVMe saw the following changes:
- The latency halved from 6 microseconds to 2.8 microseconds – the most important of these for consumer workloads
- Queue depth was changed from 1 queue with a max queue depth of 32 to 65535 queues and 65535 commands per queue
- No locking required to synchronize multiple threads
- One 64B fetch for better 4KB efficiency (previously AHCI needed a pair of serialized fetches through DRAM)
- Up to 2k of MSI-X interrupts (AHCI had only 1 with no steering)
So far most of the performance gains from new SSD technology – particularly in the case of NVMe has mostly benefited sequential performance. For us end consumers, Random is more important than Sequential. The reason being that most consumers are most often are reading and writing lots of small files, not large files. Prosumer and enterprise are of course different.
For consumers, the most important of these is the fact that the latency was halved because that is likely to have the largest effect on random performance. The main reason was because the NVMe standards streamlined the storage stack process. Unlike AHCI, NVMe did not need to read registries. That being said, it doesn’t solve the problem of the bottleneck of NAND speed. We can make a faster controller and that has been the huge driver of SSD performance, but the NAND has not gotten faster. I suppose that the fact that multiple queues can be valuable too, especially now as more cores have become more mainstream and single threaded performance improvements seem to have leveled off.
Look at the image below:
This is what really matters for us consumers. Your applications will load much faster, and your PC will boot much faster because of this. This is especially true if we can pair it with a very fast controller. Perhaps it is not as fast as the fastest RAM Disks, but that is still a huge step forward.
Remember that for consumers, random performance at QD1 is the most important performance metric because we are constantly reading and writing lots of small files.
The problem of diminishing returns
For most consumers, we get a noticeable leap in performance when we jump from the traditional hard drive to the SSD. That is because we see a jump of latency from milliseconds in hard drives to microseconds in SSDs. A jump from microseconds to milliseconds will not be nearly as noticeable. They are still there but … diminishing returns.
Actually, the problem is that for most consumers, cost is by far the most important metric for adoption of NAND or 3D XPoint. NAND currently costs several times that per unit of capacity compared to a hard drive, although NAND’s price does continue to fall. One serious problem is that SSDs for most consumers are simply “good enough” as is. Price will be the big barrier then.
I think that when 3D XPoint is introduced, we will see it introduced in the enterprise world, then for consumers. We may not see mainstream adoption for years though. Enthusiasts are likely to be the first consumers that adopt it once it becomes released to the consumer world, alongside certain prosumers.
Intel and Micron claim that 3D XPoint is compatible with Extreme Ultraviolet (EUV) lithography, so perhaps we shall see the price fall down below NAND and hard drives with smaller processes – that is assuming they can solve the EUV technical difficulties anyways. There is also the claim that unlike NAND, 3D XPoint gets better with smaller process nodes! I think that at this point, all we can do is wait with cautious optimism. In the end, if these claims are true and we should be skeptical because companies always want to overhype their products. Yet if it proves to be true, then it may be cost that causes the adoption of 3D XPoint over the alternatives.
Finally, we need to wait and see what competing technologies the other giants of storage release. I suppose that we will find out in the next couple of years.
Despite the diminishing returns, it still represents an impressive step forward for storage technology.
I think that for the end user, subjective “responsiveness” of a computer will be better with 3D XPoint than existing drives, but it will be hard to notice with SSDs the gap. Professional and enterprise users that rely heavily on the sequential performance will again, see the biggest benefits.
Perhaps if Moore’s Law ends, at least we will had our cake and had the chance to eat it – a very fast storage, that has made P/E cycles, and is affordable in terms of units per GB.