I was privileged to attend a morning of presentations from the Intel Optane team last month (September 2018), straight at Intel headquarters while in the Silicon Valley. Here are my thoughts on Intel Optane, its future, and which use cases are relevant to this storage class.
A difficult birth
Intel Optane, the commercial name for Intel’s implementation of 3D Xpoint memory (pronounced three-D crosspoint). It is a class of persistent memory used for data storage, and available in various form factors (SSD, PCIe card, and DIMM-compatible modules). It was developed by a now defunct partnership between Intel and Micron.
Optane, despite its merits, is a kind of unloved child in the storage industry. 3D XPoint has been long expected and was released with a significant delay over announced delivery times. To add insult to injury, Intel did serious mistakes in marketing the product: they were very confident on the level of performance they would achieve with Optane. Perhaps too much: 3D XPoint was over-sold on performance and under-delivered – performance was announced to be initially 1000x faster compared to NAND, while final specifications spoke only of a 10x improvement.
Most articles published during 2017 about Optane and 3D XPoint (broadly available online) still reference the 1000x faster speed. This was a serious blow to Intel’s credibility and it damaged Optane’s image in the eyes of the public. With only a 10x reduction in latency, this would make the product lose some of its flair.
Delays in delivery, a defunct partnership and significantly reduced performance: what does this means for Optane? While it’s not dead-on-arrival, do you think it is artificially maintained into life? Or does it has space to fly with its own wings?
Issues with NAND and what drove 3D XPoint development
Optane and the 3D XPoint memory class were created to address issues related to NAND memory architecture that have been crippling the flash storage industry for a long time.
Durability and Density
NAND Flash memory is known to have durability issues related to how many program/erase (P/E) cycles a flash cell can sustain.
This level varies based on the memory design. Without getting into hardware design, NAND memory uses a flood gate mechanism where cells are either positively or negatively charged to indicate the bit value of the cell. In the case of SLC (Single Level Cell) memory, the threshold is 50%. Below 50%, the bit value is set to 1 and above 50% it is set to 0. MLC (Multi Level Cell) adds more complexity to this with 4 charge levels (holding 2 bits per cell), and TLC (Three Level Cell) further increases with 8 charge levels (holding 4 bits per cell).
Trying to keep this article focused on Optane and not on NAND memory architecture, the good thing with moving from SLC to TLC is that the density increases (4 bits instead of 1 per cell), but at the cost of more stringent requirements on accuracy in readings. After a certain number of full writes over a cell, the cell would at some point wear out. The challenge with the overall media durability is that when you “lose” an SLC cell, you lose only the ability to store 1 bit. When you lose a TLC cell, you’re now “losing” 4 bits of capacity.
But wait! This is memory, it’s all electronic and digital, there are no moving parts! So why would it wear out! It isn’t one of those cranky hard drives where we use analog things such as magnetic heads to read data on a platter!
Well, that’s where things get funny. Flash media is also made of chemical compounds, and without getting into details (because I’m not a scientist and those things are waaaaay to complicated for my simple mind), programming and erasing cells leads after a certain amount of cycles to an eventual wear out. It doesn’t mean that the cell is out of function, but perhaps that the threshold between a 1 and a 0 (i.e. the electric charge level) can no longer be predicted? Looking for an analogy? Cell phone batteries and battery packs. They hold their charge well first, then they deteriorate. Just like us humans and our brain cells.
This has been an issue since the early days of flash memory and many vendors had to come up with their own implementation of a software stack that would address flash memory space in a way that would prevent or minimize this wear. And imagine that to simplify things I didn’t cover 3D NAND and QLC.
Performance
Performance on NAND Flash is also a bit of a struggle. For the sake of brevity, I cannot get into a detailed view. Let’s assume that we no longer rely on SATA but on the PCI Expres (PCIe) bus, so bus considerations are out of scope. Read speeds are faster than write speeds (in the microseconds range) but erase operations take considerably longer and are in the milliseconds range.
Why are erase operations taking more time? With NAND, single cells can be programmed but only an entire block can be deleted. Erase operations are a kind of write but writing the same value across an entire block – and fully resetting the state of each single cell. Also, add the density factor to the equation: it takes more time to erase a block containing TLC cells than SLC cells. Finally, on intensively used systems with high data modification rates, this puts weights on the program/erase cycles of the device.
3D XPoint : fundamental changes
These issues are what drove the develoment of 3D XPoint. It was important to build an extremely reliable and durable product, which is undoubtedly a success with Optane. The 3D XPoint architecture is novel compared to NAND. It doesn’t use any flood gates or electrons to indicate a cell charge. Unsurprisingly, it uses a tri-dimensional cross-point structure where cells can be addressed by determining which are the top and bottom wires relevant to it.
The material used to build 3D XPoint memory seems to have peculiar memory storage capabilities that make it able to work solely based on electrical resistance variance. A cell can be written to or read from based on the amount of voltage that is sent to it, and without the need for a transistor. This use of a new material makes 3D XPoint memory much more durable than NAND. The elimination of transistors also makes up in turn for more available space, which increases the storage density. It’s hard to write a lot on 3D XPoint design because it feels like the architecture is being obfuscated. At least, I had a real hard time trying to find technical materials on the matter. I promise however to write a more comprehensive article if I come up later with more specs.
These innovations allow Intel to market Optane devices which have astounding capabilities. The Optane SSD DC P4800X for example has a rating of 550,000 IOPS for read & writes, at 10 μs read and write latencies. In comparison, Intel’s latest SSD flagship, the SSD DC P4610 – which uses 64-Layer 3D TLC NAND, is rated at 620,000 read IOPS and 77 μs read latency, while it is rated at 200,000 write IOPS and 18 μs write latency.
Optane is ahead on consistency regardless of the workload balance between writes and reads, but it’s hard to envision it as a mainstream media that would entirely replace 3D TLC NAND.What makes Optane (and 3D XPoint in general) particularly appealing is the resistance of the media. In fact, it appears to me that it is best placed in front of 3D TLC NAND / Enterprise NVMe, but more on this below.
Market Orientation & Use Cases
3D XPoint had been thought with the word “extreme” in mind: extreme durability, extreme performance, extremely low latency. And in the storage market (just like with cars or other mundane objects) whatever is “extremely” reliable and fast usually sells at a massive premium, not just because it’s cool, but because massive investments were made to develop a new technology including new manufacturing processes and equipment.
A parallel would be to compare prices of flash media in 2008-2009, because they too at the time were game changing in terms of latency and IOPS.
Optane in the Data Center world
Optane is currently too expensive to be used as a main storage media but is still a cheaper alternative to NVDIMM (Non-Volatile DIMM), offering consistently more capacity than an NVDIMM at approx. 33% of the price. If DIMM-level latency is not desired, but if latencies in the single to double-digit microseconds range are expected, Optane is right at the sweet spot.
As said above, a hefty price tag restricts its usage mainly to data caching and tiering mechanisms. Dell EMC for example uses Optane into their PowerMax systems for the most time-critical caching operations. NetApp, on the other hand, uses Optane as the first tier of their DATA Max software defined persistent memory solution. This proves that Optane has a use case in the data center market, but for now due to the price tag it is confined to supporting Tier-0 solutions, which will still rely on more affordable tiers of data for persistent, long-term storage.
Optane in the Consumer Electronics world
The lack of interest from a sales perspective made Intel re-orient their marketing messaging (at least partially) from the data center world to the gaming world. Die-hard gamers are being massively hit at by Intel on the benefits of Optane drives. I’m personally not convinced that Optane brings a huge improvement to a gaming workstation, but please take that with a grain of salt: my favourite games are from the 90’s. Worst case it will make you look cool with your friends by having the latest available expensive piece in storage.
It doesn’t means that Optane isn’t making sense there, but again the price tag is really high, and what we’re seeing are 16 GB or 32 GB Optane modules in M.2 form factor, so while it can improve speed, it still doesn’t qualifies as a full storage tier.
Max’s Opinion
Is Optane really a lost cause? Only time and the market will tell whether Optane is promised to a bright future or not. Perhaps the biggest tragedy of Optane has been the very poor marketing that went around it. Overselling and underdelivering never makes a good impression. Assessing what went wrong here would be a topic of its own on which I will not argue.
What makes the future look gloomy for Optane is the recent development in the partnership between Intel and Micron. It’s been nearly a year since Optane products are available on the market and the partnership is gone. Also, we don’t even know if there will be a second generation of 3D XPoint products created.
My challenge with Optane is that I’m struggling very hard to imagine a mainstream use case that would lift the product up and bring the expected profitability that Intel (and Micron) had been betting on. While it does provides performance benefits, the premium price tag makes it difficult to adopt in a storage market which is dominated by the quintessential $/GB cost ratio matter.
Indeed, customers are on the hunt for great performance at affordable prices, but not all of them are looking at Tier-0 storage arrays. Because Optane’s primary attribute is its durability, it may keep lingering into niche use cases such as being used as the resilient and durable caching/tiering layer within Tier-0 performance solutions. At least for the enterprise storage market. It is not in my capacity to say whether Intel’s bet on the gamer / high performance workstation market will work or not.