Intel to Launch 3D XPoint DIMMs in 2H 2018 | Hacker News Reader

Intel to Launch 3D XPoint DIMMs in 2H 2018

anandtech.com

79 pointsp1esk41 comments

Do these really have RAM-like performance or is Intel calling them DIMMs as a marketing ploy?

I still remember when Optane was first announced Intel hyped it as a replacement for RAM. When it was finally available it turned out to be about 1,000x slower than RAM and barely faster than other SSD. So much for RAM replacement...

Barely faster than NVMe SSDs when fully saturated. At low queue depths and in mixed workloads (especially when overwriting data) Optane absolutely destroys the competition. This is a database dream drive. Only problem is the low capacities right now.

> Barely faster than NVMe SSDs when fully saturated.

The Optane products released thus far are all literally NVMe SSDs, which is why their best-case latency isn't much better than that of flash-based NVMe SSDs. They still have PCIe and NVMe protocol overhead to deal with, which the 3D XPoint DIMMs won't have.

The unimpressive peak throughput of the current Optane SSDs is largely a consequence of the Optane SSD controller having a fairly low channel count. A single 3D XPoint DIMM will probably have substantially higher throughput than a current Optane SSD.

It's much better in latency too[1].

I also wonder if the current products are bottlenecked by the controller and whether DIMMs will see vastly superior performance. Guess we'll have to wait and see.

[1]https://www.pcper.com/reviews/Storage/Intel-Optane-Memory-32...

> I also wonder if the current products are bottlenecked by the controller

There's no question that they are. NVMe drive prototypes that use DRAM as their backing memory instead of flash or some other persistent memory have about the same overall latency as Optane SSDs.

With NVMe, the storage industry is in a much better position to take advantage of 3D XPoint than if we were all still using SATA or SAS, but it's still on a peripheral bus not a memory bus.

Intel promised 512GB per DIMMs. But looking at those Optane Drives results I think its much lower bandwith may be a problem.

Just waiting to see product review, it has the potential to bring In-"Memory" computing to next level.

Optane also looks like it would make a nice ZFS log device.

Not with the comically low dwpd people have been seeing.

NAND already has reached it's peak. I don't see why people are so disappointed that the first generation of a rushed product is competitive with todays high end NAND SSDs.

There was a time that Intel was telling us that this was so much better than NAND that it represented an entirely new tier in the storage hierarchy, sitting between DRAM and SSDs.

"DIMM" stands for dual in-line memory module, it is not about speed. It means that there are pins on both sides of the module with different signals (as opposed to SIMM).

Technically you are correct (the best kind of correct). But on a standard PC motherboard/configuration, DIMM slots are used exclusively for RAM which is much faster than ROM/storage. 3dxpoint was specifically marketed as being almost like a RAM/ROM hybrid in application. You could have 32GB RAM for regular computing plus 512GB of optane in the remaining DIMM slots and it would be like having swap space with very little speed penalty.

So GP is right to be concerned about speed for these form factors. If it's not faster than an NVMe drive them it's not worth the price.

I expect them to have RAM like performance because they'll have a RAM mirror similar to existing NVDIMM-N's.

That's not how they work. The capacity is much larger than any DRAM DIMM.

Not the ones I've seen: for example: https://www.micron.com/~/media/documents/products/data-sheet...

NVDIMM-N's have more flash than RAM simply for wear leveling concerns. Micron's 32GB NVDIMM-N has 32GB of RAM and 64GB of flash, but the capacity of the NVDIMM-N is equal to the amount of RAM.

Well, 4x for the moment. And manufacturers could start making DDR4 with 8-high TSV stacks that would cut it to 2x.

And considering Intel is positioning these as an exclusive enterprise product, RAM is probably going to be competitively priced, even in terms of storage space.

This could be an amazing technology for high performance mobile devices, especially laptops.

For this first release, "3D XPoint DIMMs will require server platform support because they are unlikely to operate as standard DDR4 DIMMs. The JEDEC NVDIMM-P standard for persistent memory DIMMs has not been finalized and is expected next year".

Whether they have an $8B market or not will very much depend on the performance and price. I really hope they can make it work though. The first generations of a new technology are typically not much to get excited about, but as it gets better over time this could be the first major shakeup to the memory/storage hierarchy since the SSD - with potentially bigger ramifications because it's byte-addressable.

so, as someone naively excited about memristors I've noticed the similarity between Xpoint and memristive memory... am I totally off here though?

XPoint is a form of chalcogenide resistive memory while HP memristors are titanium dioxide resistive memory.

is this a measure to avoid patent conflicts? (or is there an appreciable difference in inner workings)

Not a patent issue. PCM is also a NOR flash device, and have been used for decades in shipping memories.

The answer to your paranthetical depends a bit on how much detail you care for. Both devices are types of resistive crosspoint memories. In XP, the resistor is a chalcogenide that has a low resistance polycrystal line state, and a high resistance amorphous state. These are tuned by the program pulse length. In TiO2, the mechanism is (presumably, perhaps debatebly) filamentary bridges between electrode, which can be eliminated by reversing the voltage.

So on one level, both are resistive memories. On another, we are comparing a bulk phase transition driven by pulse control to a filamentary transition driven by voltage.

Finally, the more important difference is the selector. PCM devices can use a chalcogenide based threshold switch. That is the OTS on the XP diagrams. That selector is the key to functional, dense memory arrays, and the existance of a matched selector is what makes PCM functional. I haven't seen what the matched selector for memristors is.

Worth reading recent material from Crossbar on these topics.

One thing I wondering is what Micron is planning on doing with 3D Xpoint. It was a joint venture between the two firms, after all, but it seems like Intel is the only one doing anything with it so far. I haven't even heard anything.

Micron has shown a few prototypes, like an SSD using a controller from CNEX Labs, who are best known for their open-channel SSD work.

I didn't really expect that Intel would end up this far ahead of Micron in bringing 3D XPoint to market, but it has been clear for quite a while that Intel has the dominant role in their relationship.

Intel NVM Systems and Micron have been partnered since the early days of both of their starts in NAND. Investment in R&D and fabs are shared between the two companies. Intel is largely pushing 3DXP down the pipe from the CPU groups as a performance advantage. If/when it really takes off, they will need added support/investment from Micron to increase supply, reduce risk, etc.

That will change things. I really would like a nice hibernate function that didn't require saving anything to disk.

Why? Hibernating to an SSD is pretty fast already.

Imagine zero-draw "hibernation mode" taking a few milliseconds to enter, and "waking up" in a few milliseconds, too. Apply to mobile devices.

Mobile devices have battery, so they can power RAM allowing sleep. Also I've read that Optane is power-hungry, so it won't be used in mobile devices anyway.

This stuff promises sub millesecond hibernation. Things like hibernating between sentences kind of hibernation. Doing that could conceivably extend the effective 'use' time of a laptop with a 99Wh battery to a week or more between charges.

Hmm...still several seconds in my neighborhood.

If it were 0.1 second I think life would be quite a bit more pleasant.

We're talking CPU idle and hibernate being nearly essentially the same. Instant.

This reminds me of core memory, which was intrinsically nonvolatile and likely has more endurance per bit than any flash-based storage today. I wonder how it handles wear leveling, since being in a DIMM format means the CPU(s) can write to it far more aggressively than an SSD.

It doesn't require wear leveling like SSDs do. It's like RAM in that regard.

I strongly recommend this talk from StrangeLoop on persisted memory: https://youtube.com/watch?v=VE1hCUMLHX4

> It doesn't require wear leveling like SSDs do. It's like RAM in that regard

Are you sure? The article lists a fixed durability of 30 drive writes over 5 years. If it wasn’t wear leveling, those 30 writes would be reached very quickly on specific locations.

30 drive writes per day, which is over 50K writes total. But at RAM speeds you could hit 50K writes in no time (XPointHammer? you heard it here first), so wear leveling would probably be a good idea. And here's my favorite wear leveling paper: http://researcher.watson.ibm.com/researcher/files/us-moinqur...

That's really interesting. At around 12:30 minutes in the talk I linked the speaker, Rob Dickinson (from Intel), says it doesn't need it. I wonder which is correct.

Edit: fixing timestamp and adding speaker's name.

sad part is you gotta combine it with an EPYC cpu otherwise people can potentially grab your DIMMs and read the data

Epyc has SEV encrypted memory and Skylake has SGX encrypted memory (although I wouldn't expect SGX to actually work with XPoint).

With free added back door!

Reply on news.ycombinator.com