News Optane's Last Gasp: Intel's Final Persistent Memory Roadmap Leaks

[Admin]

Intel's Crow Pass gets detailed: 3 GB/s random bandwidth, 6 GB/s sequential bandwidth.

Optane's Last Gasp: Intel's Final Persistent Memory Roadmap Leaks : Read more

 

[cyrusfox]

Could we get someone to play taps for Optane, last product for it here? Seems like it. Agreed with this article , Optane we hardly knew yeah

I did grab 905p 960GB for cheap, and tempted by the cheap P1600x 118gb, I should grab at least one of those. How long will it be until we see another emerging memory outside NAND and DRAM hit the market?

So Long Optane, will be a long time if ever we see anything like you again.

 

[thestryker]

Could we get someone to play taps for Optane, last product for it here? Seems like it. Agreed with this article , Optane we hardly knew yeah

I did grab 905p 960GB for cheap, and tempted by the cheap P1600x 118gb, I should grab at least one of those. How long will it be until we see another emerging memory outside NAND and DRAM hit the market?

So Long Optane, will be a long time if ever we see anything like you again.

I picked up 4 of those P1600X drives for when I build a new server box since I'm not going to be using a raid controller I want to utilize them with ZFS. I kicked around one of the 905p drives, but I don't know which platform I'm using yet so I don't know if I'll have the extra PCIe lanes needed.

The only SCM other type technology I've seen benchmarked is Kioxia's XL-Flash, but I've never seen any products make it to market. I think this has to do with the financial/legal issues the company has had, but it's hard to say. This isn't a new technology, but a modification of NAND so one would think that this shouldn't need dramatic modification of fabs to make. It also seemed to perform pretty close to Optane so at least on the storage side it could make for a viable stand in.

 

[DavidLejdar]

Next for Intel is apparently CXL - which should in particular enable memory to be attached to the CPU over a PCIe 5.0 link. And that shows that there is a push for increased data transfer rates, speed and latency, as CPUs could apparently perform way higher than what even DDR5 and Gen5 NVMes can deliver there these days.

 

[thestryker]

CXL replaces Optane Memory, but does absolutely nothing for the storage side of things. There is no non-volatile memory technology on the market that has good random performance now that Optane is a dead end (there's also the absolutely absurd endurance which no tech can come close to). I'm unsure if Intel just bet on the wrong usage for the technology, or if the storage side was just never in the picture for volume sales. I know Gelsinger never had any interest in Intel doing anything with memory tech (or as he put it making dumb chips) which is why he was quick to bail on the NAND business.

Just taking a wild guess here that the cost of spinning up a fab and developing the next iteration of 3D XPoint would have been way too big of an investment for the storage market (the memory side being a dead end due to CXL) that they would have to try to grow on top of all of the engineering costs. It's still a great shame because it leaves a gap in the market which is unlikely to be filled. I firmly believe consumers are the real losers here as enterprise can just throw money at circumventing the issue while consumers will be stuck with just increasing sequential R/W on NAND which is mostly irrelevant.

 

[DavidM012]

Question: Is there no technology that would combine memory and storage into one device? So if you could have a hard disk at the speed of ram and memory with the capacity of a hard disk wouldn't that speed up the entire system? It would be like loading your games on a ram disk and since Optane can be non volatile couldn't they just redesign the architecture a bit to have a fast hybrid memory/storage interface to the cpu and be done with it?

I'd guess reasons why they wouldn't do anything like that, to compete in the storage and memory market at the same time well I suppose you lose a bit of fault tolerance: If your memory and storage goes out at the same time you're done whilst at the moment your storage can go down but doesn't take your memory with it so then they would probably have to vastly overbuild the fault tolerance and reliability - but Optane had high endurance?

Maybe could solve the fault tolerance by having 2 hybrid modules or something.

Question is would it have high endurance as system memory?

There could still be a market for ssd's for more storage even if there were a hybrid memory/storage drive to run the OS and apps from.

Well the CPU needs a pool to dump data in and they don't make cache memory larger since it's expensive and runs hot?

It sounds like the main problems of hybridising memory and storage would be getting customers to trust it, and simply developing something at a price that can duel with dimms and nvme so what if you could produce a universally compatible interface and have hybrid memory/storage devices with capacities starting at 500gb up to 4 tb well it would sort things for customers who need large amounts of memory and sort things for consumers who only want a steam library.

It might only be sci fi though since I don't know the technical reasons why the pc arcitecture couldn't be redesigned and a new technology introduced on a new line of motherboards - maybe it's just too much of a risk, to get people on board with the idea, and put the pieces together is too complex and it all depends on whether people would buy it and if it was durable enough.

At the moment anyway we're loading data onto our drives via the network and usb since cd's, dvds and blu-ray have all receded the way of the floppy disk.

So it's just having storage and memory that can move data around the pc as fast as the cpu innit? I see no reason (besides any unknown technical hurdles) that a company couldn't simply design a pc with a new architecture, plonk it on a shelf and see if people buy it. The main selling point would be maintaining compatablity with x86/64 software. If you could keep a lid on r&d costs and market at a competitive price,

It might grow or it might flop or it might cause chaos, who knows? The main silly mistake you could make would be to produce something that underperforms against it's competitors.

You could keep some of the modularity of the PC to make space for upgrades and well maybe just cut out this whole budget/hi end approach, make everything middle so manufacturing costs would be lower and everyone pays a reasonable price for the same thing.

The question is would be technically possible to repurpose some of this tech? Well I suppose things developed the way they did simply because CPUs roll off the production line in different grades.

That's no reason you couldn't hybridise storage and memory and design yet another new mobo interface?

 

[Kamen Rider Blade]

Intel needs to sell the IP to the Memory Manufacturers and get everybody in the industry involved to use it.

 

[rluker5]

4 channel 4400 Optane is going to be a faster drive than Nand will ever be capable of in a pc form factor. I've seen the DDR4 dimms for sale, but I'll never have the hardware that has the permissions to use it. If they enabled using these DDR5 standard modules for HEDT I'm sure they would get more publicity.

But for the average consumer it isn't worth it. It works great with my 13900kf, but for my old 4980hq a sata nand ssd is good enough. I just picked up a 1TB Team sata ssd for that system for $47 to use a Christmas gift card. Basic nand is so cheap right now and isn't that bad to use.
But the same reasons for not getting Optane also apply to not getting anything but decent bargain gen3 NVME. I also have a Hynix P41 and it isn't as fast in doing anything but the right sized transfer as my Optane. Too small is slower, too big is slower, too full is slower, OS drive or gaming is slower, mixed is slower. The form factor is nice as long as it isn't used too much and throttles. That's why gen3 is nice, it is lower power, Gen 4,5 are just second best for a lot more money than basic.

 

[USAFRet]

Question: Is there no technology that would combine memory and storage into one device? So if you could have a hard disk at the speed of ram and memory with the capacity of a hard disk wouldn't that speed up the entire system?

We had those. Hybrid drives.
A small solid state segment and a large spinning drive.

Typically 8-16-32GB of SSD, and someTB of spinning.

When solid state was expensive, they were sort of a transitional thing. Sort of good for repetitive data, where the most used blocks of data ended up in the SSD portion.
But with falling SATA III prices, those went away fairly quickly.

 

[DavidM012]

I mean hybrid memory and ssd. Do away with dimms and nvme, or at least relegate nvme to secondary storage like hdd's have been, and have one optane based module that can run the OS and data storage at the speed of memory and memory with the capacity of storage. All solid state. One drive serving as memory and storage - possible since Optane is non volatile?

And if your system ram is non-volatile shouldn't it be that you don't lose data in a power cut and it kind of changes how the system behaves a bit and loads programs since memory wouldn't be cleared unless you instructed it to. So if you had an app crash or something, maybe a reset wouldn't clear it if the ram wasn't volatile if you see what I mean. You'd need to reprogram the OS a bit maybe to run a virtual machine per program or something.

So might be sci fi, and I don't really know what the technical hurdles might be just wondering if it isn't 'possible' to run memory and storage off one module with a new interface to the cpu to run storage that's as fast as memory since nvme speeds are catching up to system ram speeds anyway.

All that really matters is the speed of the various buses or how fast they need to be to load data and shuffle it around the system.

10gbps ethernet could be just as fast a way to load data as a 10gbs nvme? So what does it matter where the data is sourced from? Thin clients could possibly run off ethernet instead of having drives so long as there was infrastructure to support them so it's just a matter of the advantages and disadvantages of each data transport bus isn't it?

 

[USAFRet]

I mean hybrid memory and ssd. Do away with dimms and nvme, or at least relegate nvme to secondary storage like hdd's have been, and have one optane based module that can run the OS and data storage at the speed of memory and memory with the capacity of storage. All solid state. One drive serving as memory and storage - possible since Optane is non volatile?

Thats what Optane was supposed to be.

Very very fast (near actual RAM speed), non-volatile.

Cost, need, the rest of the needed ecosystem....

 

[Diogene7]

To my knowledge, there exist at least 2 emerging memory type technologies that are also non-volatile, while having reasonable power-consumption (I would think lower than Optane), maybe reasonable R/W cycles endurance before failing (maybe 10E10 or more), reasonable latency (below 100ns R/W access time), but not necessarily dense enough (Mbytes density at the moment…), and likely still very expensive :

1. Nantero carbon nanotube NRAM memory
2. Everspin STT-MRAM (Magnetic RAM) (it also exist SOT-MRAM, VG-SOT-MRAM, VC-MRAM,…)

Those 2 technologies have the potential to bring disruptive improvements but the 2 main reasons why those technologies are noy yet spreading much are :
1. Cost/economics/Return On Investment (ROI) : As they are not in very High Volume Manufacturing (HVM), they are costing much, much more (maybe 10x, 100x, 1000X…) than the currently mainstream technologies. Furthermore worldwide memory manufacturers (Samsung, SK Hynix, Micron,…) already have a lot of capital invested in standard memories, on which they want to get a Return On Investment.

2. Technical limitations of 2.1 memory controller and 2.2 emerging memory itself : memory controllers are integrated with the CPU or chipset, and so are gating the access to new memory. (Ex: In 2023, some Intel new CPU come with DDR5 / LPDDR5 memory controller, and so if the new emerging memory can match those exact specifications, it would not work on a DDR5/LPDDR5 bus, and so can’t be integrated).

In theory, the new CXL standard (ex: CXL 3.0) on PCIe5 or higher will open new opportunities, especially integrate the memory controller with the memory, not the CPU/chipset.

You just need a CPU that is compatible with PCIe5 / CXL (which in data-center should begin to appear in 2023/2024…), then on the PCIe5 bus, connect a card full of non-volatile memory, below 100ns R/W access NRAM (or MRAM) that have a CXL to NRAM (or CXL to MRAM) driver.

Read/Write access time to memory on a CXL bus (PCIe5 bus) will be a bit slower (maybe between 100ns to 1000ns), but still at least 10x to 100x faster than NAND flash SSD (10 000ns)

The key enabler is implementation of the CXL protocol as it will allow new emerging memory from startups (so not as good, neither as cheap as mainstream memory) to be integrated in computer systems : without CXL, they are gated / limited to the only memory types the CPU/chipset designer as integrated to its chip…

Hopefully, we may see a surge in R&D investments (by several startups or else) being done to size this new opportunity, and from there see new types of memory (non volatile memory) come to market.

 

[rluker5]

To my knowledge, there exist at least 2 emerging memory type technologies that are also non-volatile, while having reasonable power-consumption (I would think lower than Optane), maybe reasonable R/W cycles endurance before failing (maybe 10E10 or more), reasonable latency (below 100ns R/W access time), but not necessarily dense enough (Mbytes density at the moment…), and likely still very expensive :

1. Nantero carbon nanotube NRAM memory
2. Everspin STT-MRAM (Magnetic RAM) (it also exist SOT-MRAM, VG-SOT-MRAM, VC-MRAM,…)

Those 2 technologies have the potential to bring disruptive improvements but the 2 main reasons why those technologies are noy yet spreading much are :
1. Cost/economics/Return On Investment (ROI) : As they are not in very High Volume Manufacturing (HVM), they are costing much, much more (maybe 10x, 100x, 1000X…) than the currently mainstream technologies. Furthermore worldwide memory manufacturers (Samsung, SK Hynix, Micron,…) already have a lot of capital invested in standard memories, on which they want to get a Return On Investment.

2. Technical limitations of 2.1 memory controller and 2.2 emerging memory itself : memory controllers are integrated with the CPU or chipset, and so are gating the access to new memory. (Ex: In 2023, some Intel new CPU come with DDR5 / LPDDR5 memory controller, and so if the new emerging memory can match those exact specifications, it would not work on a DDR5/LPDDR5 bus, and so can’t be integrated).

In theory, the new CXL standard (ex: CXL 3.0) on PCIe5 or higher will open new opportunities, especially integrate the memory controller with the memory, not the CPU/chipset.

You just need a CPU that is compatible with PCIe5 / CXL (which in data-center should begin to appear in 2023/2024…), then on the PCIe5 bus, connect a card full of non-volatile memory, below 100ns R/W access NRAM (or MRAM) that have a CXL to NRAM (or CXL to MRAM) driver.

Read/Write access time to memory on a CXL bus (PCIe5 bus) will be a bit slower (maybe between 100ns to 1000ns), but still at least 10x to 100x faster than NAND flash SSD (10 000ns)

The key enabler is implementation of the CXL protocol as it will allow new emerging memory from startups (so not as good, neither as cheap as mainstream memory) to be integrated in computer systems : without CXL, they are gated / limited to the only memory types the CPU/chipset designer as integrated to its chip…

Hopefully, we may see a surge in R&D investments (by several startups or else) being done to size this new opportunity, and from there see new types of memory (non volatile memory) come to market.

The original Optane dimms have 250ns latency with the right format. The newest DDR5 Optane iteration seems to top out at 3 times the random bandwidth so latency seems lower.
Unless it is significantly cheaper than Optane it isn't doing anything Optane doesn't already do. And isn't profitable at.
Optane dimms even have software and hardware support to run as a hard drive, memory dimms, or both at the same time. Just look at the slide in the article. I doubt they will be as disruptive as Optane, which wasn't a lot. It just had the potential to change the file systems to the point that direct storage in games would be redundant. Everything would be direct storage. Dedicate a couple of e cores to data shuffling and your Optane/DDR mix would make ramdisks using traditional filesystems look slow and unresponsive.

The masses voted against that with their wallets.

 

[DavidM012]

Well I missed all that though I heard of Optane I thought it was just another brand name at first and then I read a few articles thought it looked interesting but Intel didn't go to great lengths to market it I thought or explain it's revolutionaryness and at the time I was seeing Optane ssds at £999 a few years ago and thought wts, no thanks.

If that translates into 'the masses voted against with their wallets' this sounds more like some sort of computer political game. If you really want to get a product out there there's plenty of scope to sell lots of it at an affordable price rather than a few of it at a high price and if they couldn't leverage economics of scale to actually get the product in systems, no wonder.

If they could just make something and price it by the cost of materials and labor rather than price the products by performance tiers maybe they'd get somewhere.

My reckoning would be that people want performance gains but they want to buy stuff at a sensible price. If you could replace an nvme ssd and dimms with an Optane based product at a competitive price maybe £150-200, people might go for it when they're building a pc. Upping the ante so high and charging £999 for 1 pci-e ssd is hopeless.

 

[USAFRet]

If that translates into 'the masses voted against with their wallets' this sounds more like some sort of computer political game. If you really want to get a product out there there's plenty of scope to sell lots of it at an affordable price rather than a few of it at a high price and if they couldn't leverage economics of scale to actually get the product in systems, no wonder.

No, just that it was:

1. Too expensive for any perceived user benefit.
2. Marketed badly.

This is/was beneficial to a datacenter.
Not everything that comes around is beneficial or cost effective to the consumer community.

If a 1TB solid state drive were 5 times as fast as whatever you use now, but cost $4,000....would you buy it for your game machine?


Price at market is not always just "Lets charge this!" There may well be an actual development cost, in that to create enough actual chips at the right quality does cost more.

 

[DavidM012]

So it simply couldn't be done practically which answers my questions about it, if it needs a miracle to cut production costs, whatever.

 

[DavidM012]

I would like my apps faster as well as games, though now I think of it the only reason we need terabytes of storage is because steam games are often 50-100gb a throw these days and might be accused of bloatware. If you could trust the industry was being as efficient as possible... Windows has grown to 30gb what are we doing with data?

Besides which disks aren't the only media by which data could be loaded since there's ethernet and and usb. If you would rather your game shipped on a usb 3.0 flash drive than download 100 gigabytes down a 4mb/s internet connection sure margins would get tighter and overheads would rise.

The problem with different approaches to moving data is yes ok the high end of the market can access 10gb/s somewhere, what about the rest of us if the infrastructure isn't there to actually support the latest and greatest technologies in practice it's all sci fi.

If the internet was really 'good enough' we might not even need disks and our pcs could really be thin clients, if data could be ported via ethernet/wifi at those kinds of speeds but it simply isn't that fast in reality for most users.

All these different data buses compete against each other and in the end it will be whichever wins out economically you might've thought.

Your PC & data could just be a virtual machine in a database in a datacentre somewhere if the remote connection was fast enough. They made a cpu out of optics didn't they? The technology could be anything that can make bits. And quantum computers are developing so some peoples are going to get to use supercomputers but it seems like humble consumers are getting the crumbs and cast offs.

 

[thestryker]

Intel had the cheaper (this is relative as even when it launched it was around 4x NAND price) models of their Optane drives in the 900/905p series. Due to the cost premium, lack of perceived difference and the SSD market pushing sequential as the end all these didn't break into consumer much. They did have almost the same endurance as the enterprise drives though so they caught on to some extent there, but this was mostly businesses saving money buying consumer over enterprise which is bad for Intel's business.

The way to drop the price was advancing the technology as from what I'd read their second revision of 3D XPoint managed to cut manufacturing prices almost in line with the increased density. The problem is that this means it'd be another 2 generations before they could get close enough to NAND to be a more broad alternative. So it's one of those things where Intel would have to keep developing it, bring a new fab online and then market the hell out of its storage capabilities. Virtually all of the Optane marketing had gone towards the Optane Memory (dead due to CXL and rapid increase in DRAM capacity) product so the storage side mostly went to those who needed exactly what it provided.

Intel is big enough that they could have kept it going, but they're a publicly traded company which hit bumps with their primary revenue source so investors have been screaming to cut as much as possible for years now. 3D XPoint could have ended up being a revolution for computing in the long term the same way SSDs were, but the benefit to the company was never going to be there any time soon. This and Gelsinger's stance on memory technology meant there wasn't an advocate for Optane so here we are.

 

[USAFRet]

I would like my apps faster as well as games,

Faster than what?
Currently, if I open a moderately complex Excel file, it takes under a second.

Your PC & data could just be a virtual machine in a database in a datacentre somewhere if the remote connection was fast enough. They made a cpu out of optics didn't they?

Physical distance. Physics.
You can't make a 100 mile connection be as fast as a 1 inch connection.

Some part of the system must be done locally.

Current games, for instance.
A lot of the brainpower lives in the game server. May be 1,000 miles away.
But a lot of it also exists on your desk. The conversation between your CPU, RAM, and GPU.

Also, consider the Forum you're looking at right now.
Hit New Posts, what happens?
You send a request to a database server on a different continent.
It retrieves what you requested, going through literally millions of rows of data.
It then sends back to you.
Your screen refreshes.

Signals back and forth across the planet, multiple times, to retrieve your particular little bit of data.

Maybe 0.6 seconds.

Your PC & data could just be a virtual machine in a database in a datacentre somewhere

And that already exists. OneDrive, GoogleDrive.
But a LOT of my personal data, I wish to keep right here. Instead of entrusting it to some faceless corporation.

 

[DavidM012]

If want to compress a steam game it takes all day. But I wouldn't need to if it wasn't bloatware or if storage shipped in petabytes for pennies.

There's already a bit of competition between nvidia and amd with video editing nvenc vs h.264 or whatever and people buying super fast desktop cpus and 'hi end' builds going in excess of $4000 dollars sometimes, well that's better than what they were selling $10000 dollar systems a few years ago which are now superseded by the next generation of cpu.

There was a moment of crazy I thought. When the last 3dfx voodoo went on ebay for $5000 and then someone recreated it for $200 including bugs - article I think there's something to be said about fools, money and being savvy in this industry. You can play old 3dfx titles with a free glide wrapper now. So that could translate as really just waiting 15-20 years to get the last generation of tech at what was the hi end at the time. Absurd but, now you can play homeworld: cataclysm with 3dfx enabled.

Yes there is a case for keeping some personal data but you when you need 30gb just to run the operating system there is a kind of baseline minimum requirement there and your pc can't be a true 'thin client' it's a heavily customised and modded client which you might argue should be done cheaper if electronics are going to go bad and can't be fixed by amateurs since people are already complaining about their $3000 dollar piles of junk even though the modularity might just mean in reality it's your $150 mobo that's a piece of junk it's still money and time and hassle to get a new mobo and then maybe the socket type has changed so you need memory and a cpu as well etc.

Whilst drives could last 5 years, 3 years or 10 years the warranty isn't 10 years anyway well you can only hope things might shape up with some concept of standards you can rely on.

 

[USAFRet]

If want to compress a steam game it takes all day. But I wouldn't need to if it wasn't bloatware or if storage shipped in petabytes for pennies.

That time is CPU/RAM, not drive speed.

 

[DavidM012]

but if your drive was super fast/hi endurance hybrid ram/storage device it might have a bit of an edge performance wise. The technology exists on paper but it's still sci fi - dreaming. And wouldn't need if it wasn't bloat ware so could really trust that data demands will grow rather than shrink if they could just code tighter instead?

now I think of it, I think it had been said that x86/x64 architecture wasn't the most efficient it just won out over risc for some reason in the mass market.

Some Apple engineer designed a risc chip that practically ran without power or something? Most of us simply don't understand enough about the tech to even know what to demand.

Sci fi would get you a portable wearable AI assistant with augmented reality or something that can time travel. Reality: Cortana was a windows service you couldn't turn off or elect not to load and suspected as spyware.

 

[rluker5]

Intel had the cheaper (this is relative as even when it launched it was around 4x NAND price) models of their Optane drives in the 900/905p series. Due to the cost premium, lack of perceived difference and the SSD market pushing sequential as the end all these didn't break into consumer much. They did have almost the same endurance as the enterprise drives though so they caught on to some extent there, but this was mostly businesses saving money buying consumer over enterprise which is bad for Intel's business.

The way to drop the price was advancing the technology as from what I'd read their second revision of 3D XPoint managed to cut manufacturing prices almost in line with the increased density. The problem is that this means it'd be another 2 generations before they could get close enough to NAND to be a more broad alternative. So it's one of those things where Intel would have to keep developing it, bring a new fab online and then market the hell out of its storage capabilities. Virtually all of the Optane marketing had gone towards the Optane Memory (dead due to CXL and rapid increase in DRAM capacity) product so the storage side mostly went to those who needed exactly what it provided.

Intel is big enough that they could have kept it going, but they're a publicly traded company which hit bumps with their primary revenue source so investors have been screaming to cut as much as possible for years now. 3D XPoint could have ended up being a revolution for computing in the long term the same way SSDs were, but the benefit to the company was never going to be there any time soon. This and Gelsinger's stance on memory technology meant there wasn't an advocate for Optane so here we are.

It doesn't sound too implausible to have had a high end gaming rig with 32GB HBM3 in package for the ram and 1-2TB of 75GB/s with near ram latency Optane dimms as your primary hard drive that basically acts like a giant pagefile with a small amount dedicated to an OS partition. As long as you don't think about the profits that would be lost charging professionals and corporations thousands extra for that.

Shame we will get M.2 ssds that need heatsinks with fans and no reason to upgrade the ram to cache instead.

My metallurgical background still has me a fan of glassy metals, and a product with a lifespan of quadrillions of atomic structure changes in infinitesimally small transition times is still pretty entertaining to have.

 

[DavidM012]

Yes you just need some substrate with tolerance to represent bits and away you go. They made an optical cpu but that didn't hit the mainstream - there's one in existence but it's sci fi.

 

[bit_user]

the modules will still use AES-256 encryption though, just like previous models.

I'll bet this is limiting bandwidth and partly responsible for the substantial power consumption. It's pretty dumb to integrate encription, when the Optane DIMMs are meant to be used with CPUs that already have a built-in memory encryption capability.