Interesting article. I wonder if the emphasis on comparing 3dxpoint to MLC/TLC flash was techinsight's idea or anandtech's.
From the original propaganda coming from Intel, I expected a memory that would barge in between DRAM and Flash, forcing DRAM to act more like a cache (presumably something like HBM connected to the CPU) and in turn caching slower, denser, and cheaper Flash (TLC or denser). This also appears to be what Intel thought with the shipping Optane, although that is pretty much moot after the introduction of pseudo-SLC 3d flash (no real caching needed, although it might make a good dram replacement in SSDs).
And of course the big questions remain: can Intel get the latency below milliseconds and get the endurance enough to use <1ms latency?
Seriously, I'm always surprised when tech types are disappointed in new technology. As it is, this provides a NJ get of major advantages, and it's just first generation. When did NAND first come out, 20 years ago? I still have a NAND card from my old Agfa minilab that cost US $250 and has 4MB storage (it was used to store settings in the machine.)
Give this a few years. I mean, the first usable SSD was just 16GB. The ones we saw back then as really usable were, by todays standards for SSD, very slow, and cost $3,600 for a 32GB drive.
Take a look at the Optane drive review from a few weeks ago and you'll see that latency is <10uS with 5-9s <100uS. The endurance is listed as 30 drive-writes per day.
Looks like the technology is delivering something new and exciting that fits between DRAM and NAND Flash.
Optane will probably never replace DRAM unless ridiculously drastic improvements are made since DRAM pretty much have a legendary endurance that optane probably could never match. As a caching scratch board it can make sense, but it's usefulness is quite limited especially for your average Joe.
You don't need it to completely replace DRAM. If the technology comes near to what they promised, then it can significantly reduce the need for DRAM. Right now the policy is to put much as possible into DRAM to mask slow speed of storage. When the storage is much faster, then it'll be different.
But final goal for these devices won't be realized for 10-15 years. You need to overhaul the computer system from top to bottom. It's something that has been stable for 70 years, so, you are talking about mentality changes of the people involved in the low level computer architecture.
Revolution, comes as a stage. Never instant. Now the days of easy scaling is at an end, that'll be true even in semiconductors where traditionally newer device always offered better in every category.
not really, as commercial Everspin Mram at 90nm is faster and can even now fit above the lower Dram for speed, then.. its also got a lot more nm room to grow and is already near infinite write Non-volatility
So they only need to stack 8 layers of this to leave the most advanced 3D NAND in the dust density/capacity-wise, without even stepping fien to a smaller process node. The only issue then would be cost per die - something that might well be resolved through higher volumes and better yields...
I can see this completely replacing NAND within the next 10 years or so...
3D XPoint doesn't scale nearly as well as 3D NAND when it comes to layers. Cross-point arrays are built layer by layer, whereas 3D NAND has a vertical channel i.e. all layers are processed at once. PCM is also not that scalable when it comes to lithography because it too suffers from cell-to-cell interference.
This is highly misleading , you focus on density when manufacturing is very different from NAND and DRAM. You do realize that most of your readers don't understand that cost is not just die size. Even you don't seem to understand the difference between costs and pricing. Cost isn't really a factor in how XPoint is priced today.
And to state the obvious, since the article fails to do so, XPoint competes with NAND as is lacks the specs to compete with DRAM. If you look at bit density vs Micron's 64L 256GB, 4.34 vs 0.62, that's 7 times. That only gets worse as the 3D NAND architecture is aimed at scaling on the vertical while XPoint isn't. WD/Toshiba likely have a more 3D NAND like architecture for their ReRAM, we'll see soon enough. As for DRAM, we'll see ,it might go 4F2, might go 3D , might even be replaced by a SCM.
It's not supposed to be directly competing with DRAM. That's something some of you posters are making up for yourselves. But it does look that it will compete favorably against DRAM. Performancewise, Intel has said that it won't equal performance, so we've know that from the beginning. But it comes far closer than NAND does, and so, for some purposes it could substitute for more RAM.
Density and die size are always important. This is also first generation. They're way down on the learning curve. I'd bet that next year, a new board will have 512GB and cost under $1,000. That would be about half the cost of today's board. I do t see anything other that larger sizes and lower costs in the future.
Will it ever surpass NAND in pricing? Likely it will. When, well we don't know that yet.
I've no doubt that the price is influenced by R&D costs, as pricing always is.
I'm excited for this technology, just not right now. Once it's economical and scalable, it could fundamentally revolutionize the way PC's function, altering (or combining) the roles of storage and RAM.
"My guess is that the companies have found a way to make the same materials behave in a slightly different way," he continued. "Chalcogenide glasses are the basis of numerous emerging memory technologies such as oxygen vacancy and silver dendrites, so they don’t necessarily have to use the phase-change mechanism to achieve bit storage."
Currently rocking a 32GB Optane as a very fast drive on my 4.7 Ghz Devil's Canyon that I use for running intensive Ghz-limited simulations. It's working brilliantly as a ludicrously fast pagefile, as I regularly bump into the memory limits of the machine. My code also uses a few GB of space to hand off large data to other workers, and the Optane is brilliant for that as wall.
Can't wait for the next generation, but this one's a giant leap above the M.2 SSD it replaced.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
22 Comments
Back to Article
wumpus - Saturday, May 27, 2017 - link
Interesting article. I wonder if the emphasis on comparing 3dxpoint to MLC/TLC flash was techinsight's idea or anandtech's.From the original propaganda coming from Intel, I expected a memory that would barge in between DRAM and Flash, forcing DRAM to act more like a cache (presumably something like HBM connected to the CPU) and in turn caching slower, denser, and cheaper Flash (TLC or denser). This also appears to be what Intel thought with the shipping Optane, although that is pretty much moot after the introduction of pseudo-SLC 3d flash (no real caching needed, although it might make a good dram replacement in SSDs).
And of course the big questions remain: can Intel get the latency below milliseconds and get the endurance enough to use <1ms latency?
vladx - Saturday, May 27, 2017 - link
It's only the first generation, I think Intel's hype numbers were for much later when the technology finally matures.melgross - Saturday, May 27, 2017 - link
Seriously, I'm always surprised when tech types are disappointed in new technology. As it is, this provides a NJ get of major advantages, and it's just first generation. When did NAND first come out, 20 years ago? I still have a NAND card from my old Agfa minilab that cost US $250 and has 4MB storage (it was used to store settings in the machine.)Give this a few years. I mean, the first usable SSD was just 16GB. The ones we saw back then as really usable were, by todays standards for SSD, very slow, and cost $3,600 for a 32GB drive.
dwrivers - Saturday, May 27, 2017 - link
Take a look at the Optane drive review from a few weeks ago and you'll see that latency is <10uS with 5-9s <100uS. The endurance is listed as 30 drive-writes per day.Looks like the technology is delivering something new and exciting that fits between DRAM and NAND Flash.
http://www.anandtech.com/show/11209/intel-optane-s...
hahmed330 - Saturday, May 27, 2017 - link
Optane will probably never replace DRAM unless ridiculously drastic improvements are made since DRAM pretty much have a legendary endurance that optane probably could never match. As a caching scratch board it can make sense, but it's usefulness is quite limited especially for your average Joe.IntelUser2000 - Sunday, May 28, 2017 - link
You don't need it to completely replace DRAM. If the technology comes near to what they promised, then it can significantly reduce the need for DRAM. Right now the policy is to put much as possible into DRAM to mask slow speed of storage. When the storage is much faster, then it'll be different.But final goal for these devices won't be realized for 10-15 years. You need to overhaul the computer system from top to bottom. It's something that has been stable for 70 years, so, you are talking about mentality changes of the people involved in the low level computer architecture.
Revolution, comes as a stage. Never instant. Now the days of easy scaling is at an end, that'll be true even in semiconductors where traditionally newer device always offered better in every category.
melgross - Sunday, May 28, 2017 - link
Intel has never claimed that optane will completely replace RAM. You're setting up a straw argument that was never made, and then knocking it downBMNify - Sunday, May 28, 2017 - link
not really, as commercial Everspin Mram at 90nm is faster and can even now fit above the lower Dram for speed, then.. its also got a lot more nm room to grow and is already near infinite write Non-volatilityboeush - Saturday, May 27, 2017 - link
So they only need to stack 8 layers of this to leave the most advanced 3D NAND in the dust density/capacity-wise, without even stepping fien to a smaller process node. The only issue then would be cost per die - something that might well be resolved through higher volumes and better yields...I can see this completely replacing NAND within the next 10 years or so...
Kristian Vättö - Monday, May 29, 2017 - link
3D XPoint doesn't scale nearly as well as 3D NAND when it comes to layers. Cross-point arrays are built layer by layer, whereas 3D NAND has a vertical channel i.e. all layers are processed at once. PCM is also not that scalable when it comes to lithography because it too suffers from cell-to-cell interference.jjj - Saturday, May 27, 2017 - link
This is highly misleading , you focus on density when manufacturing is very different from NAND and DRAM. You do realize that most of your readers don't understand that cost is not just die size. Even you don't seem to understand the difference between costs and pricing. Cost isn't really a factor in how XPoint is priced today.dwrivers - Saturday, May 27, 2017 - link
The article says that die size is a "strong predictor" of cost, which it certainly is.jjj - Saturday, May 27, 2017 - link
Not even close to that when you compare 2 very different things.jjj - Saturday, May 27, 2017 - link
And to state the obvious, since the article fails to do so, XPoint competes with NAND as is lacks the specs to compete with DRAM. If you look at bit density vs Micron's 64L 256GB, 4.34 vs 0.62, that's 7 times.That only gets worse as the 3D NAND architecture is aimed at scaling on the vertical while XPoint isn't. WD/Toshiba likely have a more 3D NAND like architecture for their ReRAM, we'll see soon enough.
As for DRAM, we'll see ,it might go 4F2, might go 3D , might even be replaced by a SCM.
melgross - Sunday, May 28, 2017 - link
It's not supposed to be directly competing with DRAM. That's something some of you posters are making up for yourselves. But it does look that it will compete favorably against DRAM. Performancewise, Intel has said that it won't equal performance, so we've know that from the beginning. But it comes far closer than NAND does, and so, for some purposes it could substitute for more RAM.melgross - Sunday, May 28, 2017 - link
Density and die size are always important. This is also first generation. They're way down on the learning curve. I'd bet that next year, a new board will have 512GB and cost under $1,000. That would be about half the cost of today's board. I do t see anything other that larger sizes and lower costs in the future.Will it ever surpass NAND in pricing? Likely it will. When, well we don't know that yet.
I've no doubt that the price is influenced by R&D costs, as pricing always is.
Samus - Saturday, May 27, 2017 - link
I'm excited for this technology, just not right now. Once it's economical and scalable, it could fundamentally revolutionize the way PC's function, altering (or combining) the roles of storage and RAM.hahmed330 - Saturday, May 27, 2017 - link
It probably never will endurance of RAM is almost infinite.Lolimaster - Sunday, May 28, 2017 - link
For consumer use the best application is to simply use it as a dedicated swap file drive + some programs installed on it. 16GB should do it.Swap inside an SSD wears the thing with constant (but small) writes on top of normal usage. When you hit an HDD based swap you feel the pain.
Optanes m.2 need to work on AMD mobos too.
Lolimaster - Sunday, May 28, 2017 - link
With the current DDR4 prices, people are mostly stuck at 8-16GB of total ram, 16GB of optane swap sounds nice.lefty2 - Sunday, May 28, 2017 - link
So, is it or isn't PCM?There's quite a funny article here: https://www.theregister.co.uk/2016/01/19/xpoint_in...
"My guess is that the companies have found a way to make the same materials behave in a slightly different way," he continued. "Chalcogenide glasses are the basis of numerous emerging memory technologies such as oxygen vacancy and silver dendrites, so they don’t necessarily have to use the phase-change mechanism to achieve bit storage."
rolfaalto - Monday, May 29, 2017 - link
Currently rocking a 32GB Optane as a very fast drive on my 4.7 Ghz Devil's Canyon that I use for running intensive Ghz-limited simulations. It's working brilliantly as a ludicrously fast pagefile, as I regularly bump into the memory limits of the machine. My code also uses a few GB of space to hand off large data to other workers, and the Optane is brilliant for that as wall.Can't wait for the next generation, but this one's a giant leap above the M.2 SSD it replaced.