I remember it like yesterday, when HP announced Memristor based DRAM replacement ready in 2016 with 1000x capacity over DRAM, same speed yet true non-volatility four years ago. 1000x Capacity was to be done via layering (linear cost) and energy not an issue because only access would involve it, not storage.
My immediate reaction was that HP might die for the business such a product would kill, before they’d be able to put it in the market: HP was still pushing a lot of storage in those days. The other instinctive reaction was that compute wasn’t going to be anyway near ready with 1000x random access memory capacity and that it would have to move into the RAM itself to have any chance of getting value out of the technology.
I also very distinctly remember Martin Fink, Meg Whitman, Brian Krzanich and Satya Nadella sitting around a table discussing how the Memristor was going to change the world. It was quite obvious right there that Intel’s enthusiasm didn’t match that of HP nor mine.
When they announced their own NV-RAM technology some time later, the first thing that was obvious was that they wanted it to fit right between low-end DRAM and high-end Flash, which would have been just perfect from a business point of view. But the Memristor at the same time proved, that brand new base technology doesn’t just grow into its assigned niche, but can step on dozens of feet right and left: The memristor seemed destined to roll up all downward storage tiers both in price and performance and kick right up into logic, where it could outperform silicon gate-by-gate using a more powerful form of logic.
Well Optane so far failed to fill its assigned niche and the Memristor failed to leave the labs, but somehow I missed the story behind: What happened to the Memristor? Is Optane/X-point in as much trouble as the 10nm process?
Probably not. They are on the verge of Gen 2 and have a dedicated Fab. You don't dedicate a multi-billion dollar Fab if there is no foreseeable path for improvement. You can also buy Optane right now.
What you need to remember about HP (and this is the reason I love HP...in fact, the ONLY reason) is they are an engineering house not that different from IBM.
HP's focus on engineering is completely different than IBM though, where the later is focused on patents, and former is focused on concept.
HP is an incredible place to work for this. When you hear about the famous Meg Whitman\Mark Hurd layoffs though the last two decades, it was literally trimming the fat. Engineers aren't the fat. Marketing, sales, service\support, and other low-level positions are the fat. Sure engineers makes substantially more than someone in service, but there are far less engineers consuming the budget, and engineers, in HP's (and my) mind are not expendable. The company has excellent retention and values their real employees.
Case in point, look at Compaq. NOT A SINGLE ENGINEER WAS LET GO during that merger. HP wanted those guys, and a lot of them are still there. Even the Itanium guys have been shifted to new projects, and a lot of them actually worked on Memristor going back nearly two decades when HP was looking to make faster flash storage and higher density DRAM.
Maybe if IBM wasn't so obsessed with Quantum computing, they would have had a working NV storage solution by now, but that doesn't appear to be their focus.
HP hasn't been an engineering house since Carly took over and fired all the engineers in favor of selling cheap inkjets with super-expensive ink. Witman just continued that and HP's been in a death spiral ever since.
The issue why we are seeing this (and also not seeing Micron's own QuantX) is because Micron does not yet own the technology. There was an event last week that caused Intel to breakup with Micron. Micron has done a great job of keeping this a secret: https://www.scribd.com/document/383955630/54308736...
First: Not sure when your memory is from, but I doubt they had to contend with pseudo-SLC 3d NAND flash. That's basically eliminated the niche between NAND and DRAM, at least as far as SSD caching goes. 3Dxpoint would have taken off a lot better without that tech, although I'd love to see the stuff cache the NAND drives as sort of a nonvolitile memory buffer (assuming the optane was close enough to the CPU to act as memory. See problems with that below).
Second: Intel has real competition now, but is still clearly on top. 3Dxpoint may well be a fairly disruptive tech, assuming it can begin to replace DRAM. If the cost of populating a server with memory drops by half or more, you don't have to buy the absolute highest performing CPU for it at a 10x premium. An EPYC or possibly an ARM server will make a ton of sense.
I'm fairly sure I've seen reports of less than honest OEMs providing 3dxpoint and calling it "memory", and this would even more or less work (say 80% performance of DRAM) assuming you filled all the channels with DRAM (this doesn't seem possible for said OEMs). The big gamble in doing this "for real" is that you are either stuck with using the DRAM as "real memory" and the optane as "virtual memory" or you have to fully build a proper memory system with a huge set of tags for all your DRAM and fill it like cache (which is certainly not what we are seeing from Xeon motherboards). Those tags are *expensive*, and stuffing them into DRAM means that you have to have built the DRAM interface to use it as such (your memory gets all offset). You also have lots of issues that you can't use multi-way caches with tags in the DRAM.
I think nearly all of Micron's grief comes from Intel's two strategies: 1. Enterprise. Actually this one makes sense. There must be plenty of systems with databases in SSDs for the simple reason that they couldn't find a system that would hold that much memory (and it would be a tad expensive). I'd expect Optane to sell in plenty of SSD-limited systems. 2. Consumer. This unfortunately is a joke. Accelerating rotating hard drives with faster nonvolitile stuff made sense 10-20 years ago (and Intel's indifference to their old SSD-based system provided on motherboards shows just how important this is) but is mostly a joke now. I may like AMD's system, but that is based on the idea that you can get your entire working set of storage on SSD and keep the rarely used stuff on rotating storage.
I'd love to see 2019 computers [i.e. 10nm Intel and "7nm" AMD] built around 3dxpoint memory. I'm not holding my breath (except for high end Xeons that either require manual partitioning or simply have a faster means to send 4k pages to DRAM. And expect to pay a stiff premium for that).
1. Intel will stop 3D NAND production after third gen (96-Layers) because they are losing money due to scale of production (unlike Micron). Micron under Sanjay doesn't want to give it's technology to a competitor (yes, Intel is paying Micron for NAND development until third gen is finished). 2. Intel only has one big fab for NAND and that's the china one. Intel is going to convert that to 3D Xpoint. 3. Micron has the right to buy Intel's stake in IMFT for a very low price beginning 2019 (it's in every 10-Q and 10-K). They are probably going to buy it because Micron's timeline for 3DXpoint products hints at it. Micron is going to finish it's product development in second half 19 (just when 2nd gen XPoint is ready and Micron doesn't have to share development and production with Intel) and products will come to market in early 2020. 4. Intel wants very high speed 3DX to replace or expand very large RAM based databases but it's not clear what Micron wants (maybe embedded stuff like phase change on MCUs; look at STMicro).
I'm pretty sure Micron and Intel both want 3D XPoint in the same spaces. Micron wants to go after the data center just like Intel. Micron has been moving more aggressively into the data center with their NAND recently.
As far as why the dissolution of the 3D XPoint partnership, it could just be because both Micron and Intel feel comfortable with the technology now and since they are the only two companies with access to it they will be competitors in that space.
I wouldn't be surprised if you were right about Intel's NAND business, though. Micron and Intel announced the dissolution of their NAND collaboration earlier this year. Chinese companies are going to start moving into the NAND space and Samsung is increasing capacity. Intel might not find the space to their liking any more and a pivot to 3D XPoint would make sense.
Yeah, but Micron probably won't get access to Intel's proprietary DIMM interfaces for Optane (non-Jedec). So data center for Micron can only mean PCIe or U.2 but that's a tough market...
I'd rather see XPoint as cache in UFS packages and SSDs replacing DRAM and pseudo-SLC mode.
I don't see why Micron would enter into a joint venture with Intel and develop a memory technology with a roadmap for two methods of deployment, both DIMMs and SSDs, without gaining access to the IP necessary for both deployments. Just because Micron has thus far chosen not to release any 3D XPoint products doesn't mean they aren't fully capable of doing so. I'd be willing to bet that Micron will be able to market 3D XPoint DIMMs on both Intel and non-Intel platforms (such as OpenPower).
More likely Intel will gladly give access to such DIMMS in return for promises to never sell anything but "Intel only" memory or current PCIe access. The last thing they want is Micron publishing a DDRish spec that AMD and ARM could use.
If Intel is trying to use a server CPU monopoly to extend to a server memory monopoly it could easily backfire. On the other hand they could easily sacrifice half the potential server memory market to maintain their server CPU market.
I'd be very surprised if Micron doesn't have the power to bring 3D XPoint fully to non-Intel platforms. Micron is a founding member of the OpenPower Consortium and surely would want to sell 3D XPoint in OpenPower servers and supercomputers.
Intel does not have a server CPU monopoly, anyway. They simply have a dominant market share. There's a difference. I really doubt Micron entered into the joint venture in a way that severely limited its ability to profit from what it was developing. If Intel loses market share, a possibility Micron would have been astutely aware of, Micron would have wanted to have the ability to market 3D XPoint to the non-Intel market as well.
Intel absolutely doesn't want Xpoint to replace DRAM (which is what it would be doing in the datacenter). The price of RAM in those servers is huge, which means you pay for the most powerful server you can buy and you don't look at the price tag (until it gets higher than the DRAM). This allows Intel to jack their prices sky-high as nobody has competed with them for a long time. Now they not only need to fear EPYC, but suddenly the performance bar that ARM needs to clear to make a "server worthy" CPU for "optane memory" is far lower (and ARM servers are showing up which have some serious performance).
Intel is trying to figure out how to put the genie back in the bottle (they may have to buy Micron out. Or maybe the whole thing won't work after all.)
Exactly. Early NAND wasnt very good either. If they can double the performance and halve the cost in one generation, we will have something VERY interesting on our hands....
No way SCM can replace DRAM. It can just complement. Maybe MCS can do but only time will tell. When QuantX, Optane news came up, all and sundry expect a lot. But it falters. MRA, ReRam, PCM may look to be more at an embryonic stage, yet to be hatched into a sea of monstrous competitions.
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.
Facebook
Twitter
RSS
19 Comments
Back to Article
abufrejoval - Monday, July 16, 2018 - link
I remember it like yesterday, when HP announced Memristor based DRAM replacement ready in 2016 with 1000x capacity over DRAM, same speed yet true non-volatility four years ago. 1000x Capacity was to be done via layering (linear cost) and energy not an issue because only access would involve it, not storage.My immediate reaction was that HP might die for the business such a product would kill, before they’d be able to put it in the market: HP was still pushing a lot of storage in those days. The other instinctive reaction was that compute wasn’t going to be anyway near ready with 1000x random access memory capacity and that it would have to move into the RAM itself to have any chance of getting value out of the technology.
I also very distinctly remember Martin Fink, Meg Whitman, Brian Krzanich and Satya Nadella sitting around a table discussing how the Memristor was going to change the world. It was quite obvious right there that Intel’s enthusiasm didn’t match that of HP nor mine.
When they announced their own NV-RAM technology some time later, the first thing that was obvious was that they wanted it to fit right between low-end DRAM and high-end Flash, which would have been just perfect from a business point of view. But the Memristor at the same time proved, that brand new base technology doesn’t just grow into its assigned niche, but can step on dozens of feet right and left: The memristor seemed destined to roll up all downward storage tiers both in price and performance and kick right up into logic, where it could outperform silicon gate-by-gate using a more powerful form of logic.
Well Optane so far failed to fill its assigned niche and the Memristor failed to leave the labs, but somehow I missed the story behind: What happened to the Memristor? Is Optane/X-point in as much trouble as the 10nm process?
Anything, please?
Adramtech - Monday, July 16, 2018 - link
Probably not. They are on the verge of Gen 2 and have a dedicated Fab. You don't dedicate a multi-billion dollar Fab if there is no foreseeable path for improvement. You can also buy Optane right now.Samus - Tuesday, July 17, 2018 - link
What you need to remember about HP (and this is the reason I love HP...in fact, the ONLY reason) is they are an engineering house not that different from IBM.HP's focus on engineering is completely different than IBM though, where the later is focused on patents, and former is focused on concept.
HP is an incredible place to work for this. When you hear about the famous Meg Whitman\Mark Hurd layoffs though the last two decades, it was literally trimming the fat. Engineers aren't the fat. Marketing, sales, service\support, and other low-level positions are the fat. Sure engineers makes substantially more than someone in service, but there are far less engineers consuming the budget, and engineers, in HP's (and my) mind are not expendable. The company has excellent retention and values their real employees.
Case in point, look at Compaq. NOT A SINGLE ENGINEER WAS LET GO during that merger. HP wanted those guys, and a lot of them are still there. Even the Itanium guys have been shifted to new projects, and a lot of them actually worked on Memristor going back nearly two decades when HP was looking to make faster flash storage and higher density DRAM.
Maybe if IBM wasn't so obsessed with Quantum computing, they would have had a working NV storage solution by now, but that doesn't appear to be their focus.
close - Tuesday, July 17, 2018 - link
Sounds like a "full disclosure" might be needed :).rahvin - Tuesday, July 17, 2018 - link
HP hasn't been an engineering house since Carly took over and fired all the engineers in favor of selling cheap inkjets with super-expensive ink. Witman just continued that and HP's been in a death spiral ever since.Rocket321 - Thursday, July 19, 2018 - link
I will never forgive Carly for signing the death warrant of Palm, just a few months after Mark bought them.lincoln.revere - Tuesday, July 17, 2018 - link
The issue why we are seeing this (and also not seeing Micron's own QuantX) is because Micron does not yet own the technology. There was an event last week that caused Intel to breakup with Micron. Micron has done a great job of keeping this a secret: https://www.scribd.com/document/383955630/54308736...wumpus - Wednesday, July 18, 2018 - link
First: Not sure when your memory is from, but I doubt they had to contend with pseudo-SLC 3d NAND flash. That's basically eliminated the niche between NAND and DRAM, at least as far as SSD caching goes. 3Dxpoint would have taken off a lot better without that tech, although I'd love to see the stuff cache the NAND drives as sort of a nonvolitile memory buffer (assuming the optane was close enough to the CPU to act as memory. See problems with that below).Second: Intel has real competition now, but is still clearly on top. 3Dxpoint may well be a fairly disruptive tech, assuming it can begin to replace DRAM. If the cost of populating a server with memory drops by half or more, you don't have to buy the absolute highest performing CPU for it at a 10x premium. An EPYC or possibly an ARM server will make a ton of sense.
I'm fairly sure I've seen reports of less than honest OEMs providing 3dxpoint and calling it "memory", and this would even more or less work (say 80% performance of DRAM) assuming you filled all the channels with DRAM (this doesn't seem possible for said OEMs). The big gamble in doing this "for real" is that you are either stuck with using the DRAM as "real memory" and the optane as "virtual memory" or you have to fully build a proper memory system with a huge set of tags for all your DRAM and fill it like cache (which is certainly not what we are seeing from Xeon motherboards). Those tags are *expensive*, and stuffing them into DRAM means that you have to have built the DRAM interface to use it as such (your memory gets all offset). You also have lots of issues that you can't use multi-way caches with tags in the DRAM.
I think nearly all of Micron's grief comes from Intel's two strategies:
1. Enterprise. Actually this one makes sense. There must be plenty of systems with databases in SSDs for the simple reason that they couldn't find a system that would hold that much memory (and it would be a tad expensive). I'd expect Optane to sell in plenty of SSD-limited systems.
2. Consumer. This unfortunately is a joke. Accelerating rotating hard drives with faster nonvolitile stuff made sense 10-20 years ago (and Intel's indifference to their old SSD-based system provided on motherboards shows just how important this is) but is mostly a joke now. I may like AMD's system, but that is based on the idea that you can get your entire working set of storage on SSD and keep the rarely used stuff on rotating storage.
I'd love to see 2019 computers [i.e. 10nm Intel and "7nm" AMD] built around 3dxpoint memory. I'm not holding my breath (except for high end Xeons that either require manual partitioning or simply have a faster means to send 4k pages to DRAM. And expect to pay a stiff premium for that).
brakdoo - Tuesday, July 17, 2018 - link
It's obvious what's going on:1. Intel will stop 3D NAND production after third gen (96-Layers) because they are losing money due to scale of production (unlike Micron). Micron under Sanjay doesn't want to give it's technology to a competitor (yes, Intel is paying Micron for NAND development until third gen is finished).
2. Intel only has one big fab for NAND and that's the china one. Intel is going to convert that to 3D Xpoint.
3. Micron has the right to buy Intel's stake in IMFT for a very low price beginning 2019 (it's in every 10-Q and 10-K). They are probably going to buy it because Micron's timeline for 3DXpoint products hints at it. Micron is going to finish it's product development in second half 19 (just when 2nd gen XPoint is ready and Micron doesn't have to share development and production with Intel) and products will come to market in early 2020.
4. Intel wants very high speed 3DX to replace or expand very large RAM based databases but it's not clear what Micron wants (maybe embedded stuff like phase change on MCUs; look at STMicro).
Yojimbo - Tuesday, July 17, 2018 - link
I'm pretty sure Micron and Intel both want 3D XPoint in the same spaces. Micron wants to go after the data center just like Intel. Micron has been moving more aggressively into the data center with their NAND recently.As far as why the dissolution of the 3D XPoint partnership, it could just be because both Micron and Intel feel comfortable with the technology now and since they are the only two companies with access to it they will be competitors in that space.
I wouldn't be surprised if you were right about Intel's NAND business, though. Micron and Intel announced the dissolution of their NAND collaboration earlier this year. Chinese companies are going to start moving into the NAND space and Samsung is increasing capacity. Intel might not find the space to their liking any more and a pivot to 3D XPoint would make sense.
brakdoo - Tuesday, July 17, 2018 - link
Yeah, but Micron probably won't get access to Intel's proprietary DIMM interfaces for Optane (non-Jedec). So data center for Micron can only mean PCIe or U.2 but that's a tough market...I'd rather see XPoint as cache in UFS packages and SSDs replacing DRAM and pseudo-SLC mode.
Yojimbo - Tuesday, July 17, 2018 - link
I don't see why Micron would enter into a joint venture with Intel and develop a memory technology with a roadmap for two methods of deployment, both DIMMs and SSDs, without gaining access to the IP necessary for both deployments. Just because Micron has thus far chosen not to release any 3D XPoint products doesn't mean they aren't fully capable of doing so. I'd be willing to bet that Micron will be able to market 3D XPoint DIMMs on both Intel and non-Intel platforms (such as OpenPower).wumpus - Wednesday, July 18, 2018 - link
More likely Intel will gladly give access to such DIMMS in return for promises to never sell anything but "Intel only" memory or current PCIe access. The last thing they want is Micron publishing a DDRish spec that AMD and ARM could use.If Intel is trying to use a server CPU monopoly to extend to a server memory monopoly it could easily backfire. On the other hand they could easily sacrifice half the potential server memory market to maintain their server CPU market.
"A bird in the hand is worth two in the bush"
Yojimbo - Wednesday, July 18, 2018 - link
I'd be very surprised if Micron doesn't have the power to bring 3D XPoint fully to non-Intel platforms. Micron is a founding member of the OpenPower Consortium and surely would want to sell 3D XPoint in OpenPower servers and supercomputers.Intel does not have a server CPU monopoly, anyway. They simply have a dominant market share. There's a difference. I really doubt Micron entered into the joint venture in a way that severely limited its ability to profit from what it was developing. If Intel loses market share, a possibility Micron would have been astutely aware of, Micron would have wanted to have the ability to market 3D XPoint to the non-Intel market as well.
wumpus - Wednesday, July 18, 2018 - link
Intel absolutely doesn't want Xpoint to replace DRAM (which is what it would be doing in the datacenter). The price of RAM in those servers is huge, which means you pay for the most powerful server you can buy and you don't look at the price tag (until it gets higher than the DRAM). This allows Intel to jack their prices sky-high as nobody has competed with them for a long time. Now they not only need to fear EPYC, but suddenly the performance bar that ARM needs to clear to make a "server worthy" CPU for "optane memory" is far lower (and ARM servers are showing up which have some serious performance).Intel is trying to figure out how to put the genie back in the bottle (they may have to buy Micron out. Or maybe the whole thing won't work after all.)
iwod - Tuesday, July 17, 2018 - link
So what exactly is new in the 2nd Gen Optane? Speed ? Power? Price?brakdoo - Tuesday, July 17, 2018 - link
More than two (mirrored) layers. It was supposed to go much higher than now to bring costs significantly down.danielfranklin - Wednesday, July 18, 2018 - link
Exactly.Early NAND wasnt very good either.
If they can double the performance and halve the cost in one generation, we will have something VERY interesting on our hands....
[email protected] - Wednesday, July 25, 2018 - link
No way SCM can replace DRAM. It can just complement. Maybe MCS can do but only time will tell. When QuantX, Optane news came up, all and sundry expect a lot. But it falters. MRA, ReRam, PCM may look to be more at an embryonic stage, yet to be hatched into a sea of monstrous competitions.