[ThePlatform] Intel Shows Off 3d-Xpoint Memory Performance

[dealcorn]

At Oracle's OpenWorld conference, Intel revealed comparative Optane (SSD) performance on two unspecified Oracle application stacks. http://www.theplatform.net/2015/10/28/intel-shows-off-3d-xpoint-memory-performance/.

“This is a huge performance improvement,” Krzanich said. “I think all of us, for any kind of I/O operations, that could see a 5X to 8X improvement in speed – that’s what we have been looking for. It is an improvement in both performance and latency. But there is really more to this technology. Intel Optane SSDs provide about 200X less variability, and that is an additional benefit that you can count on for your datacenter.”

Re Optane DIMMs:

Krzanich said that Intel will have working Optane DIMMs ready later this year for early testers, and will combine the performance of DRAM with the capacity and cost of flash. What this means is that a mix of DDR4 and Optane DIMMs in a two-socket server with a total of 6 TB of addressable memory, “virtually eliminating paging between memory and storage, taking performance truly to a whole new level.” Krzanich added that data encryption in the DIMM, so that data at rest on the DIMM – it will take us all a while to get used to that – is secured.

 

[ShintaiDK]

Cant wait for 3DXPoint SSDs.

 

[MrTeal]

Cant wait for 3DXPoint SSDs.

It will be interesting to see how the XPoint SSDs slot in with current systems. The PCIe x4 link in current m.2 slots is good for ~4GB/s bandwidth. The P420m and P3700 come reasonably close to saturating that already, it's not hard to imagine a similar capacity XPoint needing an x16 link to really stretch its legs.

 

[LoveMachine]

I can understand how this is great for servers and datacenters with huge I/O demands. But at least in the short term, will this offer anything meaningful at the consumer level? It's slower than DDR ram and more expensive than SSDs, so at least until the costs come down to current SSD levels, will home users or standard office users see much benefit by going to 3DXPoint?

Would this potentially be something transformative for say phones and tablets? I can see how having a single block of non-volatile but fast storage would simplify small form factor devices.

 

[MajinCry]

I can understand how this is great for servers and datacenters with huge I/O demands. But at least in the short term, will this offer anything meaningful at the consumer level? It's slower than DDR ram and more expensive than SSDs, so at least until the costs come down to current SSD levels, will home users or standard office users see much benefit by going to 3DXPoint?

Would this potentially be something transformative for say phones and tablets? I can see how having a single block of non-volatile but fast storage would simplify small form factor devices.

We will be able to finally play Oblivion without any stutter.

 

[imported_ats]

I can understand how this is great for servers and datacenters with huge I/O demands. But at least in the short term, will this offer anything meaningful at the consumer level? It's slower than DDR ram and more expensive than SSDs, so at least until the costs come down to current SSD levels, will home users or standard office users see much benefit by going to 3DXPoint?

Would this potentially be something transformative for say phones and tablets? I can see how having a single block of non-volatile but fast storage would simplify small form factor devices.

Considering that in the majority of real world workloads there is almost no difference between a decent SATA3 SSD and a NVMe based SSD? probably not. The main advantage in the server space is the low latency on read based workloads allowing significantly higher IOPS at low queue depths.

They'll be some improvement for consumer workloads but it is highly unlikely that it will justify the costs initially.

 

[dealcorn]

Would this potentially be something transformative for say phones and tablets? I can see how having a single block of non-volatile but fast storage would simplify small form factor devices.

Communications from Intel have been mixed on how quickly mobile applications will be targeted. I recall that at announcement, phones were mentioned as one target market. The talking points then switched to data centers and high end gaming applications. Around the announcement of the Dalian investment, phones popped up again as a target market. Faster never hurts but is less important in today's phone applications. Initial phone adoption will be driven by (1) superior batter life because unlike DRAM, 3DXP is non volatile, (2) lower assembly costs because one 3DXP module (likely delivered on package with SoC) can replace the functions of both DRAM and flash NAND memory, and (3) reduced exposure to thermal management glitches. Prior to a substantial production ramp, it is unknown how the cost of 3DXP compares to the cost of separate memory and flash NAND modules. As the ramp commences, Intel may gain some insight regarding the slope of 3DXP's learning curve vs the slope of 3D NAND's learning curve. Cost matters.

Watch Broxton announcements. There is a 6 month lag between the release of Broxton for TV sticks and the general release of Broxton with Kaby Lake graphics. I do not think anyone has speculated that Broxton's SoC might support 3DXP DIMMs (as a test case prior to the release of Purley). If Broxton does not support 3DXP DIMMs, Intel sees no short term opportunity for 3DXP in phones. Leaked Broxton specification suggest it is not precluded from supporting 3DXP DIMMs, if chip set support was built in. It is not like Intel has 1,000 employees working to bring in a high end mobile customer and wants to use a 3DXP exclusive as a sweetener to make it happen.

Tablets have less pressing thermal issues than phones and can accept larger batteries. Tablets should be a low priority market.

 

[imported_ats]

Communications from Intel have been mixed on how quickly mobile applications will be targeted. I recall that at announcement, phones were mentioned as one target market. The talking points then switched to data centers and high end gaming applications. Around the announcement of the Dalian investment, phones popped up again as a target market. Faster never hurts but is less important in today's phone applications. Initial phone adoption will be driven by (1) superior batter life because unlike DRAM, 3DXP is non volatile, (2) lower assembly costs because one 3DXP module (likely delivered on package with SoC) can replace the functions of both DRAM and flash NAND memory, and (3) reduced exposure to thermal management glitches. Prior to a substantial production ramp, it is unknown how the cost of 3DXP compares to the cost of separate memory and flash NAND modules. As the ramp commences, Intel may gain some insight regarding the slope of 3DXP's learning curve vs the slope of 3D NAND's learning curve. Cost matters.

WRT (2), its highly unlikely that you'll see 3DXP replace RAM in smart phones. It isn't replacing ram in any other market FYI, it is augmenting it. In the case of servers, it basically a trade off between 1TB to 1.5TB per socket. Instead of adding an additional .5TB, you end up adding 4-8TB of 3DXP. It operates slower (likely 2-4x increase in latency and 2-3x effective decrease in bandwidth) but significantly reduces the need for backing store (aka disks/SSDs/etc) or in other cases enables enough capacity to take a disk based DB directly into memory. The base level of DRAM will still be present and be where the vast majority of work takes place, the 3DXP will be effectively "db cache" or "pagefile". AFAIK, no one is proposing to use the 3DXP as just more DRAM as it has significant disadvantages vs standard DRAM as far as performance: AKA there is a reason you see it on the pyramid below DRAM and not split with DRAM.

Unless people are ok with a significant performance hit, 3DXP won't be replacing DRAM in phones.

 

[ShintaiDK]

It will be interesting to see how the XPoint SSDs slot in with current systems. The PCIe x4 link in current m.2 slots is good for ~4GB/s bandwidth. The P420m and P3700 come reasonably close to saturating that already, it's not hard to imagine a similar capacity XPoint needing an x16 link to really stretch its legs.

The main benefit as I see it is IOPS and Latency. Not raw bandwidth.

A bit like the same with HDs to SSDs. It didn't really matter about sequel transfer. But the seek time.

Some 500GB 3DXpoint M.2 2260 and I jump

 

[ShintaiDK]

Considering that in the majority of real world workloads there is almost no difference between a decent SATA3 SSD and a NVMe based SSD? probably not. The main advantage in the server space is the low latency on read based workloads allowing significantly higher IOPS at low queue depths.

They'll be some improvement for consumer workloads but it is highly unlikely that it will justify the costs initially.

Its certainly not going to be any kind like HD to NAND SSD. But NVME is an example of the limitation of the current NAND based SSDs. So I do see consumer improvements, specially in the random I/O. But with that said, its more evolution and not night and day.

With the 1000x higher endurance its also getting awfully close to "The last SSD you ever need" if we look beyond size.

 

[dealcorn]

WRT (2), its highly unlikely that you'll see 3DXP replace RAM in smart phones....

Unless people are ok with a significant performance hit, 3DXP won't be replacing DRAM in phones.

Your point is well taken and your knowledge of 3DXP performance characteristics is greater than mine, but Intel has patented this exact configuration. See the Purple Swan article: http://seekingalpha.com/article/3253655-intel-and-micron-the-purple-swan.

Dynamic partial power down of memory-side cache in a 2-level memory hierarchy

• The focus of this approach is on providing performance with a relatively small amount of a relatively higher-speed memory such as DRAM while implementing the bulk of the system memory using significantly cheaper and denser non-volatile random access memory (NVRAM).

One embodiment of a tablet computer or cellular telephony device is configured with PCMS memory but potentially no near memory and no additional mass storage (for cost/power savings). However, the tablet/telephone may be configured with a removable mass storage device such as a flash or PCMS memory stick.

I do not dispute you assessment of 3DXP performance characteristics, but what is your source? I recall but can not now locate a discussion that suggested a sliver of DRAM cache could substantially mask the unknown performance detriment associated with 3DXP. If accurate, this approach may open the high end phone niche to 3DXP.

EDIT: From the comments to the above cited article:

Author’s reply » The point is not that phase change memory is slower but rather that eDRAM-cached PCM (with the associated Intel and Micron patents to compensate) is just as fast as pure DRAM with lower cost and power.

It is just early to eliminate high end phones from the target markets.

 

[Khato]

It will be interesting to see how the XPoint SSDs slot in with current systems. The PCIe x4 link in current m.2 slots is good for ~4GB/s bandwidth. The P420m and P3700 come reasonably close to saturating that already, it's not hard to imagine a similar capacity XPoint needing an x16 link to really stretch its legs.

I wouldn't say that. I'd be quite happy with a 3D XPoint SSD which could maintain 1GB/s regardless of transfer size. 'Cause sure an NVMe 950 pro can hit 1GB/s write and 2GB/s read sequential, but on 4KB random those figures drop to 250 MB/s write and 125 MB/s read.

Besides, what we should be looking forward to is having our 3D XPoint storage directly connected to the CPU just like memory no? Since PCI-E is too slow both in terms of bandwidth and latency.

 

[jpiniero]

Intel is still not talking about active power consumption, which is a bad sign. Seems that the SSDs will be fast but not amazingly faster... you would be bottlenecked by the storage controller. You'd need the DIMM version for it to really shine.

Besides, what we should be looking forward to is having our 3D XPoint storage directly connected to the CPU just like memory no? Since PCI-E is too slow both in terms of bandwidth and latency.

The endgame is onpackage. It's gonna be awhile before we get there though.

 

[PCRider]

Most of the posters in this thread are missing the big picture. Imagine having a system that has 4 Tb of 3d memory that boots from the memory, because the memory is also the primary storage device. Latency is such a big deal due only to the current system architecture needing to load code from storage into memory to execute. Remove that lag time and you get a big boostin performance that will more than make up for the speed drop in DRAM to 3d. This will take a few years to be fine tuned, but OS's will need to be recoded to be able to understand and take full advantage of memory data that doubles as primary storage. When that happens lookout because the speed jump is going to push processor development hard to speed up to max out the capabilies of this change.

Add in the additional factors that limit DRAM gaining speed from further hardware development and that 3d has astronomical room for more speed throught continued hardware development, and 3d is going to crush DRAM in the long run. The only question is when 3d will over take DRAM in standard performance because it is already cheaper to manufacture.

3d also has some key advantages in operating cost that no one seems to be pointing out. 3d doesn't need the power cycle to maintain data at the transistors. That means the memory locations don't get powered unless you use them. This can, depending heavily on system demand, reduce power usage, heat generation, and allow for quicker memory response over DRAM. The biggest challenges to data centers is expanding power for facilities and managing the heat systems generate at large scales. The another is system recovery after power loss. Systems won't need to reboot after power loss. They'll simply pickup exactly were they left off and perform error checks to prevent data corruptions. No more rebuilding virtual systems from backups on ESX hosts and minimizing downtime.

Computer forensics will also go through a huge change. Since the memory in non-volatile, you can pull the power and you have a possible cyber attack frozen in mid process cycle. After some evolution in legal evidence rules, this will allow for much easier evidence gathering of a suspected cyber attack.

I have no doubt that 3d xPoint memory will drive major changes for the computer industry for years to come.

 

[TechGod123]

Most of the posters in this thread are missing the big picture. Imagine having a system that has 4 Tb of 3d memory that boots from the memory, because the memory is also the primary storage device. Latency is such a big deal due only to the current system architecture needing to load code from storage into memory to execute. Remove that lag time and you get a big boostin performance that will more than make up for the speed drop in DRAM to 3d. This will take a few years to be fine tuned, but OS's will need to be recoded to be able to understand and take full advantage of memory data that doubles as primary storage. When that happens lookout because the speed jump is going to push processor development hard to speed up to max out the capabilies of this change.

Add in the additional factors that limit DRAM gaining speed from further hardware development and that 3d has astronomical room for more speed throught continued hardware development, and 3d is going to crush DRAM in the long run. The only question is when 3d will over take DRAM in standard performance because it is already cheaper to manufacture.

3d also has some key advantages in operating cost that no one seems to be pointing out. 3d doesn't need the power cycle to maintain data at the transistors. That means the memory locations don't get powered unless you use them. This can, depending heavily on system demand, reduce power usage, heat generation, and allow for quicker memory response over DRAM. The biggest challenges to data centers is expanding power for facilities and managing the heat systems generate at large scales. The another is system recovery after power loss. Systems won't need to reboot after power loss. They'll simply pickup exactly were they left off and perform error checks to prevent data corruptions. No more rebuilding virtual systems from backups on ESX hosts and minimizing downtime.

Computer forensics will also go through a huge change. Since the memory in non-volatile, you can pull the power and you have a possible cyber attack frozen in mid process cycle. After some evolution in legal evidence rules, this will allow for much easier evidence gathering of a suspected cyber attack.

I have no doubt that 3d xPoint memory will drive major changes for the computer industry for years to come.

If all this pans out, then we have very exciting times up ahead indeed.