Welcome to the era of unforgettable memories, courtesy of nonvolatile DIMMs and SCMs
Increasing processor power and the rise of IO-heavy applications such as analytics are driving demand for higher capacity and cheaper solid-state memory. But DRAM has lagged behind Moore's law, and there are major challenges ahead in reducing its cost and boosting its capacity. Although new types of memory, such as Intel's and Micron's jointly developed 3D XPoint, are waiting in the wings, they will not replace DRAM for some while yet. That is mostly because, although they outperform flash, they cannot match the performance of DRAM.
However, the emerging memories have advantages over DRAM in other areas. They use less power, and they're set to be cheaper and to offer greater data capacities. Importantly, they're persistent or nonvolatile, unlike DRAM. That is why the collective label for the emerging memories is 'storage-class memories' (SCMs). Until more recently, 'memory' has mostly meant volatile DRAM, while 'storage' has meant nonvolatile disk. DRAM's volatility has been a major drawback because it requires DRAM contents to be protected against power loss, and that regularly involves extra IO steps writing data into safe, nonvolatile storage. Those steps reduce overall server and storage system performance.
Both software and hardware will need to be changed to fully exploit SCMs. On the hardware front, the new memories will need to use the DDR buses in server DIMM memory slots, to make the most of their performance. The devices that do that will be called nonvolatile DIMMs (NVDIMMs.) In some cases, they will combine DRAM and an SCM in the same device. Standards to cover this activity are being created by JEDEC, the microelectronics body that standardized the DDR interface inside DIMM slots. We will cover the software changes needed to fully exploit SCMs in a separate report.
The 451 Take
It's hard to exaggerate the coming impact of new memories. In Intel's words, they will drive 'mega' changes in system architecture, and we think that is a perfectly reasonable word to use. But the process will take time, and alongside the transformation to system architectures, application and OS software will also need to change. Setting standards for the way SCMs can be used in DIMM slots shows foresight, and will smooth the transition to SCMs. More importantly, perhaps, it also shows JEDEC's confidence that, despite false starts and delays in development, new memories are at last poised to enter mainstream IT. As an important side note, JEDEC's work is not just about new memories – the industry still has not finished exploring ways of using flash, which, by some definitions, also deserves to be called an SCM.
While NVDIMMs are set to become a major part of the future landscape for servers and storage, the term has been in existence for many years. The first NVDIMMs appeared around 10 years ago, and comprised DRAM DIMM modules that had been made nonvolatile by using flash inside the modules as a backup medium. Since then, such devices have established a small foothold in the market, mostly within storage systems.
Until now, the devices have been called NVDIMMs, but because other types of NVDIMM have already appeared, and others are on the horizon, new nomenclature is needed. In 2016 JEDEC published a standard covering this long-existing type of NVDIMM, which it relabeled as an NVDIMM-N. More standards describing different types of NVDIMM and using different suffixes are in the pipeline (see below).
In normal operation, NVDIMM-Ns store data in DRAM in the same way as conventional DIMMs. When power is lost, the contents of the volatile DRAM are copied into safety in nonvolatile flash, using energy from a backup capacitor or battery. The flash is only used when there is a power loss, and in normal operation the NVDIMM-N shows the same performance as standard DRAM-powered DIMMs.
Anticipating more use of NVDIMM-Ns
NVDIMM-Ns cost two to three times as much per GB as equivalent DIMMs, and their capacities have so far been modest – although they are increasing. This year Micron will begin shipping 16GB and 32GB NVDIMM-Ns, which are the same capacities as the two most commonly used types of conventional DIMMs. Other NVDIMM-N suppliers include AgigA Tech, Netlist, PNY Technologies, SMART Modular Technologies and Viking Technology.
NVDIMM-Ns eliminate the need to write data to flash or disk during normal operations or IO flow, for protection against power loss. Instead, they only write data to flash if or when power has been lost. That can significantly boost performance. For example, Netlist says it has demonstrated its devices increasing performance fivefold for a MySQL database, when compared with using standard DIMMs and writing data to PCIe flash drives to protect it against power loss.
Unlike NVDIMMs powered by future SCMs, NVDIMM-Ns do not increase memory capacity. But they run at the speed of DRAM, unlike future SCM-powered NVDIMMs will. For that reason, some vendors are predicting not only that NVDIMM-Ns will coexist with other, SCM-powered NVDIMMs, but that usage of NVDIMM-Ns will increase. JEDEC's work creating the NVDIMM-N standard shows that it shares that view.
In 2016 Hewlett Packard Enterprise became the first major server maker to offer NVDIMM-Ns under its own brand. HPE's devices are being OEMed from Micron, and are offered by HPE as an option for its new Proliant Gen9 servers. Micron estimates that NVDIMM-Ns are currently fitted to a low single-digit percentage of servers, but believes that could grow to beyond 10% over the next two to three years. HPE's sizeable share of the server market no doubt factored into this forecast. Among other server makers, some Lenovo servers may support NVDIMM-Ns, while Dell and Cisco say none of their servers do so – although this could change if customer demand develops.
The NVDIMM-N standards created by JEDEC cover the physical connections between NVDIMM-Ns and DIMM slots. They also cover the system software interface, eliminating the need for customers to install proprietary NVDIMM drivers and BIOSs. JEDEC-compliant NVDIMM-N drivers are now part of the Linux kernel, and are in Windows Server 2016. Although the NVDIMM-N standard currently only covers the use of flash with DRAM, JEDEC says it might be updated in future to cover other nonvolatile memories.
JEDEC has two other NVDIMM standards in play. The first is a standard being developed for what JEDEC has labelled as NVDIMM-Ps, which will be powered entirely by an SCM, or by a combination of DRAM and an SCM. In the latter case, they will use DRAM as a front-end cache to a back-end primary data store consisting of an SCM. It is not yet clear, but the options for the back-end SCM could include flash, as well as emerging memories.
The NVDIMM-P standard is being developed behind closed doors, and JEDEC says most of the technology leaders in the industry are taking part, with 15-20 JEDEC member companies meeting twice a week as part of an NVDIMM-P task force. 451 Research believes the NVDIMM-P standard will cover fast, memory-style byte-level access to data, as well as slower, storage-style block access to data.
Although the standard has not yet been published, Netlist is set to release such a device this year. Netlist has backing from flash giant Samsung, and in 2016 unveiled plans for a device like an NVDIMM-P that uses DRAM as a front-end cache to a larger amount of flash inside the same device. At that time, Netlist did not make any quantitative claims about performance, but said the devices will allow a server to be loaded with up to 12TB of memory. Even when there is a cache miss and data has to be accessed from flash, that access will be at byte level, rather than the slower block access used by other flash devices. The target applications include big-data analytics, as well as in-memory and other high-performance databases.
Intel's and Micron's 3D XPoint memory is claimed to be 10 times denser than DRAM, about 40% cheaper and 1,000 times faster than flash. That performance claim has been undermined by demonstrations of 3D XPoint-powered PCIe-NVMe cards that have shown a far smaller advantage over flash. However, 451 Research believes that this is in large part because PCIe links are not fast enough to make the most of 3D XPoint's performance. Indeed, Intel itself has predicted that 3D XPoint will be used more often in NVDIMMs. So far the two companies have only demonstrated 3D XPoint working in PCIe-NVMe drives, and Intel has said those will ship first, before 3D XPoint NVDIMMs. At present, that makes it unlikely that 3D XPoint NVDIMMs will ship any sooner than late this year, and probably later.
Elsewhere, Sony and Viking said this summer that they are jointly developing an NVDIMM powered by Sony ReRAM memory. The duo did not give details of the device, which might be powered purely by ReRAM, or could combine ReRAM with DRAM. It is very unlikely to be an NVDIMM-N, which would use ReRAM simply as an alternative to flash as a backup medium. One reason is that the cost of ReRAM compared with flash would be very hard to justify when used only in a backup role.
The second standard in play is not in development yet, but JEDEC say it has been 'defined.' If developed, it would be for devices classified as NVDIMM-Fs, which are powered entirely by flash. 451 Research believes that if or when this standard is developed, it will also cover both block and byte-level data access.
Diablo Technologies has already shipped two such devices, with the second device providing byte-level access to data. Diablo claims that for a Spark workload at a Fortune 250 financial services company, that second device has reduced three-year TCO by 65% while boosting performance by 29%, and reducing server count by a factor of three. Late in 2016, Diablo said its server maker, SuperMicro, had qualified the device, and that other major server makers were set to follow suit.