Flash Memory Summit 2018: a summary of new types of storage devices

Source: 　　Editor: Aron Update Time :2018-09-09

Persistent Memory has always been a hot topic of storage field, whose technology and drug development are similar and require a lot of time and huge investment. Technology like NAND Flash, which was developed successfully in 1985, has been coming down to where it is today with application.

For The current PM summary, we use The Memory Guys’ (http://thememoryguy.com/) comparison.

First, let’s look at the companies behind each technology. Currently, MRAM is Everspin+GF, and the previous Avalanche is completely silent this time. IBM and Samsung both have reports on MRAM, but they also have reports on other PM. For ReRAM, Crossbar + UMC’s investment seems to have disappeared with Martin's departure, compared with Saundersk and HPE before acquisition.

https://www.theregister.co.uk/2017/01/25/cutting_hewlett_packard_labs_down_to_size/

For FRAM, it seems that TI and Fujitsu are the most positive. But the two companies’ manufacturers and data centers are a bit far apart. For most new technologies, the existing tycoons of the industry are in a wait-and-see attitude. Besides not wanting to kill a thousand and injure eight hundred, there is also a question of innovators’ dilemma. Therefore, it can be seen that the status quo of MRAM and ReRAM is most suitable with this situation, which small companies struggles in front but big companies wait behind.

From the perspective of products, the three technologies of MRAM, ReRAM and Xpoint are the focus of the current market. However, Xpoint's status is basically similar to that of X86 in data centers. ARM and Power are both working hard, but the current situation is not looking that good.

FMS2018 also has an MRAM’s Developer Day this year, and it is not hard to observe from the topic that now MRAM mainly focuses on the Embedded market, which has some deviations from the data center. We do see that the FPGA of Xilinx is used by Everspin and several small factories as the controller of MRAM, and MRAM products based on PCIE interface is produced. However, due to the cause of capacity, it is definitely not used as the main storage, so it isn’t threatening to Xpoint at all. At the beginning of this year, the summary of MRAM is more comprehensive, which was made at the SNIA's NVM Summit.

ReRAM had been hot for a while because of HPE's "The Machine". In The 2016 NVM Summit, Sandisk staff confidently talked about the cooperation with HPE. They thought that Samsung did not care about the progress of NVM because they had 128G DIMM products. Therefore, they hoped that ReRAM could become the strongest opponent of Intel 3DXpoint. Later, because Sandisk was purchased by WD, HPE disbanded the team of "The Machine" and most of the members are working on a new deal, Gen-Z at present. Gen Z, like ZFS, shares the deep desire of the engineer Nerd for technology. "I'm the ultimate product and other people should sleep and wash," but I hope it's not ZFS.

In contrast to RAM, ReRAM is more closely related to the current AI. Making use of its own structure, there are many studies on PIM and NN.

3DXpoint is undoubtedly the winner of the current market. Although both Intel and Micron have been tight-lipped about what 3DXpoint is, there are many companies in the market have a realization on the whole.

The technology of PCM, made up of 20nm

Cross Point's array, stacked selectors

Intel, Micron, Numonyx and IP’s vendor have mentioned that it nearly took 10 years to achieve.

While the current market’s AEP doesn't achieve what Intel preheated in 2015, compared with NAND, it is definitely the winner with 1000X performance and endurance. After the production of the M.2/U.2 PCIE interface, it launched the DIMM slot.

Interestingly, the technology is a collaboration between Intel and Micron. It is about to end, but so far Micron has not launched any products. Especially in this year's FMS Micron exhibition, there is no news about 3D Xpoint, which is really a strange thing. Does Micron think that Intel’s loss-leader is only suitable for Intel? In the latest news, Google Cloud took the lead in deploying the DIMM of 3D Xpoint. Whether it could bring more applications to NVDIMM-P?

For single 3D Xpoint chips, the speculations are as follows:

l 128Gbit。>10 percent of OP

Read delay is ~ 125ns, write delay is larger

l Endurance: ~200K cycles

It can be seen that this is not an alternative to DRAM, nor is it an alternative to NAND Flash.

It can be seen that the life cycle of 3D NAND is much longer than expected, which may not a good news for the other two PM.

Additionally, an interesting phenomenon was reported. During a Session the next morning, newm-201a-1:3D XPoint: Current Implementations and Future Technologies Track. Although it is relatively early, it is indeed a crowded market. Because in the second link, the session of applied materials was really hard to understand, so directly went to pmem-201-1: Persistent Memory Software and Applications (Persistent Memory Track), where there was only less than 10 people. It is indeed strange that the interest in PM manufacturing is greater than the interest in how to use it, which may not be a good thing for PM.