Get email delivery of the Cadence blog featured here
SNIA, the Storage Networking Industry Association, organized a webinar recently with industry analysts Tom Coughlin (of Coughlin Associates) and Jim Handy (of Objective Analysis). The webinar was about emerging memories, meaning MRAM (magnetic), ReRAM (resistive), PCM (phase-change), and FRAM (ferroelectric). The main focus was PCM aka 3D-Xpoint, aka Optane, and MRAM. The key thing about all of these memory technologies is that they are persistent.
Persistent memory (henceforth PM in this post) requires support to get the most out of it. On the hardware side, JEDEC and others are supporting early developments, and crystallizing requirements and interfaces. On the software side, the SNIA and others are working on operating system support and application program support.
Some early groundwork was achieved with NVDIMM-N, which are DRAM DIMMs with a flash backup on the DIMM (and a battery). When the power fails, the DRAM is copied into the flash, when it is restored, the DRAM is reloaded. This makes it appear as if the DRAM is PM. Two changes that this required are bringing a power fail signal out to the DIMMs so that the DIMM knows to copy the DRAM to flash, and BIOS changes so that the operating system can be brought up without reloading memory.
Ongoing work on the hardware side is Non-Uniform Memory Architecture (NUMA) because PM typically has different read and write times. This requires an MMU redesign. Intel has a proprietary redesign of DDR4 to be transactional, in the sense that, following a write, some reads can be performed before the write completes. This is focused on their 3DXpoint memory that they sell under the name Optane.
On the software side, there has been a lot of work done on the operating system side, but PM is "useless if its advantage is untapped" and the persistence is unknown in most software. Closed systems like hyperscale data centers and storage area networks can use it now, but open systems will evolve gradually and it will take time. There is the usual two-sided-market problem that nobody wants to update their software until PM is widely deployed, and there is not a lot of value in deploying it if everyone simply treats it as volatile DRAM.
If you have ever seen one of Jim's presentations, you will have seen this graph. I first showed it in a post from 2015 Memory and Storage: the Wall Is Coming Down. It was only my second week back at Cadence.
The graph shows price horizontally and memory bandwidth (performance) vertically. So the L1 cache is really expensive but really fast at the top right, and tape is really slow but cheap in the bottom left. In the middle is where persistent memory sits, with higher performance and cost compared to an SSD, but lower performance and cost than DRAM. Jim's point is that for 3D-Xpoint/Optane to take off, it has to be priced below DRAM, otherwise people will just buy DRAM. Intel now has Optane DIMMs available and they are a key selling point for the next-generation server Cascade Lake CPUs. For more details on that, see my post about Intel's presentation at HOT CHIPS, Intel's Cascade Lake: Deep Learning, Spectre/Meltdown, Storage Class Memory. Optane SSDs are gaining what Jim calls "modest acceptance".
Historically, there have been two types of MRAM (M for magnetic): Toggle Mode and Spin Transfer Torque (STT). Everspin makes both of these (GLOBALFOUNDRIES is the actual manufacturer) and they are currently the only vendor of standalone MRAM. They have shipped over 123 million units. Another company, Avalanche, is sampling. STT MRAM is used by all the foundries for their embedded MRAM solutions. Markets for MRAM today are space, high-uptime systems, and caches/buffers (for example, IBM SSDs and some RAID controllers use MRAM).
There is a new technology called Spin-Orbit Torque (SOT) that has higher reliability, is faster than STT—in fact, it is as fast as SRAM. But there are no commercial products yet. It has the potential to eventually replace SRAM. Since MRAM is built in the metal stack, that could allow you to get embedded "SRAM" without requiring silicon area.
First, who wants PM if it costs more than DRAM? Obviously anyone who is using NVDIMMs, which are early adopters in areas like high-availability systems and financial databases. MRAM DIMMs are interesting in some of these markets since they are faster (but more expensive).
But Tom anticipates that memory prices will move past established technologies, as shown in the graph. The critical inflection is when PM gets cheaper than DRAM. It seems it is unlikely ever to catch 3D NAND flash, where more and more layers are continually being added. Note that PM is already cheaper than EEPROM, SRAM, and NOR flash. Way back in 2004, NAND got cheaper than DRAM and the market growth exploded. The same thing will happen with PM and, in particular, 3D-Xpoint. Once it gets cheaper than DRAM, everyone will want it (and, at least initially, the persistence will be of secondary importance).
But it's a moving target since DRAM prices have fallen 55% in the last two years (January 2018 to January 2020). This is part of the commodity cycle at work (which applies to things other than memories):
But the future is rosy. Tom and Jim's forecast is for over $37B in emerging memories (MRAM and 3D-Xpoint) by 2029.
There is a PDF of the presentation (but the webinar itself was not recorded). Tom and Jim also have a report Emerging Memories Ramp Up for sale.
Tom and Jim mentioned that Cadence now supports DDR4 MRAM. To be more precise, last August, Cadence announced that our DDR4 memory controller now supports Everspin's 1Gb STT-MRAM. I didn't write about it at the time but Anandtech did. Here is the Cadence DDR4 page.
If persistent memory is a topic of special interest to you, then take a look at the Persistent Memory Summit in Santa Clara on January 23. Here are full details including a link for registration (it's free). I will attend.
Sign up for Sunday Brunch, the weekly Breakfast Bytes email.