The first of a planned series of Tech Talk presentations by Intel focused on NVM Express (NVMe) and was given to a worldwide web audience by Amber Huffman, senior principal engineer and NVM Express Workgroup chair. The presentation reinforced the message given by the same speaker at IDF13.
NVMe is the standardized high-performance host controller interface for PCI Express-based SSDs with an architecture designed for non-volatile memory scaling from the enterprise to the client. While ready to offer performance gains with the present generation of NV memory SSDs based on NAND, it has also been designed with the next-generation NV memory in mind. Its architecture focuses on latency, parallelism, and performance, and the details have been well documented. (See Expansion of NVMe Support Signals Growth and PCIe Storage Spec Group Incorporates for Added Clout .)
NVMe is now supported by 90+ leading companies and a 30-company promoter group. The first products began shipping in 2014.
The driving force for NVMe acceptance is its ability to move NV memory from the legacy of the rotating disc and serial operation to parallel operation. Standardization of the register, feature, and command sets is a means of moving from proprietary to product interoperability.
Flash memory multilevel cell (MLC) NAND provides the primary non-volatile part of the present proof of performance of NVMe-based designs, which is impressive. The appearance of emerging NV memories would appear to hold the key to a dramatic factor of 4x improvement in future performance. However, the saving grace is NVMe has been designed to be memory independent or “agnostic” to the memory technology that eventually wins the NV emerging technology race.
This writer hoped this Tech Talk would provide some insight into Intel's viewpoint as to which of the competing technologies might win the emerging NV memory race and, more importantly, the timing of its emergence -- even though, as stated earlier, the NVMe controller has been designed on a “don't care” basis in that respect.
Huffman's list of emerging memory types is shown in my Figure 1. It included the usual suspects with a brief description of the mechanisms involved. The list included as mechanisms: phase change, spin polarized electrons, electro-chemistry, oxidation-reduction, and oxygen vacancy drift diffusion and barrier modulation.
Figure 1: Emerging memory types
What was interesting is that a multilayer stacked memory matrix (see Figure 1) was used for illustrative purposes rather than one that was MLC based. In the past, Intel, with STMicroelectronics and, more recently, at IEDM 2009 with Numonyx, has looked at and demonstrated both multilevel cell (MLC-PCM) and stacked matrix, albeit the latter only a single-layer matrix where the thin-film memory isolation or selection device is stacked vertically above the memory device. As far as we are aware, those projects have disappeared into the PCM technology amortization trail of Intel-STMicroelectronics-Ovonyx-Numonyx-Micron.
It could well be that Intel recognizes that a stacked array may offer the only solution to any scaling problems, especially as Huffman included the word “scalable” in her slides referring to next-generation NVM. The development of a stacked thin-film matrix isolation device is a non-trivial problem that will require a solution before the next generation of multilayer NV memory can emerge.
This list in Figure 1 omits correlated electron memory devices, although oxidation-reduction might cover one case. CeRAM, without the filament; and, more recently, the change in tack to correlated electron device for the group developing PCMO-based memory will cover the other under both the oxidation reduction and modulated interface banners.
Huffman specifically mentioned only MRAM as an example of an emerging NV memory. A followup question by this writer inquiring if that is Intel's view of the most likely near-term emerging NV technology winner was rebuffed with a reply along the lines of: Intel does not wish to comment on emerging memory technologies or which one will win.
Figure 2 provides Huffman's comparative performance of the NVMe illustrating the dramatic gains that will be obtained when emerging NV memory eventually reaches maturity and is able to offer the access times comparable with today's DRAMs.
Figure 2: Comparative performance of NVMe
It is clear that when the next-generation memory appears (and with the assumption that the NV memory is no longer the bottleneck) there will be a need to optimize platform storage interconnect (electro-optic?) and the software storage access methods in order to obtain further performance gains.
For those with an interest in emerging memory development and solving the associated problems, this presentation defined what you will need to bring to the table to meet future NVMe performance expectations. As well as non volatility, the acceptable memory will require DRAM-like write/erase lifetime and access time, low power, and a thin-film matrix isolation device for matrix stacking.
We need a device with write/erase power dissipation that does not cause the core of a multilevel matrix to cook. The ability to scale to sub 20 nm will be essential, although some relief will be available in that respect if multilevel matrix stacking is possible. Clearly, emerging memory write/erase performance must become the focus of attention for the emerging memory wannabes.