Will SCM come to the market and disrupt traditional memory hierarchy?
Storage class memory (SCM) is aiming at the wide spectrum of storage and main-memory market segments with non-volatile, short access times, and low cost-per-bit. Yet so many technologies are being developed to replace today’s NAND flash, DRAM, or fill the gap in between DRAM and SSD that, despite having heard these claims for long time, ideas like SCM still remain just that — just an idea — for now. Our question is “Will SCM come to the market and disrupt traditional memory hierarchy?”
So, let’s check the reality of SCM.
As we understand it, a big gap exists between DRAM as a main memory and storage memories, such as SSD and HDD, in terms of performance and cost. However, current systems are designed for this traditional memory hierarchy, and we do not need to change the system unless a compelling reason arises, such as ultra-low cost, high-performance, or both.
There are many new memories. I categorize them into three groups:
- Emerging memory, which includes MRAM, CMOx, and RRAM.
- NVDIMM-P, which combines NAND and predictive software.
- NOR-based NVDIMM, which includes 3D XPoint and 3D Super-NOR.
Various emerging memories have been developed as alternative solutions in case scaling of silicon memories reach the limitation. However, emerging memories did not close the gap with silicon memories for long time. For your information, recent progress, status, and challenges of emerging memories are well summarized by Storage Class Industry Association (SNIA).
In case of MRAM, cell size is about 50 times larger (which comes from effective cell size considering F2 difference and bit-per-cell) compared with planar NAND cell size. To overcome cell size issue, many emerging memories, such as RRAM, CMOx, and 3D XPoint, are adopting cross-point structure. It looks simple, reasonable, and good for low cost. However, cross-point structure is very expensive because it heavily depends on advanced lithography tools. Therefore, as Micron explained, fab investment and manufacturing cost of cross point structure will be significantly higher than planar structure. Considering material issue, cell size issue, and manufacturing cost issue, it will take long time for emerging memories to compete with silicon memories or to find out new market opportunities as SCM.
In case of NVDIMM-P, predictive software allocates data in advance between DRAM and NAND in anticipation of being needed. However, limitation of NVDIMM-P is obvious. Even though NAND flash is physically located at DIMM along with DRAM, traditional memory hierarchy still remains the same. The NAND still works as a storage device and DRAM works as a main memory. Because read/write of NAND is substantially slow, CPU should interact with DRAM only for fast operation. Predictive software for NVDIMM-P may work or may not work well occasionally. If necessary data is not ready at DRAM, then, the data read from NAND (i.e. ~1,000x slower than DRAM) will significantly slow down system performance. In other words, system performance cannot be maintained with NVDIMM-P at all time. Even though the predictive software works perfectly well, it is not practically possible to use NAND as a main memory because of limited endurance of MLC/TLC NAND (i.e. ~10K endurance). CPU exchanges data with main memory frequently. Simply NAND cannot withstand such frequent program/erase. So, it is fair to say that NVDIMM-P is an improved SSD, not a SCM.
NOR-based NVDIMM is non-volatile and could work as fast as DRAM thanks to low ready latency (i.e. ~100 ns) of NOR flash. It does not need predictive software. Therefore, high performance could be maintained at all time. Both 3D XPoint (3DXP) and 3D Super-NOR (3DSN) have high endurance thanks to single-level cell (SLC). Therefore, both 3DXP and 3DSN could work as main memory, significantly replace DRAM, and boost system performance abruptly with large amount of main memory. As a result, NOR-based NVDIMM could disrupt traditional memory hierarchy. If NOR-based NVDIMM could be cheaper than NAND-based SSD, it is expected to change the rule of the game in the semiconductor industry and take large portion of SSD/HDD and DRAM market segments.
3D XPoint (3DXP) is well-known phase change memory and switch (PCMS) which is not NAND, but NOR flash memory. A while ago, PCMS tried to replace NOR in the market. However, high cost was the main issue. Now, 3D XPoint enables multiple cell-stacking. Though, it seems not easy to bring down the cost rapidly. There are many technical and operational challenges associated with mass production of 3DXP (see “3D XPoint — Reality, Opportunity, and Competition” at EE Times).
According to Micron, the 2nd generation of 3DXP with 4-layer stacking is expected to be 5x expensive than NAND and 2x cheaper than DRAM. Therefore, 3DXP will not be an affordable storage device, but could be an affordable main memory which replaces DRAM. Due to the cost issue, it seems difficult for 3DXP to disrupt traditional memory hierarchy.
3D Super-NOR (3DSN) is silicon NOR flash memory. Unlike planar NOR which is expensive with large cell size (i.e. 10F2), 3DSN is very affordable with small cell size (i.e. 4F2) and vertical cell stacking. Expected dollar-per-GB of 3DSN with SLC is about 6 cents-per-GB. It shares the same cell structure with 3D Super-NAND which is expected to be 2 cents-per-GB. For more information about how 3DSN enables low cost, please check “3D NAND Flash at 2 Cents per GB” at EE Times. Because 3DSN is significantly cheaper than SSD and DRAM, it could disrupt traditional memory hierarchy. 3DSN prefers to use SLC instead of MLC/TLC in order to achieve low read latency and high endurance, which is essential for main memory. It has about half billion effective R/W endurance cycles (assuming 1,000 to 1 memory size ratio between 3DSN and DRAM). NOR is fast (i.e. ~100 ns read latency). The same as to DRAM, NOR has capability of random access. As explained, it does not need predictive software. Therefore, high performance could be maintained at all time.
Summary and comparison of SCMs are shown in the graph. Emerging memory has many issues, such as materials, large cell size, and high manufacturing cost along with cross-point structure. Considering all these, the chance for emerging memory to close the gap with silicon memory seems not high. NVDIMM-P is not storage class memory, but could be a better SSD. Performance of NVDIMM-P cannot be maintained at all time with predictive software. NOR-based NVDIMM has a potential to replace both NAND and DRAM if it has competitive cost-per-bit. It is as fast as DRAM and has high endurance. Also, high performance could be maintained at all time. It has random access capability. Most important matter is cost-per-bit. As explained, 3D XPoint will be difficult to be affordable because there are many technical and operational challenges associated with mass production of 3D XPoint. So, it will compete only with DRAM. 3D Super-NOR is silicon NOR flash memory and has ultra-low cost. It has been successfully developed. So, it is expected to provide ultimate solution for storage class memory and disrupt traditional memory hierarchy.
—Sang-Yun Lee is the CEO of BeSang. Inc. BeSang is open to collaborate with system makers to introduce 3D Super-NOR NVDIMM to the market.