3D XPoint was debuted with big claims in 2015. However, there are many wild guess and speculations because details have not been shared in public domain
Early this year, details about 3D XPoint came out from EE Times interview with Guy Blalock, co-CEO of IM Flash. 3D XPoint is well-known phase change memory and switch (PCMS).
(Picture source: Intel)
According to Micron, 3D XPoint has many technical and operational challenges, such as 100 new materials raising supply chain issues, cutting fab throughput by 15%, 3x-5x increase in capital expenses compared to planar NAND, and heavily depending on lithography tools. Therefore, 3D XPoint becomes expensive. The 2nd generation 3D XPoint with 4-layer stacking is expected to be about 5 times more expensive than planar NAND. So, it seems difficult to be an affordable storage device. Instead, 3D NAND will serve the storage market.
According to Intel/Micron, 3D XPoint is aiming at high-end SSD and DDR4 NVDIMM markets. Though, 3D XPoint-based SSD will serve niche market because of cost issue. The sweet spot of 3D XPoint should be the DDR4 NVDIMM because of low read latency (i.e. ~100ns). The configuration of DDR4 NVDIMM is shown below. Data comes to DRAM from CPU and copied to 3D XPoint. Then, the data goes back to CPU from 3D XPoint. It is possible because 3D XPoint has low read latency. However, because 3D XPoint has much less endurance compared to DRAM and its write speed is very slow, it cannot directly replace DRAM. Instead, as a shadow of DRAM, data in DRAM is copied to 3D XPoint in order to expand the size of main memory. In this case, DRAM works as a buffer for 3D XPoint. As a result, 3D XPoint will significantly replace DRAM if it works well.
One important consideration is the cost of 3D XPoint. The 2nd generation of 3D XPoint with 4-layer stack is expected to be about half price of DRAM. If the cost advantage of 3D XPoint over DRAM is less than 2x, then, customers may prefer to use DRAM instead of 3D XPoint even though 3D XPoint is non-volatile. DRAM still continues about 30% bit growth every year. So, considering the technical and operational difficulties associated with mass production of 3D XPoint, it will not be easy for 3D XPoint to keep up with 2x cost advantage over DRAM.
(Picture source: BeSang Inc.)
Software-based NVDIMM-P is emerging as a competitor against 3D XPoint. As shown below, predictive software allocates data between NAND and DRAM in anticipation of being needed. In fact, such prediction may work or may not work well. In order to increase the chance of success, NVDIMM-P should increase the size of DRAM. Therefore, NVDIMM-P strongly depends on DRAM while 3D XPoint reduces the dependence on DRAM.
(Picture source: The Register)
3D Super-NOR is also emerging and could replace 3D XPoint. Similar to 3D XPoint, 3D Super-NOR is NOR-based non-volatile memory having low ready latency (i.e. ~100ns). It does not need huge fab investment. Memory cell is vertically stackable and it has very small cell size (i.e. 4F2). Unlike 3D XPoint, no new material and no extreme process are needed at all. Most importantly, 3D Super-NOR is about 10 times cheaper than 3D XPoint and 20 time cheaper than DRAM. Therefore, 3D Super-NOR will be attractive as a replacement of DRAM at DDR4 NVDIMM. As shown below, 3D Super-NOR shares the same memory cell structure with 3D Super-NAND which is expected to be 2 cents per GB.
(Picture source: BeSang Inc.)
3D XPoint has huge market potential as a main memory. However, the technical and operational difficulties associated with mass production of 3D XPoint may make it difficult to be an affordable solution over DRAM. Software-based NVDIMM-P could be a low cost alternative solution. However, there should be limitation as a software driven technology. It also increases DRAM dependence. 3D Super-NOR could replace 3D XPoint and provide the most affordable storage class memory solution.
—Sang-Yun Lee is the CEO of BeSang Inc. (based in Beaverton, OR), which designs, developms, and delivers TRUE
3-dimensional (3D) integrated circuit (IC) solutions."