Time to Look For Low-Cost DRAM Alternatives
#DRAM is turning into a seller’s market, and it’s time to look at low-cost alternatives. DRAM prices are heating up, and there is no easy solution to relieve this high price issue because it does not come from the imbalance between supply and demand, but instead from the end of #MooresLaw for planar DRAM (see related article, Why Memory Prices Are Heating Up). This year, bit growth of DRAM will be the strong 23 years. Planar DRAM scaling has also slowed considerably in the last three years, as shown in figure 1. Therefore, DRAM is transforming into a seller’s market and DRAM vendors are making record profits this year. Similar to oil crisis, customers are paying more for DRAM under the DRAM crisis. Therefore, I would like to discuss what kind of solutions we could find for low cost DRAM. ￼ ￼ ADVERTISING ￼ Figure 1: DRAM Roadmap Plan vs. Reality and DRAM ASP For the time being, emerging memories such as MRAM and phase-change memory (PCM) have been challenging planar DRAM. However, MRAM goes to embedded applications and PCM (i.e. 3D XPoint) is being used as high-end SSD applications. Realistically, it is difficult for them to directly replace planar DRAM considering cost-per-bit, performance and reliability. Thus, emerging memories are searching for their own niche market segments now. DRAM fab expansion is one method to increase DRAM supply. However, DRAM vendors may be hesitant because the cost of a new fab is about six times higher than a fab upgrade. For the time being, DRAM vendors have increased DRAM output though fab upgrades for the lateral scaling of DRAM, which increased DRAM bit output exponentially by a power of two. In contrast, a new fab without lateral scaling would require signficantly more investment for just a linear increase in DRAM bit output, which, coupled with higher manufacturing costs, will make DRAM more expensive. Additionally, in a seller’s market, it would be difficult for DRAM vendors to have strong reasons to increase DRAM fab capability through new fab investment without lateral scaling. So, switching from planar DRAM to 3D DRAM is necessary. As shown in the figure 2, more die-per-wafer could be produced using 3D DRAM. As long as the wafer processing cost of 3D DRAM is reasonable, and it is easy to generate 3D DRAM, it is beneficial to adopt 3D DRAM technology for low cost (see related article, Why 3D Super-DRAM). Certainly, 3D DRAM is crucial for the continuation of DRAM scaling. But, unfortunately, DRAM vendors do not have their own 3D DRAM technologies yet. ￼ Figure 2: 3D DRAM enables more die-per-wafer and low cost DRAM Another suggestion is DDR4 NVDIMM to replace DRAM as a main memory. There are many kinds of NVDIMMs. Unlike other storage-oriented NVDIMM, DDR4 NVDIMM should be as fast as DRAM and work as a main memory. Intel 3D XPoint suggested DDR4 NVDIMM a few years ago. Because 3D XPoint has NOR-type flash, it could work as fast as DRAM thanks to low read latency (i.e. ~100 ns) of NOR flash. It also does not need predictive software. Therefore, high performance could be maintained at all times. SLC NOR flash has high endurance thanks to the single-level cell (SLC). Therefore, NOR flash in DDR4 NVDIMM could replace DRAM, and boost system performance significantly with a large amount of main memory as shown in figure 3. If an affordable NOR flash could be generated through 3D NOR, DDR4 NVDIMM would be much more attractive. Second generation 3D XPoint (i.e. 4-ayer) is expected to be about half the cost of DRAM and has some mass production issues (see related story, 3D XPoint – Reality, Opportunity, and Competition). Conventional planar NOR is not as affordable because of the large cell size (i.e. about 10F2 to 12F2). Therefore, 3D NOR should be essential for DDR4 NVDIMM. In the case of 3D NOR, it could be 6 cents/GB (see related story, Will Storage Class Memory Disrupt Memory Hierarchy?). ￼ Figure 3: DDR4 NVDIMM configuration along with 3D DRAM and 3D NOR. For 10GB 3D DRAM plus 1TB 3D NOR, the cost of DDR NVDIMM is to be $10+$60=$70. Today, the industry is being crippled by soaring DRAM prices, and is just waiting idly for next downturn. Many analysts consider this upturn to be just another boom time for DRAM, and anticipate the next downturn as the market balances out between supply and demand. However, at this time, we should consider more factors to predict the DRAM market in the future. If we look into the history of DRAM, the downturns came when DRAM bit growth was more than 45 percent and caused oversupply. Now, Moore’s Law does not work for planar DRAM anymore and bit growth will be the lowest it has been for the past 23 years. Because we cannot see more than 45 percent DRAM bit growth from planar DRAM, there will not be any downturn in the future. Certainly, 3D DRAM is not an option today because the technology is not there. NVDIMM using 3D NOR could significantly reduce the total ownership cost of memory subsystems, because 6 cents/GB of 3D NOR could replace DRAM and SSD at the same time as a main memory as well as storage memory. As we know, planar DRAM has not been changed at all for the last several decades. We must start innovating DRAM now.