Reading this article made me dizzy with all of the different variants of RAM that are being worked on. But I'm sure that some readers will find it useful and interesting.
electronicdesign.com
Here are Handy's concluding remarks, which seem to me to be on the mark. His comments on Intel and XPoint are a cogent response to the question some people have about why Intel seems to use XP much more than Micron.
I am often asked which memory technology has the best prospects for success, and which applications are better matched to which technology. My short reply is that none of these technologies is particularly better suited for an application than is any other, and that cost is the one overriding factor that will determine their success.
A glowing example can be found with today’s mainstream technologies: NAND vs. NOR vs. DRAM vs. SRAM. Each of these technologies competes against the others in many applications, even though they are extremely different from one another. Code can be executed directly out of NOR but not from NAND, but NAND is substantially less costly, and NAND/DRAM combinations have successfully taken the bulk of the cell phone market from a combination of NOR flash and SRAM. Server users have found that they can often cost-reduce their systems by reducing DRAM and increasing the use of NAND-based SSDs to achieve the same performance. NOR and SRAM win out for reasons of speed or economy when the size of the memory is small enough that a cheaper NOR flash or SRAM can replace a significantly larger (and slightly more expensive) NAND or DRAM.
Decisions of which technology to use most often are based on the total cost of the system and the performance it must achieve. At the moment all of the emerging memory technologies defined above are challenged to approach the cost of entrenched alternatives – today there is a difference of a few orders of magnitude for most systems. As volume shipments increase that situation should change.
Which Will Win?
Objective Analysis expects for the memory that races down the cost curve the fastest will stand the greatest chance of winning the bulk of the market.
Intel is forcing its 3D XPoint Memory down the cost curve by manufacturing significant volumes at a loss. The company can afford to do this because it believes that 3D XPoint is required to support the company’s higher-performance CPUs. Any money that Intel loses on 3D XPoint sales will be recovered through the sale of more costly processors.
Other technologies don’t have this advantage, and must be sold at a profit to make any sense. This is a chicken-and-egg problem: To reach high volumes any emerging memory must sell at competitive prices, but costs will prohibit that until high volumes are reached.
If DRAM and NAND flash do finally hit their scaling limits, then these technologies can become cost competitive after a couple of scaling nodes have been missed, as was shown in Fig. 2, but until that time they will be relegated to those niches that can tolerate their higher costs in return for some indispensable technical advantage, and those applications are very few, and have very low unit volume requirements. |
