DRAMs are the highest-volume commodity semiconductors built today, with about 11 percent of the total semiconductor market. Nevertheless, DRAM sales last year totaled about $14 billion, down from $40 billion three years earlier, even as the number of parts and bits shipped continues to grow. So why the downturn in the value of the market?
To understand this better, it is necessary to comprehend the four forces that affect DRAM evolution: technology trends, system designers, producers and consumers. Although all four are at work, it is system requirements that determine the need for the evolution and, ultimately, which solutions are used.
Technology trends have a huge impact on the DRAM market. They directly affect the evolution of DRAMs and thus the market itself. Of course, while process shrinks are constantly occurring, these processes tend to be behind advanced processes that are used for microprocessors and they tend to have many fewer levels of metal. DRAM manufacturers are constantly trying to use technology to reduce the cost of production and increase yield. The number of bytes of components per wafer and cost of fabrication all relate to reducing this cost, as does saving a mask layer. DRAM makers will do what they can to keep the costs down; currently, the processes that most use enable them to very cost-effectively produce 64-Mbyte and 128-Mbyte DRAM densities. By the end of this year or the beginning of 2001, 256 Mbytes should be cost-effective.
It is the very nature of the DRAM market that memory storage increases as system costs decrease-applications continue to drive the need for more memory while system costs stay the same or drop. As processes continue to shrink, one of the biggest future trends that will occur is diminishing Vdd. The industry was lucky that the 5-volt standard lived as long as it did-even 3.3-V components were mainstream for a long time. But as processes moved below 0.25 micron, lower Vdd values have become inevitable.
Rambus DRAMs were the first DRAMs to run on 2.5 Vdd and we can expect to see 1.8 V in the near future. ASICs that connect to these DRAMs are already running at 1.5 V, with future processes dropping Vdd down close to 1 V in the coming year or two. Maintaining signaling compatibility among ASICs, processors and DRAM will become a serious challenge as Vdd continues to drop.
Changes in the interface
Interface evolution has also started changing more rapidly. As recently as just a few years ago, the standard page-mode DRAM interface was nearly identical to the original interface going back to the 4-kbit memory generation. Over time, devices went from by-1-bit to by-8 or by-16-bit configurations, adding page mode and fast page mode as standard. But essentially the interface was nearly the same. Access times decreased in the last 20 years, but not by much. Early DRAMs had access times of approximately 250 ns; SDRAMs today are around 50 ns.
In 1995, the first Rambus DRAMs shipped in volume. These were the first synchronous DRAMs and were also the first DRAMs to use double-data-rate technology. They were used in the Nintendo 64 and have a data rate of 500 MHz. The Rambus DRAMs that are shipping in PCs today have an 800-MHz data rate on the pins and are 2 bytes wide, providing 1.6 Gbytes/second of data from a single DRAM. This is eight times the data rate from a typical SDRAM used in a 100-MHz DIMM today and twice the peak module bandwidth. Rambus is working on technology that will double the data rate of the DRAMs again and quadruple memory module bandwidth to 6.4 Gbytes/s.
SDRAMs continue to evolve as well. The 133-MHz SDRAMs are already going to market, with speedier 166-MHz and even 200-MHz parts on the horizon. DDR SDRAMs add a double-data-rate feature to the DRAMs, with data rates of 266 MHz being discussed-in fact, they may be shipping by the time this article is published.
There are many other interfaces available on DRAMs, including specialty interfaces tailored to a particular application, such as graphics. In general, these parts are available from only a small subset of all manufacturers, have larger die sizes than standard DRAMS and are priced higher than standard DRAMs because of the increased cost and lower volume.
Ultimately, the system or ASIC designers determine what memory types will be used. DRAM has been the memory of choice for system designers because it offers the lowest cost per bit of any semiconductor memory. Static memory is much easier to design with, but it is significantly more expensive and is not available in densities as high as those of DRAMs. However, DRAMs add complexity and require occasional refresh and also have a nonuniform time access pattern that can be difficult to optimize for every application.
Memory is usually a small part of the overall project, but often the design of the memory subsystem and the selection of the type is key to meeting performance or cost objectives. Memory can affect the system in several ways: It can limit performance, it has an effect on total cost and the choice of memory will affect everything from board space to power to radiated emissions. The decisions these designers make will determine which memories will be needed in a year or so if the product they are designing is successful on its own in the market.
The range of application space for DRAM is daunting. It covers the gamut from embedded memory in hard-disk and graphics controllers, to small memory systems for game machines, set-top boxes, voice recorders, high-end graphics controllers, printers and many other products. Today, PCs consume the most memories and require expandability in addition to low cost. High-end memory applications such as servers and mainframes can literally use thousands of DRAMs in a system. Packaging, power and performance in addition to reliability and scalability are key to designers' goals.
There are many interesting products on the horizon, part of the so-called post-PC era. Companies are continuing to innovate with new and exciting products and platforms such as the Sony Playstation 2, the Palm Pilot series of products, MP3 players and recorders, and digital video products, including HDTV and hard-disk-based VCRs. The profusion of combinations of fixed-function devices, as opposed to general-purpose PCs, seems to point to the future new markets. These products, coupled with the servers and networking equipment needed to completely rebuild the Internet to meet an almost insatiable need for bandwidth, will continue to demand more DRAM and more high-bandwidth DRAM.
One of the most significant forces on the DRAM market is the DRAM companies themselves. In the past several years, many companies have quit the DRAM business, been sold or merged with other companies. Texas Instruments Inc.'s business was acquired by Micron Technology Inc., LG was forcibly sold to Hyundai, Motorola Inc. has left the DRAM business entirely and Siemens AG has spun off its semiconductor operations-now called Infineon. Several other companies have reduced their output while new companies, primarily in emerging markets such as Taiwan, are just beginning to produce DRAMs.
Hyundai, Micron and Samsung are now the big three producers-and fierce competitors-with Hyundai in the position of acquiring a company about its own size and also a large debt. Beyond the top three, Fujitsu, Hitachi, IBM, Mitsubishi, Mosel-Vitelic, NEC, Oki, Toshiba, Vanguard and several more are the remaining DRAM producers.
DRAM development is in general a longer cycle than that of ASICs, taking more than a year from start to first production of a DRAM. Production processes are different as well. Because DRAMs are designed with row and column redundancy, yield for DRAMs exceeds that of ASICs or logic devices. DRAMS are also different from many other ICs in that much of their operation and characteristics are analog in nature. Many parameters of DRAMs, like precharge time, access time, column cycle time and even interface parameters like setup and hold can have an effect on yield. So it is a very complex process to determine the best trade-off to meet the system designer's need for more performance while keeping yields high and costs low.
Typical processes today are at about 0.22, going to 0.18 in the coming year. The key to a process is how the storage capacitor is designed. Significant R&D goes into designing next-generation process technology; it has become so expensive that it is now common for DRAM companies to work together.
Building a fabrication plant is a big, expensive decision. A new state-of-the-art facility could easily cost over $1 billion, and there are almost constant expenditures needed for new equipment to build parts to ever-shrinking geometries. DRAM companies have to predict more than a year in advance if there will be sufficient worldwide capacity and demand to justify the capital expenditures. In some cases, those companies have built facilities and have let them remain unoccupied because excess capacity existed.
However, with the significant reduction in the number of suppliers and an apparent increase in demand from users, supply cannot meet demand right now, which has been causing DRAM prices to increase in the last several months. In December 1995, a 16-Mbyte DRAM sold for about $55; earlier this year, the same device sold for less than $3. This dramatic drop followed a period when DRAM prices remained higher than the normal trend line would have predicted. Pricing has since firmed to slightly higher levels, but for a while prices were well below the long-term trend line.
Who's buying what
The largest market for DRAMs is the PC industry, where an estimated 55 percent of all DRAMs go-to such OEMs as Apple, Compaq, Dell, HP and IBM. Commonly, they buy directly from the DRAM companies through contract pricing that is constantly renegotiated.
The next largest market is the workstation/server market with companies such as HP, IBM, Sun and others. Most of the PC and workstation markets use memory modules, buying or subcontracting them from companies such as Kingston, Smart, Viking and dozens of other memory module companies.
Few if any system companies build their own.
Of the non-PC, non-main-memory applications, most of the remaining DRAMs are sold as individual devices for diverse applications such as game machines, peripherals (such as printers), communications equipment (such as networking boxes) and thousands of other products.For the most part, the companies purchasing the DRAMs have little choice in the DRAM they buy-the decision was made by the system designer months or even years ago. Sometimes they have a choice-a system designer may have been able to accommodate two or more choices in the design-but normally a design is fixed with a particular memory type in mind. For main-memory applications, this means the purchasers are limited to finding interchangeable parts from the various manufacturers, if they have been qualified as a supplier.
Technology is an ever persistent driver of increasing density and smaller die size, leading to lower-cost-per-bit memories. Few industries have seen the constant reduction in cost experienced by the electronics industry, primarily because semiconductor production is like a printing business, with the fonts getting smaller each year.
The system designers face the difficult task of building ever more complex and interesting products while juggling the huge amounts of information and misinformation concerning technology trends, costs and market dynamics. Their selections often border on religious belief, since hard facts are often hard to come by. These designers select which memory will be the ones used in future products and thus affect the buying habits of the OEMs.
The dynamics in the marketplace between the DRAM companies and the OEMs that build the products determine the supply-and-demand equations and ultimately set the price. Small imbalances in supply tend to greatly affect price, and the DRAM market seems to go through these cycles every four or five years. Now, we appear to be coming out of a low point where capital spending has not kept up with increasing demand, resulting in more demand than supply. Thus, prices have firmed and are expected to possibly go up or at least remain stable for a while. Meanwhile, the DRAM companies will continue to reduce their cost by aggressively shrinking die size and moving to denser products.
Ultimately the consumers benefit from this cycle of forces that play on the DRAM market, although for those of us close to the action it sometimes seems like a relentless race to deliver ever more bits and bandwidth for fewer dollars.
See related chart