Innovative packaging techniques squeeze more RAM into tighter spaces
Rick Cook
Unless you're trying to upgrade a laptop, you don't generally think about how the memory for computers is physically packaged. But computer makers care a great deal about memory packaging. Computers, especially notebooks, are shrinking at the same time that memory requirements are rapidly escalating. Memory-hogging OSes like Windows NT and OS/2 are running memory-hungry graphics-oriented applications.
The market-research firm Dataquest (San Jose, CA) says the amount of DRAM found in the average business computer has grown from 4 MB in 1993 to 16 MB today. It will jump to 24 MB by 1997. Dataquest found that computer makers, trying to hold down prices in a fiercely competitive market, are shipping most sy
stems with just enough memory for the basics--typically 8 MB. But customers know that isn't enough, and they're doubling the memory--often before leaving the store.
Dataquest memory analyst Ron Bohn says, "Companies try to ship with as little memory as they can without losing sales and while retaining quality." Herein lies the second part of the memory-packaging challenge: not just to cram more memory into smaller boxes but to do it in a way that allows users to add even more memory later on.
Skyscraper RAM
In early personal computers, RAM came in the form of DIP (dual in-line pin) chips--those little black centipedes with the silvery legs so beloved of experimenters and kit builders--that were soldered or socketed onto the motherboard. Later on, these chips were installed vertically onto small circuit boards that took up less space. While there's still plenty of life left in the SIMM format and its emerging successor, DIMM (dual in-line memory module), there's still rea
l estate above those packages waiting to be grabbed. In an exciting shift, several companies have announced memory-packaging technologies that take advantage of this third dimension by stacking memory chips closely on top of one another.
Among the companies that have announced 3-D memory modules are Cubic Memory, Staktek, and Dense-Pac Microsystems. A number of other companies, including Intel, Hitachi, and Texas Instruments, reportedly have development projects under way for 3-D modules. (Let's clarify one point of terminology: 3-D has two distinctly different meanings in relation to RAM chips. The various technologies for stacking or layering wafers are generally called 3-D, referring to a physical construction. Matsushita, however, uses the name 3D RAM for its specialized graphics-oriented memory chip, referring to what it's used for. The 3-D memory modules discussed in this article are 3-D in form, not function.)
Memory stacking has been tried before, albeit more crudely, by several companie
s that piggybacked several DIP chips into a single package. But all those early attempts, such as the 512-Kb modules on the first IBM ATs, used ICs that were already packaged. More recently, some companies have made special 3-D memory modules for such low-volume, high-cost applications as defense and aerospace. The newest generation of memory modules puts naked chips--bare silicon--into special packages. The thinness of the silicon wafers means you can cram more of them together, and maybe even at a competitive price.
The packaging potential of these new techniques is awesome.
Staktek's Ribcage Stakpak DRAM module
gets 16 MB into a volume of one-eighth of a cubic inch.
Cubic Memory
does even better, packing up to 2 GB of DRAM per cubic inch in its 3-D modules. That kind of packaging density promises even smaller systems and even larger memory capacities in workstations and servers.
Ingenious as they are, 3-D memory modules pose some special problems
for their creators. For example, to make the most of them, you need more address lines from the module to the motherboard than a standard SIMM or DIMM provides. That means a nonstandard motherboard layout, which in turn implies proprietary packaging. The small computer market has shown a considerable aversion to proprietary hardware, but the advantages of 3-D modules are so compelling that many computer makers have announced plans to use them.
While heat is a definite problem with logic chips (remember the first 5-V Pentiums that introduced us to the fan-on-a-chip?), it's a lot easier to handle in DRAM. Nonetheless, 3-D module makers take care to ensure adequate cooling via ventilation ribs and by using signal leads to conduct away heat.
Staktek has been building 3-D memory modules since 1992 and has patented several packaging techniques. Staktek's products use either standard memory dies or TSOPs (thin small outline packages) in thin plastic-leaded packages with an external metal mounting struc
ture that provides both heat and signal transfer. To make clear that its technology is strictly a matter of packaging, the company's Stakpak products are available with flash RAM, up to 256 MB per module; DRAM, up to 128 MB; or SRAM (static RAM), up to 32 MB.
Cubic Memory is one of the most active companies making 3-D memory modules. Its process stacks DRAM chips that have gold wires running to the wafer edges. They are then interconnected by a conductive silver epoxy (see the figure
"Wafer Stacking"
). Cubic Memory is building 64-MB SIMMs for the 100-MHz Pentium-based laptop from Tadpole (Austin, TX), as well as other modules for Tadpole's SparcBook 3. For the Pentium model, Tadpole wanted to fit 128 MB of memory in just two SIMM slots. Panasonic is another Cubic Memory customer, offering a 32-MB DRAM PC Card for its notebook computers.
Clearly, 3-D modules will carry a price premium for some time to come. Cubic Memory claims that its devices cost only slightly more than
conventional packaging--$44 to $56 per megabyte, compared to $35 to $45 per megabyte for standard SIMMs, according to its president Chet Brown--but the difference is real and will matter to most computer makers. That extra cost will probably limit the use of 3-D modules to situations where space is especially tight and memory requirements are particularly high. It's no accident that Tadpole is using the modules in high-performance notebooks.
However, not everyone thinks that 3-D modules need be costly. Staktek recently introduced a lower-cost package it calls the Uniframe Stakpak. According to James Cady, executive vice president for engineering, the Uniframe DRAM modules are available in OEM production volumes for under $35 per megabyte now, and they should be under $25 by the end of the year.
At the high end of the computing spectrum, Cray Research is shipping its T90 supercomputers with SRAM stacks. Forty 4-Mb SRAM chips are soldered between two printed circuit boards to provide a dense stack
. Each memory module includes 16 stacks. Such dense packaging lowers the distance that signals must travel and stabilizes the capacitance of the signal lines. An expensive technology, it's well suited to supercomputer applications--computation-intensive operations that need massive amounts of memory. (For more on supercomputing, see "The Grand Challenges," February BYTE.)
The Old Reliables
When DIP chips lost out in the memory wars long ago, they were replaced first by the SIPP (single in-line pin package) and later by the SIMM, a module designed for DRAM in computers (and other applications) that is today's de facto standard for desktop computers. Essentially, a SIMM (and also a DIMM) is a small printed circuit board holding several DRAMs that plugs into a socket on the motherboard.
Perhaps the most obvious thing about SIMMs is that they stand sideways--perpendicular to the motherboard. This has two clear advantages: It saves motherboard real estate and permits better a
ir circulation around the chips. Also, the modules are easier to handle than memory chips. Thus, nearly anyone can install them.
SIMMs come in two flavors: 30- and 72-pin. The 30-pin SIMMs use 8-bit DRAMs, and the 72-pin SIMMs use 16- or 32-bit parts. Generally, it takes four 8-bit SIMMs to give the same capacity as one 32-bit SIMM. There is also the matter of organization. SIMMs are usually set up to read 8, 9, 32, or 36 bits at a time. The 9- and 36-bit SIMMs include an extra bit per byte for parity checking. If your computer expects parity memory and you add nonparity SIMMs, the machine won't recognize the extra memory.
In practice, this is less of a problem than it appears. While there is a rampant proliferation of part numbers, there are relatively few kinds of SIMMs out there. According to David Sun, vice president of engineering at Kingston Technology (Fountain Valley, CA), the many part numbers actually simplify the situation for users, letting them order memory by computer make and mode
l rather than having to know the specifications of the memory modules.
One result of the wide adoption of SIMMs is that add-on memory boards, once quite popular, are much less common today. Kingston, one of the biggest aftermarket memory suppliers, lists over 1000 memory modules but only 107 memory boards in its catalog, and just 13 memory-chip upgrades for computers.
Of course, the downside to memory modules is that you are limited to whatever kind of memory modules the manufacturer decided to support in whatever quantity the company decided to support them. An upgrader also must make sure he or she has the correct module for the system.
DIMMs Brighten
SIMMs are now giving way to DIMMs, which can pack twice as much memory into the same space. They achieve this by mounting DRAMs on both sides of the module and by using two sets of contacts, one on each side of the module board (SIMMs use only one set).
According to In-Stat, a Scottsdale, Arizona, semiconduc
tor consulting firm, SIMMs should account for 82 percent of the small computer memory market this year. By 1998, however, their market share will drop to 39 percent, and most of the rest will be DIMMs.
DIMMs are more expensive than SIMMs of the same speed and capacity, but manufacturers and analysts expect this to change as DIMM volumes overtake SIMMs. In-Stat analyst Connie Batchelder says that in the long run costs will be the same.
Is 3-D Memory in Your Future?
For all its technological ingenuity, not everyone is convinced that 3-D packaging is the way to go, at least not in the short term. Kingston's Sun says that 3-D modules are basically a way of getting a half-generation advance on new memory chips--say, from the current 16-Mb chips to the equivalent of 64-Mb chips. The problem that 3-D modules solve, Sun continues, is only temporary. When the higher-capacity part becomes more widely available, it automatically takes over from the inherently more expensive 3-D mod
ule.
Sun says that Kingston looked into developing its own 3-D module a few years ago, wanting to create a 16-MB SIMM equivalent out of 4-MB chips. In addition to the extra cost and limited life span, he says that a major deterrent was that Kingston would have to buy memory wafers from one of the major DRAM fabricators. But once the higher-capacity chips became widely available, the DRAM supplier would have no economic incentive to cut into the sales of regular DRAM by selling the smaller wafer.
Still, Sun admits that 3-D modules have their place in systems where space is extremely tight and the manufacturer must pack in a lot of memory. He suggests that 3-D modules might make good sense in a PDA (personal digital assistant) or a graphics workstation doing high-end animation work.
But they may see even broader use in the future. The primary argument against 3-D modules is that the next generation of DRAMs will be along fairly soon, which will quadruple capacity per chip at the same or low
er cost per megabit. This is based on historical trends, but this extrapolation is no longer a sure thing. And if the progression does break down, alternative solutions such as 3-D modules could become the new standard for memory modules. (For more on why the market hasn't yet seen higher-capacity DRAMs--and may not see them for a while--see "Why RAM Prices Stay High".)
More in Less
Although we can't be sure what technology they will use, we can be absolutely certain that computer makers and users will pack more and more memory into computers at an increasing pace. What is ample at one point in time turns out to be unacceptable in a year or two.
Today's notebook computers can already use amounts of RAM that were unheard of in mainframes even 20 years ago, and they need every bit of the RAM they have to support modern applications and OSes. There's no reason to think that our memory requirements will diminish in the foreseeable future, so it looks like we're each going to
need a lot more RAM in our computers just to see the century out.
WHERE TO FIND
Cray Research
Eagan, MN
(800) 284-2729
(612) 683-7100
fax: (612) 683-7199
Cubic Memory
Scotts Valley, CA
(408) 438-1887
fax: (408) 438-1890
Dense-Pac Microsystems, Inc.
Garden Grove, CA
(714) 898-0007
fax: (714) 897-1772
Hitachi America, Ltd.
Semiconductor and IC Division
Brisbane, CA
(800) 285-1601
(415) 589-8300
fax: (415) 583-4207
Staktek Corp.
Austin, TX
(512) 454-9531
fax: (512) 454-9409
illustration_link (24 Kbytes)
An eight-layer stack of DRAM wafers holds 64 MB and is 0.08 inch thick
Here's how Cubic Memory layers DRAM wafers into a high-density silicon stack.
photo_link (12 Kbytes)
This 128-MB 3-D module from Cubic Memory is surrounded by the 64 separate 16-Mb DRAM chips it replaces.
photo_link (19 Kbytes)
Two 3-D memory modules from Staktek. At the left, the Ribcage Stakpak incorporates 16 MB of fast page mode DRAM in a package that's only 0.5 inch wide, 0.9 inch high, and 0.3 inch thick. At the right is the newer, less expensive Uniframe Stakpak module.
Rick Cook is a freelance writer who lives in Phoenix, Arizona. He can be reached on the Internet or BIX at
rcook@bix.com
.
More Memory In Less Space
