Individual chips will also be able to hold 512 megabits of memory, twice as much as current chips do, a shift in density that occurs every few years and can be tricky to manage.
And DDR2 chips will have to be inserted in new types of packages. Not only does this introduce another element of risk, but manufacturers will also be required to pay royalties.
Finally, DDR2 is fundamentally different from DDR in its design, introducing a greater chance for error. If anything goes wrong, a manufacturer could end up with a higher-than-acceptable number of chips from each wafer that doesn't work.
"There is a difference in this transition," Donabedian said. "There is a lot to bite off."
Memory holes
More potholes await PC makers. Because of a packaging change, DDR2 won't fit on old motherboards. "The layout is different," said Elliot, a fact that's prompting hardware engineers to tinker with their designs.
Besides navigating the manufacturing and design issues, memory makers will also have to contend with the rapid and often seemingly illogical shifts in prices for which the memory market is known.
Right now, for example, 128-megabit SDRAM chips, made with an older version of memory that doesn't provide the same performance as DDR does, sell for more than the equivalent DDR chips. Why? So much manufacturing has shifted to DDR that SDRAM has become somewhat more scarce.
If most manufacturers execute the transitions flawlessly, which would increase supply, DDR2 prices could plunge prematurely. During various points in the recent past, memory makers had to sell fresh products for less than they cost to make because of oversupply.
Conversely, widespread problems could lead to high prices and a lack of demand from PC makers. The current high prices of DDR could also postpone adoption of DDR2.
"Unless demand forecasting is spot-on (and it never is), it is highly likely that we will see severe product mix issues," Gordon said.
The time to enjoy any price premiums will also be short. Currently, 512-megabit DDR2 chips sell for 84 percent more than 256-megabit DDR. The premium will drop to 24 percent by the end of the year and then disappear entirely by the end of 2005.
To ameliorate these factors, Samsung has instituted a program it calls common-die manufacturing, which allows the company to postpone the decision on what kind of chips to make. Generally, it takes about three months to go from a blank wafer to an individual memory chip.
Samsung and others have figured out how to delay the decision on what to use a wafer for -- DDR or DDR2 -- until the final four weeks of the process, Elliot said. In an industry in which customer contracts are renegotiated every two weeks, postponing the decision will let Samsung alter its output to better match demand.
Manufacturers will also heavily target the most profitable markets. "To succeed in DDR2, you need lots of sales into the server market," Elliot said.
