Hynix is shaking things up in the memory market with its decision to license Innovative Silicon's Z-RAM technology.
The two companies jointly announced the agreement on Monday.
Z-RAM is a twist on the traditional makeup of a memory cell. Almost all PCs use DRAM to temporarily store information while the system is running, to avoid delays accessing that data from the hard drive every time it is needed. And each DRAM cell needs a transistor and a capacitor, which stores electrical charge, to represent a bit of data. But Innovative Silicon figured out a way to take advantage of the "floating-body effect" to make a transistor with the capacitor built right in there, reducing the size and complexity of a memory cell.
That results in a memory cell that is easier to scale into the future, as chip companies work to keep Intel co-founder Gordon Moore honest, and that also uses less power than a conventional DRAM cell. AMD already has a license for Z-RAM for use as the cache memory in microprocessors.
Hynix didn't say when it plans to have Z-RAM-based computer memory ready for sale, but it's likely to take some time as the undulating DRAM market rolls forward over the next several years. Perhaps this would also make it cheaper for PC companies to put more than 4GBs of memory into their systems, finally kicking off the 64-bit era of personal computing.
SANTA CLARA, Calif.--Flash memory will take over the world, if the laws of physics don't get it first, according to one of the technology's biggest backers.
On one hand, NAND flash--the kind of flash found inside digital cameras and MP3 players--is crushing the competition, Eli Harari, CEO of SanDisk, said on Wednesday at the Flash Memory Summit here.
Eli Harari, CEO, SanDisk
(Credit: SanDisk)NAND has made 1-inch hard drives obsolete. Then it's on to 1.8-inch hard drives and possibly larger 2.5-inch notebook hard drives, Harari said.
"The next big market will be video. You should expect next year to see video cameras and video camera phones that use more flash," he said. More flash notebooks will start appearing next year too.
NAND will also start replacing DRAM, the main memory used in PCs, over the next five to seven years. Because of rapid improvements, NAND has gone from being more expensive--in cost per memory bit--than DRAM in 2001 to being significantly cheaper than DRAM in 2007.
IBM and others are already working on solid-state servers that rely on NAND.
Now the bad news: it's going to become much harder to continue to improve NAND chips.
"NAND is definitely going to slow down. There are some fundamental practical limits and some fundamental physics limits," Harari said. "We have three or four generations, five at most. Beyond that, we may need a different technology."
The problem is that the different subcomponents inside these chips can't be shrunk much further. One layer, called a tunnel oxide, can probably be shrunk to about 80 angstroms. In other types of chips, this layer can be shrunk to 12 angstroms.
Data corruption is another problem. Flash memory records data by storing electrons in cells. "At 32 nanometers, you can't afford to lose more than 30 electrons" before data corruption becomes an elevated risk, he said. "Thirty electrons sounds like a lot, but it is very small." The debut of 32-nanometer chips will be sometime toward the turn of the decade.
The turning point may come with the 20-nanometer generation of chips, which will start coming out a few years later. At that point, NAND chips will be capable of holding 256 gigabits of data, and the cost per bit will be about a tenth what it is today, but it will be extremely difficult with current manufacturing technologies and materials to go further.
What are some of the alternatives? SanDisk will try to stack transistors in 3D arrays. "When land is expensive, you go up in skyscrapers," he said. The company bought Matrix Semiconductor, which invented a 3D memory chip, a few years ago. Matrix's chips, however, are not rewritable--you can't erase old data and put new stuff on it. Solving that problem will be a focus of the company's research.
Manufacturers will also put more bits in each cell. Right now, the densest chips have two memory bits in each cell. Chips with three bits per cell will probably come out in the next two to three years, and four bits after that. To date, most manufacturers haven't talked much, if at all, about three bits of memory per cell.
Other companies are also working on things like phase change memory, or mechanical actuators.
Cost will become another major issue. Fabs cost about $5 billion, and nearly 50 percent of the world's NAND factory capacity will have to be replaced in the next two years, Harari said. NAND makers thus will have to put a lot of money into construction.
The new fabs will let the most efficient manufacturers cut prices by 30 percent to 45 percent a year, he said. Unfortunately, prices have declined by 60 percent annually in the last two years.
"If the best you can do is 40 percent cost reduction, you can't have 60 percent reductions in price," he said. Consolidation may become inevitable.
Memory makers, of course, have overcome or at least postponed these problems before. Back in 2003, flash executives said the industry would start running out of tricks with 45-nanometer chips--which will start coming out in the relatively near future.
Making a supercomputer used to require teams of dedicated scientists, millions in federal research grants, and lots of specialized components that took years to design.
Thanks to clustering and other advances, a group of well-trained grad students can build one from off-the-shelf parts. As a result, the rankings in the Top 500 Supercomputers list changes more rapidly than the standings on Dancing with the Stars.
So with commoditization, what's the most expensive thing? The cooling system? The processors?
Weirdly, it's the memory, says Andy Bechtolsheim, senior vice president of Sun Microsystems' systems group and a co-founder of Sun, who revealed the company's Constellation System this week.
"The dominant cost of the systems is memory, DRAM, believe it or not," he said. A typical supercomputer will require millions of DRAM chips. DRAM drops in price like crazy--the average selling price of DRAM dropped 35 percent from December to April. But millions of chips are millions of chips.
The cables aren't cheap either. In fact, they cost more than the silicon inside switching systems.
Electricity is a problem, too.
"The cost of electricity is now showing up in the bids," he said. It varies by geography. Sandia's Oak Ridge lab in Tennessee has an advantage in this area because it's located in a hydroelectric hotbed.
- prev
- 1
- next
