DRAM debate

DRAM debateIntel has invested in Rambus DRAM manufacturers to help the high speed PC memory technology succeed. But other firms prefer the alternative – namely double data rate DRAM. David Manners reports
Last year PC microprocessor speeds went from 233MHz to 450MHz. If they do the same this year, we’ll have gigahertz PCs for Christmas. Do we want them?
Intel, unusually, seems not entirely sure. While ostensibly backing its vision of 600MHz-800MHz Rambus-based machines expected to cost between $1,500 and $3,000 in the second half of the year, it is hedging its bets by chopping the price of Celeron microprocessors to very low levels for sub-$1,000 machines.  
  New era… Rambus is a new way to drive PC memory.
400MHz Celerons have been introduced at $170 and 333MHz versions are under $70.
But Intel’s Celeron strategy looks like Plan B, a fall-back position if Plan A – the Rambus high-end vision – goes astray. Certainly Plan A is being very expensively promoted.
Intel has invested $500m in Micron Technology and $100m in Samsung to persuade them to keep making Rambus DRAMs – an unusual procedure reflecting the great distaste for Rambus among DRAM manufacturers.
DRAM manufacturers dislike Rambus because they will have to pay royalties on the chips they sell and will have to make larger die which are less economic.
The DRAM companies are already pushing hard for an alternative to the Rambus model by promoting the PC133 double date rate (DDR) DRAM which, they argue, is a less revolutionary approach than Rambus.
DDR – an evolution from synchronous DRAM – delivers two data words per clock cycle. The advent of PC133 is a consequence of moving to sub-0.2?m processing which produces a larger proportion of 133MHz die.
IBM and Silicon Graphics have said they’ll be using DDR in workstations and servers – a strategic position from which IBM could move DDR into PCs, while JEDEC has agreed a DDR standard so giving impetus to DRAM manufacturers to combine to push the technology.
The higher frequencies of Rambus are thought to involve problems with impedance control and extra shielding from radiation. The advantages of DDR over Rambus are better latency and better bandwidth per module, lower device cost because the die size is smaller than the Rambus die size, and a lower adoption cost in putting it into equipment because it is an evolutionary step, not a radical change like Rambus.
“Rambus is a new way to drive memory. That hasn’t changed for twenty years so people are uncomfortable with it,” says Helmut Schock, Toshiba’s European memory boss.
So this year’s split in the PC architecture could see the low-end PC evolving along the PC133 route, and the high-end going with Rambus.
“Some people like the PC133 option because they don’t want to go to Rambus/Intel,” says Franz Exenberger, marketing manager for memory at Hitachi Semiconductor.
Intel is not supporting 133MHz DRAM by producing a chipset, because that would interfere with its Rambus vision. However Reliance of Taiwan is already sampling a PC133 chipset, while Standard Microsystems Corp, Acer Laboratories, Silicon Integrated Systems and Via Technologies are all developing versions of the chipset.
Intel says that its Rambus 820 Camino chipset, due in Q3 ‘99, will also support a 133MHz system bus.
“PC100 is very widely used and moving from PC100 to PC133 is not so difficult – no big design changes are needed – but Rambus is much more complex. I see PC133 in the low to medium end PC segment,” says Toshiba’s Schock.
Whether Rambus or DDR wins in the market depends, eventually, on what people want.
If people want PCs to provide e-mail, word processing and Web browsing, then they can be made very cheaply or even incorporated into cheap consumer items called ‘information appliances’ or integrated into TVs, set-top boxes, or even telephones.
However, if PCs are to handle multi-tasking, manipulate digital pictures and generate large amounts of graphics needed for games playing, if they are to handle new features such as videocommunications, speech recognition and speech control, voice processing and even language translation, then gigahertz processors will be needed.
For these people the Intel vision of the future will work out with Rambus PCs boasting a data transfer rate between the microprocessor and the DRAM of 1.2Gbits per second (at 600MHz) and 1.6Gbits/sec per second at 800MHz, these machines will make an appreciable difference to the running of graphics-intensive or multi-tasking applications.
But for people who want only e-mail, Web browsing and word processing, micro-DRAM transfer rates of 1.6Gbits/sec will not make much difference to the perceived performance of their applications.
So consumers may go for radically different types of PC, and the interest in 1999 will come from whether there’s going to be a basic division in the way PCs are made – whether, after twenty years of architectural uniformity, a split could open up between the high-end and the low-end PC.
While Intel intends to dominate both ends, the other x86 microprocessor suppliers – IBM, AMD, National Semiconductor and IDT – will have an opportunity to take control of the low-end, but only if they can get their production volumes up.

Leave a Reply

Your email address will not be published. Required fields are marked *