Rambus shines in Intel-Micron deal

Rambus--which designs high-speed memory chips--will increasingly move to center stage in the world of computer design.

Michael Kanellos Staff Writer, CNET News.com
Michael Kanellos is editor at large at CNET News.com, where he covers hardware, research and development, start-ups and the tech industry overseas.
Michael Kanellos
5 min read
With Intel's $500 million investment in Micron Technology, memory designer Rambus will increasingly move to center stage in the world of computer design.

Rambus--which designs, but does not make, high-speed memory chips--is the premier company working on a solution to what is perhaps the biggest bottleneck in computer performance today. Namely, the speed at which data can be delivered to the processor.

Processors are getting faster every quarter, but the speed at which they get data is not. The resulting gap often means that a high performance processor is idling as it waits for memory to come from the processor, say analysts and chip designers, which means servers and desktops aren't working at peak performance.

Memory based on the Rambus design, called RDRAM, will reduce the time gap, because it gives memory makers a blueprint for making DRAM chips that can deliver up to 10 times the amount of data per processor clock cycle as current DRAM chips at a faster rate.

While other companies are working on high-speed memory designs, Rambus effectively has become the market leader. Not only have most processor vendors said they will make chips and chipsets that will "speak" Rambus, the Intel investment in Micron helps ensure that there will be an adequate supply of the high-speed chips starting toward the middle of next year. (See related story.)

Why is the industry moving from today's DRAM technology to RDRAM? Constrained by the laws of physics, current DRAM just can't run much faster or send data through a pipeline that's much wider.

RDRAM, on the other hand, runs eight times faster than current high-speed DRAM by imposing tight constraints on memory connection hardware.

But though RDRAM comes closer, it still can't keep up with today's processors.

"They are all fairly similar [the competing high-speed memory designs]. The key thing is Intel's endorsement," said Linley Gwennap, vice president of MicroDesign Resources. "Sixteen of the top 16 memory makers have all committed to Rambus.

"Microprocessor designers are realizing more and more that faster and faster CPUs are not going to be the answer unless you address the memory bottlenecks," he added.

The Micron investment is essentially designed to ensure that Rambus' designs work to Intel's advantage, according to some observers. Intel appears to be one of the more aggressive companies when it comes to adopting Rambus. At the same time, adopting Rambus will give Intel a way to boost desktop performance, and hence market share.

"Intel needs Rambus to happen at a reasonable price in a timely manner," said Jim Handy of Dataquest. "[RDRAM] is key to keeping Intel's processor speeds ahead of their competition."

"The Intel investment is primarily to assure some minimum level of supply to Intel for these memories," said Dean McCarron of Mercury Research.

A vexing problem
Memory latency, or the time gap between when the processors needs the information and when memory can deliver it, has been a continuing problem since the 1960's, said Nathan Brookwood, semiconductor analyst with Insight 64. Processor designers have circumvented the problem to a certain degree through caches--discreet reservoirs of memory located close to the processor that sit between the processor and main memory. Over the years, caches have not only grown larger, they have moved closer to the processor and increased in number. Intel's upcoming Merced processor will contain three caches, he noted.

Still, a tremendous gap exists, he said. Commercially available processors run as fast as 600 MHz. Memory, however, runs at 100 MHz.

"You give the memory an address, and it may pop back with a response in 40 nanoseconds," Brookwood said. "A processor running at 500 MHz would like to get something in two nanoseconds."

Rambus has designed a memory technology that leaves behind many of the technical limitations of today's prevailing memory technology, SDRAM (synchronous dynamic random access memory).

Widening the bus
There are two ways to get information out of RAM faster, McCarron said: Either speed up the RAM itself or widen the channel, or "bus," that communicates with RAM.

But SDRAM can't be sped up much more. The most recent switch from 66 MHz to 100 MHz was a difficult engineering feat. The only avenue remaining was to widen the bus to SDRAM chips--currently up to 64 and even 128 bits.

Rambus has taken a different tack, though, by focusing on speed improvements. Rambus chips currently operate at a speed of 800 MHz, but the bus is a relatively narrow 16 bits.

With current technology, that gives Rambus a two-to-one advantage over SDRAM. Rambus' technology allows a peak bandwidth of 1.6GB per second, compared to 800MB per second for SDRAM.

But while Rambus' RDRAM technology is an improvement over SDRAM, it's still not fast enough to keep up with processors of today and tomorrow. "Rambus only catches up a little bit," McCarron said, so processor designers still need increasingly large high-speed memory caches that can respond more quickly to the processor's demands for data.

Different vendors are taking independent approaches to supporting Rambus. Earlier this week, Compaq said it would incorporate a Rambus memory controller into its 21364 Alpha processor coming in mid-2000. By incorporating the memory controller, Compaq has effectively set out on a course where all of its Alpha-based servers will depend upon Rambus-style memory.

In addition, National Semiconductor will embed a Rambus controller into its "Jalapeno" desktop chips coming in 2000.

Intel will focus on chipsets
Intel will not integrate a controller into its chips, said a company spokesman. Instead, it will concentrate on chipsets that support Rambus standards. The first chipsets will come out in the fall of 1999, said a company spokesman, although observers said it could occur earlier.

Intel may also support other high speed memory standards, but "this is the only one we are being pro-active on," he said.

For computer vendors, the fact that Intel is switching to Rambus technology is at least as important as any of Rambus' technological merits, said Mike Fiebus, another analyst at Mercury.

"You're going to use Rambus, or you're going to have to jump through some hoops. That, in a nutshell, is why you'd better use Rambus."

Intel is an investor in CNET: The Computer Network, publisher of News.com.