Paul,
The sure way to make RAMBUS a winner, however, is to move the RAMBUS DRDRAM Controller on to the CPU itself - to reduce latencies
Apropos your comment, was this in today's NY Times:
An excerpt:
the largest performance constraint in computer systems today is the mismatch in speed between the microprocessor and the slower memory chips. The Berkeley researchers predict that in the next decade processors and memory will be merged onto a single chip
Barry
-------------
nytimes.com
9chip.html
July 19, 1999
Chip Designers Search for Life After Silicon
By JOHN MARKOFF
It was a chance meeting between a self-described "physicist gone bad" and a chemist. And it may someday lead to the creation of a new breed of computers based on tiny electronic switches, each no thicker than a single molecule.
Three years ago Phil Kuekes, a Hewlett Packard physicist with three decades of experience designing computers, was pondering new ways to use a computer he had developed using massively parallel architecture -- technology that attacks computing problems by breaking them into hundreds or even thousands of pieces to be simultaneously processed by many individual chips.
At about this same time, Kuekes (pronounced KEE-kus) happened to make the acquaintance of James Heath, a chemist at the University of California at Los Angeles whose lab had been experimenting with tiny structures based on molecules of a synthetic substance called rotazane. It seemed that these molecular structures might be able to function as digital switches -- the basic, binary on-off information gateways of modern computing.
Soon the two scientists were brainstorming about how it might be possible to blend Kuekes' computer design with Heath's Lilliputian switches. In the fashioning of tiny switches, or transistors, from small clusters of molecules a single layer deep, the researchers see a coming era of computers that would be 100 billion times as fast as today's fastest PCs.
The work, detailed in a paper published Friday by the Hewlett-UCLA teams in Science magazine, is a noteworthy example of the groundbreaking research that suddenly seems to be flourishing at computer laboratories around the country -- a flowering of ideas that some leaders in the field have begun to consider a renaissance in computer science and design.
In corporate and academic labs, researchers appear to be at or near breakthroughs that could vastly raise the power and ubiquity of the machines that have insinuated themselves into almost every facet of modern society. And since many of these efforts, like the Hewlett-UCLA research, are focused at the microscopic level, computing could become an increasingly invisible part of everyday life.
"A lot of exciting stuff is happening outside the mainstream PC market," Marc Snir, a senior research manager at IBM Corp., said. "We are entering a world where there will be billions of small devices spread everywhere."
The Hewlett-UCLA team is just one of six groups around the country working under an ambitious program of the federal Defense Advanced Research Projects Agency that is trying to create a new kind of molecular-scale electronics -- known as moletronics -- that researchers hope will one day surpass the power and capacity of the silicon-based technology used in today's computers. Last year researchers at Yale and Rice universities took earlier steps toward the same goal of assembling computers from molecular components.
Meanwhile, separate from the military research program, researchers at the Massachusetts Institute of Technology's Laboratory of Computer Science are trying to meld the digital with the biological by "hacking" the common E. coli bacterium so that it would be able to function as an electronic circuit -- though one with the marvelous ability to reproduce itself.
It may all sound esoteric. And even the most enthusiastic researchers concede that practical applications of their theories and methods may be a decade or more away. But researchers seem intent on renewing the emphasis on the science in computer science, venturing beyond electrical engineering and physics to draw upon diverse disciplines like biochemistry as they form their hypotheses and test them with empirical evidence.
MIT's computer scientists, for example, are pursuing the idea of building -- or possibly growing, if the E. coli experiments pan out -- vast numbers of almost identical processors that might act as sensors or even as the task-performing devices called actuators.
"We would like to be able to make processors by the wheelbarrow-load," said Harold Abelson, an MIT computer scientist.
Abelson and his colleagues, who call their approach amorphous computing, are experimenting with the idea of mapping circuitry onto biological material. That might let living cells function, for example, as digital logic circuits. Such circuits are information pathways that, whatever their complexity, ultimately involve a multitude of quite simple, binary choices: 0 or 1, on or off, this but not that.
Biological cells, of course, would be able to compute only as long as they remained alive. But the premise is the same as in the molecular-scale work: Pack as many millions or billions of these tiny decision switches as possible into the smallest spaces conceivable.
The resulting "smart" materials might be used for new types of paints or gels, for example, that would enable a highway to be "painted" with computerlike sensors to issue traffic reports, or let floors be given a surface of infinitesimal actuators that could detect dirt and dust and silently and instantly whisk it away.
In the case of the Hewlett-UCLA work, researchers have successfully created digital logic circuits, but not yet any in which the molecular switches can be restored to their original state -- returning to the off position, for example, after having switched to on. And still to be developed are the molecular-scale wires that would be needed to interconnect the switches.
The really significant implication of the Science article is that for the first time researchers have built molecular-scale computing components using chemistry rather than the time-honored technology of photolithography, the ultraviolet-light etching of circuitry onto silicon that is the process for making today's chips. The chip industry has not yet reached the theoretical limits of photolithography, but the day may come when it is no longer possible to etch circuits any closer together. That is where molecular chemistry could take over -- and possibly wring more computing power from a single chip than exists today in the biggest, fastest supercomputers.
Last week, Kuekes said his team had been in contact with the MIT group and was now discussing the possibility of combining the Hewlett-UCLA molecular switching technology with the MIT lab's biological processor work with an eye toward future computer designs.
"Think of us as the Sherwin-Williams of the Information Age," Kuekes said, referring to the vision of suspending billions of tiny processors in a paint. "This is the raw material for super-intelligent materials."
The scientists acknowledge that their projects are gambles and that any practical applications may be a decade or more away. But the work on both coasts indicates the breadth of the renaissance now sweeping through computer-design circles.
To some extent, the most recent work is a continuation of efforts that began about five years ago and quickly grew into the commercial field of microelectromechanical systems, or MEMS chips. MEMS are microscopic mechanical structures etched into the surface of silicon chips; they have spawned a multibillion-dollar business, largely around the chips in silicon accelerometers that are now standard equipment as sensors for air-bag collision systems in cars.
But as the Hewlett-UCLA and MIT research makes clear, even the conventional circuitry of a silicon chip -- or the silicon, for that matter -- can no longer be assumed. Indeed, computer architects are rethinking the whole concept of the microprocessor, the master chip that begat the PC and that has been the building block of modern computers for a quarter of a century.
"It's time to do something different. It's time to look at the things we've ignored," said David Patterson, a computer scientist at the University of California at Berkeley.
In the early 1980s, Patterson helped pioneer one of the most significant computer design innovations of its era, a technology known as reduced instruction set computing, or RISC.
Building on ideas first advanced during the 1970s by a team of researchers working under the computer scientist John Cocke at IBM's Thomas J. Watson Laboratory, Patterson and his Berkeley graduate students proved that sharp increases in the speed of processor chips could be achieved by simplifying computer hardware and shifting many functions to software.
For almost a decade after the success of the Berkeley RISC project, many experts in the computer industry believed that RISC would ultimately displace Intel Corp.'s X86 computer chip design, on which the industry-standard PC was based.
Hoping to unseat Intel, dozens of new RISC-inspired companies sprang up from the mid-'80s through the mid-'90s. But beginning with its 486 chip, introduced in 1989, and continuing in its Pentium series, Intel was already incorporating the best RISC ideas into its chips, keeping their performance close enough to RISC's potential to let the company retain its commanding market lead.
The combination of Intel's hardware and Microsoft's software proved invincible, and one by one the RISC challengers collapsed. The end of the RISC era almost a decade ago left computer designers with few fundamentally new ideas for improving computer performance.
But even as the Intel juggernaut moved on, the rising importance of the Internet and a growing consensus that computing's future lies in inexpensive consumer-oriented devices helped fuel the renaissance in chip design.
Patterson, the RISC pioneer, has now embarked on a new design approach, known as intelligent RAM, or IRAM, that has generated great interest among consumer electronics companies.
RAM stands for random access memory, the semiconductor memory that is used as a kind of temporary scratch pad by software programs. Patterson and his Berkeley colleagues have noted that the largest performance constraint in computer systems today is the mismatch in speed between the microprocessor and the slower memory chips. The Berkeley researchers predict that in the next decade processors and memory will be merged onto a single chip.
The IRAM chips would embed computer processors in vast seas of memory transistors. Instead of stressing pure processor speed, these new chips would place the emphasis on avoiding bottlenecks that slow the data traffic inside a processor. Such an approach would be especially attractive to makers of memory chips -- companies eager to find new ways to distinguish themselves in what is now a commodity market.
At the same time, some consumer electronics companies are pursuing ideas similar to those behind the IRAM project. For example, Sony Corp.'s Emotion Engine chip, which the company is designing in cooperation with Toshiba Corp. for the coming Sony Playstation II game console, blends memory and processor logic as a way to create faster, more vivid on-screen game action.
But no single computer architecture is optimal for every kind of problem. That is why many researchers are exploring the idea of reconfigurable chips -- chips whose circuitry can be reorganized for each specific computing problem.
One of the most extreme efforts in this direction is a processor approach known as RAW, for raw-architecture work station, that is being pursued by a group of researchers at MIT's Laboratory of Computer Science, working separately from the E. coli project.
RAW pushes the idea of RISC to a new extreme. The processor would not have an "instruction set" in the conventional sense of the term, which refers to the types of instructions that let software programmers direct a chip to do things like add or subtract, or compare and move blocks of data. Instead of using preordained instruction sets, RAW would present the entire chip to the programmer as a veritable blank slate on which tens of millions of transistors might be individually programmed.
Such unprecedented flexibility would mean that the same basic chip might be used for an infinite variety of purposes -- like creating three-dimensional animated graphics, performing supercomputer calculations or carrying out tasks not yet conceived.
"We're trying to determine whether this is simply a challenge for programmers or a complicated nightmare," said Anant Agarwal, the MIT computer designer who heads the RAW project.
As all these research efforts proceed, computer scientists are toiling against the perceived limits of Moore's Law -- a guiding principle for chip design since the Intel co-founder, Gordon Moore, observed in 1964 that the number of transistors that could fit onto a silicon chip doubles approximately every 18 months.
But ultimately, there are physical limits to how closely together circuits can be etched with light onto the surface of a silicon chip and still function -- a reality that is expected to mean the end of the Moore's Law paradigm sometime around the year 2012. That is why forward-looking work like the molecular-scale circuitry research of the Hewlett-UCLA team is so important.
"Clearly a technology this radically different won't tip over a trillion-dollar industry" like today's computer-chip industry, Kuekes said. "But we're looking considerably ahead of where silicon runs out of steam." |
