Original URL: https://www.theregister.com/2007/05/04/sun_proximity_communications/
Sun's chip gurus theorize about obliterating IBM and Intel
Nearing proximity
Posted in Systems, 4th May 2007 19:47 GMT
When you have companies such as IBM and Intel looking to destroy your business, it's nice to have a fella like Ivan Sutherland stored away in a back room.
Sutherland, considered the father of computer graphics, developed a method of linking processors together called face-to-face computing, as part of his work at Sun Microsystems Labs. Over the years, Sun has changed the name of Sutherland's technology to “proximity communication” and has inched closer to rolling it out in actual product. Should the proximity communication bet pay off as expected, Sun could enjoy a massive performance boost over rivals in the server game and possibly in areas such as networking as well.
Sun claims to be doing nothing short of rewriting Moore's Law with the proximity communication technology.
The main concept behind proximity communication proves easy enough to grasp, as you can see from the diagram. You have downward facing chips that sit on top of upward facing chips, allowing the two sets of chips to make a direct, electrical connection. This removes the need for wires between chips and awkward multi-chip modules employed today by some vendors.
We turn to one of Ivan Sutherland's face-to-face patents to explain some of the drawbacks associated with wiring chips in today's configurations.
This elaborate arrangement of connectors from one chip to another has two drawbacks. First, it is costly. There are many parts involved and many assembly steps to put them together. The steps include making the packages, installing the integrated circuit chips in them, bonding the pads of the integrated circuit to the conductors in the package, and fastening the packages to the printed circuit board. Although each of these steps is highly automated, nevertheless they remain a major cost factor in many system designs.
Chip Fornication
Second, it is electrically undesirable. The wires on the printed circuit board are about 1000 times as large as the wires on the integrated circuit. Therefore, to send a signal from one integrated circuit to another requires a large amplifier on the sending integrated circuit. Moreover, the conductors involved have a good deal of electrical capacitance and electrical inductance, both of which limit the speed at which communication can take place. Perhaps worst of all, much energy is required to send a signal through such large conductors, which causes the driving integrated circuit to dissipate considerable power. The cooling mechanisms required to get rid of the resulting heat add cost and complexity to the system.
Several methods have evolved to improve chip to chip interconnect. One way is to avoid several packages for the separate integrated circuits. Instead of a package for each circuit, several chips are mounted in a "multi-chip module," a kind of communal package for the chips. The multi-chip module (MCM) contains wiring that carries some of the chip-to-chip communication circuits. The size of the wires in the MCM is smaller than the wires on a printed circuit, but not yet so small as the wiring on the chips themselves. Electrical capacitance and inductance in the wires between chips remains a problem even in MCMs.
To pull off Sun's proximity communications arrangement, the company must find a way to align the chips for capacitive coupling with 15 microns of accuracy for top performance. Thus far, that has proved pretty tough, although Sun is making strides in the right direction.
Sun's researchers, for example, showed off a prototype switch last week, during a Labs open house for reporters.
The switch code-named Sedna relies on two logic chips linked to three so-called bridge chips with the proximity communications technology.
“One benefit is the I/O bandwidth increase that we get from this technology,” said Sun researcher Hans Eberle. “We achieve up to 10 terabits per second per square mm of chip overlap.”
Sun's prototype system is based on the ACTA telco standard and has four 10GigE ports and layer 2 switching. It works with off-the-shelf line cards and Sun's proprietary switch blade.
Should all go according to plan, Sun could build a flat, single-stage switch fabric capable of scaling to thousands of ports with multiple Tbps of throughput.
(It should be noted that, er, the switch wasn't in working order when we saw it and that Sun reckons a really solid test machine won't arrive for 18 months or so.)
A few microns between victory and defeat
As mentioned, Sun is still concerned about the packaging issues related to the proximity communications technology.
On the plus side, Sun's technology would allow for less wires, which would translate into smaller chips. In addition, Sun would not need the ceramic and plastic packaging used with today's chips. So, it could remove faulty chips or insert faster chips with relative ease.
On the downside is the alignment issue and concerns about removing heat form the systems.
Sun had promised the proximity communications as part of its bid for a lucrative DARPA high performance computing contract. That promise included linking second generation Rock processors with the technology. By our estimates, that puts Sun three to four years away from using proximity communications. It would also seem to indicate quite a bit of confidence on Sun's part for the technology.
One researcher described this technology as "a really big deal” for Sun. The company has Sutherland's sweeping patents and some patents purchased from MIT researchers backing up its attack. So, if the technology works as hoped, Sun would enjoy a big time lead over rivals.
The same researcher confessed that Sun could find itself well behind the competition if the proximity communications technology fails. Sun would need to catch up with IBM and others' multi-chip module expertise in a hurry.
That Moore's Law thing
In the best of all possible worlds, Sun thinks that proximity communications technology could take some pressure off chip makers.
Moore's Law gives us more transistor density every 18 to 24 months. Keeping up with the law requires chip makers to spend billions of dollars and seemingly rewrite the laws of physics ever two to four years.
Sun argues that Fabs could worry less about, say, what follows 45nm and focus more on linking 45nm parts together with proximity communications to keep up with Moore's Law.
"Instead of moving along with the Fab technology, you can stay at 45nm and just put more silicon in your system,” said Robert Drost, a Sun researcher.
We see Drost's point. Intel, however, would certainly argue for adding more transistors per chip at the same time as linking chips together. But then, Intel is so damned ambitious.
Sun also plans to make use of technology such as 3-D memory and optical networking with its proximity technology-based parts. The company expects to reveal more about its packaging progress next year. ®
Bootnote
A number of other research projects caught our eye while at Sun.
Coders will want to check out Fortess - a programming language for high performance computing applications. This work also grew out of Sun's failed DARPA bid and should help developers write multi-threaded code.
The Celeste peer-to-peer storage project seemed moving as well, although there was no one around at the booth to tell us about the work. Tsk, tsk.
And then there was the transactional memory stuff that went way over our head but is "very important," according to some smart folks.