Mixed Signals
Any sufficiently advanced information is indistinguishable from noise
Thursday 12 April 2007, 6:10 PM
Stacks of heat...
IBM has announced that it can stack chips. Building 3-D circuits by putting chips directly on top of each other is something lots of people have been looking at for a long time - I can remember getting excited about it in the 1980s - but to no good effect. IBM's breakthrough, the company says, is that it can make the interconnections using standard chip fab techniques, meaning you can have a lot more inter-chip links running a lot faster than before, but costing effectively nothing.
Which is good, but by no means the only problem to solve. The big one is power and its most common incarnation - heat. It's hard enough to keep one high performance chip cool when you can mount it directly next to a heatsink: how will that work when it's mounted between two other high-performance chips? Anything extra you add to improve cooling is going to negate the advantage of IBM's fabby new interconnects. And if you decide to limit the technique to slower, cooler chips... well, we could do that before with the other 3D interconnect techniques, and nobody really found it worthwhile.
Doubtless IBM has a solution. But it's not mentioned in the press release, which rather gives the impression that IBM has now gone "from lab to fab". It does indeed promise product later this year - but only for "wireless communication chips that will go into [radio] power amps", where the technique is probably best suited for reasons of short line lengths running in a relatively low power configuration. The really exciting stuff it's talking about, coupling fast processors to memory - is going to be... challenging.
So, interesting. But not worth the universal uncritical response that seems the consensus elsewhere in the media. More details needed!
(I would make the call and talk to IBM myself, except we're still unpacking after our office move and I'm off to Beijing tomorrow for Intel's first Chinese IDF. More on that elsewhere...)
Which is good, but by no means the only problem to solve. The big one is power and its most common incarnation - heat. It's hard enough to keep one high performance chip cool when you can mount it directly next to a heatsink: how will that work when it's mounted between two other high-performance chips? Anything extra you add to improve cooling is going to negate the advantage of IBM's fabby new interconnects. And if you decide to limit the technique to slower, cooler chips... well, we could do that before with the other 3D interconnect techniques, and nobody really found it worthwhile.
Doubtless IBM has a solution. But it's not mentioned in the press release, which rather gives the impression that IBM has now gone "from lab to fab". It does indeed promise product later this year - but only for "wireless communication chips that will go into [radio] power amps", where the technique is probably best suited for reasons of short line lengths running in a relatively low power configuration. The really exciting stuff it's talking about, coupling fast processors to memory - is going to be... challenging.
So, interesting. But not worth the universal uncritical response that seems the consensus elsewhere in the media. More details needed!
(I would make the call and talk to IBM myself, except we're still unpacking after our office move and I'm off to Beijing tomorrow for Intel's first Chinese IDF. More on that elsewhere...)
Previous
