IBM Discovery May Lead To Exascale Supercomputers 135
alphadogg writes "IBM researchers have made a breakthrough in using pulses of light to accelerate data transfer between chips, something they say could boost the performance of supercomputers by more than a thousand times. The new technology, called CMOS Integrated Silicon Nanophotonics, integrates electrical and optical modules on a single piece of silicon, allowing electrical signals created at the transistor level to be converted into pulses of light that allow chips to communicate at faster speeds, said Will Green, silicon photonics research scientist at IBM. The technology could lead to massive advances in the power of supercomputers, according to IBM."
Re:GPU = supercomputer? (Score:5, Informative)
GPUs are indeed an inexpensive way to boost speed in some cases. But they have been rather oversold; while some specific types of problems benefit a lot from them, many problems do not. If you need to frequently share data with other computing nodes (neural network simulations come to mind), then the communications latency between card and main node eats up much of the speed increase. And as much of the software you run on this kind of system is customized or one-off stuff, the added development time in using GPUs is a real factor in determining the relative value. If you gain two weeks of simulation time but spend an extra month on the programming, you're losing time, not gaining it.
Think about it this way: GPU's are really the same thing as specialized vector processors, long used in supercomputing. And they have fallen in and out of favour over the years depending on the kind of problem you try to solve, the relative speed boost, cost and difficulty in using them. The GPU resource at the computing center is used much less than the general clusters themselves, indicating most users do not find it worth the extra time and trouble to use.
It is a good idea, but it's not some magic bullet.
Re:Who would've thought... (Score:4, Informative)
Not sure this guy understands the problem. (Score:4, Informative)
He's sped up links between chips from something like one-third c to c.
Architecturally that reduces inter-chip latency by 66%, which does indeed open up a new overall speed range for applications that are bandwidth-limited by interconnects. But in no sense does it imply a 1000-fold increase in overall performance. It's only a 3X improvement in bandwidth of the physical layer of the interconnect to which the speedup applies.
It may allow architectures that pack in more computing units, since light beams don't interfere physically or electrically the way wires do. And light can carry multiple channels in the same beam if multiple frequency or phase or polarization accesses can be added. Those will further improve bandwidth and possibly allow a further increase in the number of computing units, which could help get to the 1000X number.
BTW, didn't Intel have an announcement on optical interconnects just a while ago? Yes. [intel.com] They did [bit-tech.net].