The Banana Peel Project

Today, IBM announced a breakthrough in processor design. If you don’t want to read through the entire press release, the basic point is that they’ve figured out a way to make wires that are short enough (we’re talking nanometers here) to connect computer chips that are stacked literally one on top of the other. Until now, the only way to fab chips (that’s fabricate, for those who aren’t geeky enough to spend a lot of time reading about this stuff) was to lay them out side-by-side, which not only takes up a lot of physical space but also presents tremendous synchronization problems when there are enough of them working together.

Before going into this, I’ll note as a disclaimer that a lot of the stuff I understand about supercomputing architecture and programming comes from the Web sites of the San Diego Supercomputer Center and Lawrence Livermore National Laboratory, both of which are amazing resources. Anything I don’t understand about microprocessing I’ll leave out rather than invent, or I’ll wait until I get through a few more chapters of my new microprocessing textbook. Maybe some day I’ll have a life.

When it comes to supercomputers, both physical space and synchronization are huge problems. Physical space is obvious — more space means more need for machine room space, more HVAC (heating, ventilating, air conditioning) problems , and more money. Synchronization is a programming problem, an architectural problem, and a connectivity problem. This is how I think about it: in our brains, our neurons typically have no problem operating in sync (except in the case of epileptic seizures, which can be partially controlled by forcing neurons to sync up). This is in part because our neurons fire and communicate with each other based on chemical reactions, which are very slow when compared to electronic pulses through silicon (at most once per millisecond, or a thousand times per second). The elegant synchrony of neurons is also helped by how densely they’re packed into our skulls. (Most estimates guess that we have around 100 billion neurons.)

Computer chips, on the other hand, are really, really fast. The chips that are typically used in modern supercomputers are capable of processing multiple data streams around 512,000 times per second. The San Diego Supercomputer Center’s DataStar system has more than 2,500 of those processors, but the most powerful supercomputers today have at least 10,000.

Our brains are highly specialized. No matter how hard you try, you are not going to be able to commit all of your neurons to trying to remember your girlfriend’s birthday. Many — maybe even most — of your neurons are involved in neural networks that allow you to smell, see, taste, breathe, walk, etc. You just can’t recruit these “specialty neurons” for things that they aren’t used to doing, at least not without a great deal of training.

Computer chips are totally unspecialized. Any chip can be used for any purpose at any time depending on the kinds of instructions you give it. For this reason, supercomputers can enlist every single solitary microprocessor to work together on a single task. But this is where the problem arises: take any two chips in the supercomputer array, and they are supposed to be working on the same task and contributing the the solution of the same problem, be it a Grand Challenge math problem or a complex 4D visualization. This means that they must operate in concert if the program is to assemble any meaningful information.

Computer chips communicate by electronic pulses through some sort of medium, usually copper or silicon wires. Those pulses, however, do not travel from one circuit board to another instantaneously. They are limited by many factors, including wire conductivity, input/output speed, frequency, and (last but not least) the speed of light. To synchronize chips that are located on opposite sides of the room, technicians need to set them out of sync by as much as several hundred nanoseconds. Ken Batcher, among others, has said that the limit of supercomputing power is not limited by computing power, but by how well and how quickly input and output occur.

OK, enough technical stuff. I get carried away. In short, the IBM announcement about three-dimensional stacked chips impresses me because it’s adds, literally, a whole new dimension to what we can do with parallel processing. The two-dimensional circuit board will soon be three-dimensional, which is going to fundamentally change how we talk about how well computers represent and simulate the real world.

And processors aren’t the only thing that went 3D this week:

• Researchers at Georgia Tech unveiled a new 3D solar cell, which they claim will “boost the efficiency of photovoltaic (PV) systems while reducing their size, weight and mechanical complexity.”
• Evan Malone, an engineering grad student at Cornell, invented the first desktop printer that actually prints things. Like model houses made of cheese. Yes, that’s right. Tea. Earl Gray. Hot. Check out the article at Ars Technica.

Onward.

[the banana peel project]