Tera. They're talking TeraHertz, or actually Tera-operations-per-second. Potentially into the hundreds with the same technology, they hope. ("My CPU clock's 800Mhz." "Big whoop, mine's 8,000,000Mhz. I finish 400* work units of Folding@home per second." *Precise numbers extracted from my butt.)
I knew it obviously wasn't chemical, but I went and checked it out for you and it's more as I thought: almost mechanical. Mirror-like laser arrays. You don't change the "state" of the light any more than you change the "state" of electrons, you just move it around from place to place. Actually optical switches can be physically huge compared to electrical ones and still function much faster because the medium can move so quickly.
Since it's a private company that wants to sell technology it'll be harder to get the precise details out of them than out of the Technion Institute, for example. Besides, the technology is hard to understand to someone not well-versed in the relevant physical sciences. But their Executive Summary is quite revealing: (skipping the introduction and straight to the meat)
Quote:
The Architecture
The heart of the ODSPE architecture is an optical Vector-Matrix Multiplication core. It is the key to massive data parallelism and on-the-fly reconfigurability. The optical VMM core converts electronic inputs into light by using Vertical Cavity Surface Emitting Lasers (VCSELs) and then performs a VMM operation by directing this light through programmable internal optics. The light that emerges is sensed by an array of detectors and translated back into electronic signals. By processing incoming signals at the photon level, Vector-Matrix Multiplication is performed at the speed of light.
ODSPE Architecture
The optical core has an input array of 256 Vertical Cavity Surface Emitting Lasers, as well as a Spatial Light Modulator (SLM) that is 256 x 256 in size and a Photo Detector Array (PDA) of 256 detectors. Each one of the VCSELs shines on one column of the SLM, and each one of the detectors collects the energy from one row of the SLM. As the light transmitted by a VCSEL passes through an SLM element, its intensity is attenuated by the transmissivity level of that SLM element, an effect equivalent to multiplication. Each one of the VCSELs shines on one column in the SLM, and each one of the detectors collects the energy from one row in the SLM yielding a Vector-Matrix Multiplication operation.
During each cycle, an input vector of 256 one-byte wide elements is multiplied by a matrix of 256x256 one-byte wide matrix elements (64K elements in the matrix), generating an output vector of 256 one-byte wide elements. This is the key to the massive data parallelism of the ODSPE. At the initial product with transform rate of 125 Megahertz and 64K MAC (Multiply-Accumulate) equivalent operations in the matrix yields a total performance equivalent to 8 Tera MAC operations per second.
|
Of course, I didn't know what a VCSEL was, so I looked it up and this site gives a brief, readable description.
The speed, as you can see in the quote above, isn't all due to using light instead of electricity to transfer information, but also because they exploit a multiplication matrix instead of repeated individual multiplication operations which allows great parallel processing - each clock cycle here would be essentially equivalent to about 256 clock cycles in a linear DSP, if I'm reading right.
My recommendation? Buy stock in this company when you can, and consider it a long-term investment.
But the development of the DNA transistor has been compared to the invention of the transistor as a breakthrough in computers. Let's just see whether it takes as many decades to come into widespread use as the silicon transistors did.