So Many Transistors, So Little Time -

So Many Transistors, So Little Time

ATI's Radeon 9700 graphics processor sports something like 110 million transistors, or twice what you'll find in a Pentium 4. Nvidia's GeForce4 Ti 4600, by contrast, has a paltry 63 million transistors.

Let's just skip right over the peculiar reality that it's games rather than, say, CAD apps that drive graphics, and that we have, as Jim Turley acerbically commented, “descended into a world where high-end computers are just adjuncts to games.”

What is amazing is the sheer number of transistors that you can pack onto a die the size of your fingernail. It's not so amazing why you would want to do so. You reduce the number of interconnects on your assembly, so you get better reliability. You lower power consumption, and best of all, you reduce cost, well except for nonrecurring engineering costs, which unfortunately skyrocket.

Is this a bandwagon onto which everyone ought to jump? Are we moving toward a world in which everything is a system-level ASIC? I don't think so. The cost for a mask set in a leading-edge technology is well over a million dollars. If the device is not perfect, and you have to do a respin, the dollars just keep flowing.

It's hard to doubt that the future is system-level silicon. The question is, how do you engineer it so that the broadest range of systems companies can participate? Programmability has always been a solution — that's what led to the popularity of microcontrollers in electronics systems in the first place. Throw in a microcontroller and you eliminate all the parts that make up a hard-wired controller.

Will FPGAs with processor cores and vast numbers of configurable logic blocks be the platform of choice for all but the highest volume products? Maybe after several generations. Unfortunately, FPGAs are fundamentally more expensive to manufacture than mask-programmable devices, and they consume more power.

Xilinx recently introduced an 8-million gate FPGA that can be clocked at over 400-MHz and features I/O data rates of up to 840-Mbit/s. It's based on a 0.15-micron process technology. According a story in EE Times, the company won't quote a price on the device now, but in the second half of 2003, when it's being shipped in volume, the price is expected to be $3,960 in quantities of 10,000. That's still a little rich for deployment in a microwave oven, PDA, or sprinkler system. But what happens down the road when 0.15 micron technology is trailing edge?

If Xilinx's approach is not economical for production volumes of system-level devices, watch for other solutions from companies like Ubicom, which favors a “software system-on-chip” approach and Quicksilver Technology, which is exploring adaptive computing.

One can only hope that a technology will prevail that features reasonable development and unit costs and allows a broad range of companies to participate in system-level chip design.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.