Moore's Law Matures

First, a confession; for a while there, it was getting pretty boring covering the semiconductor industry as a technical editor. The new product introductions were so predictable, especially for heavily lithography-influenced chips such as memory and programmable logic devices. A conversion from 0.18 micron to 0.13 micron processing might result in….shock and amazement….a doubling of the memory capacity, or of the user-configurable logic gatecount. Stop the presses; I didn't see that coming from a mile (and a year) away!

The above-mentioned memory trends influenced other chips, too. A latest-and-greatest microprocessor might be nearly identical to its predecessor, aside from twice the on-chip cache, for example. And even non-memory-centric chips were slaves to predictability. For several iterations in a row, ATI and Nvidia's graphics processor generational jumps basically (I'm over-generalizing, but not too much) involved doubling or otherwise incrementing the number of available parallel vertex and pixel processing pipelines. And of course, whatever the chip's category, it almost always ran faster than its precursor, thanks to the Moore's Law corollary that shrunk-down transistors were also faster-switching transistors.

Leakage current sure put the kibosh on switching speed, didn't it? As you already know (unless, perhaps, you've been living in a cave), the past few years have seen the CPU suppliers switching en masse from a rabid focus on ratcheting up clock frequency to an equally rabid focus on ratcheting up the number of processor cores they can cost-effectively squeeze on one piece of silicon. More generally, I've noticed a dramatic transition away from brute-force 'throw more transistors at the problem' design methodologies towards a more elegant 'get the most return out of the transistor investment' approach.

Take, for example, Nvidia's GeForce 7800 GTX GPU, unveiled in mid-June at an ESP of $599 for a 256 MByte frame buffer-inclusive board. It's built on a 0.11 micron process, versus its 6800 Ultra predecessor's 0.13 micron foundation; a transition which, yes, enabled Nvidia to squeeze 302 million transistors onboard the 7800 GTX (and at roughly the same die size as the 6800 Ultra).And yes, it's got 8 vertex pipelines and 24 pixel pipelines, versus its predecessor's 6 and 16. And yes, it runs at 430 MHz core and 600 MHz memory clocks speeds, versus its predecessor's 400 MHz core and 550 MHz memory clocks (note, Nvidia has also just introduced a less-expensive variant called the 7800 GT, running at 400 Mhz core and 500 Mhz memory speeds, and with 7 vertex and 20 pixel pipelines onboard).

But consider that Nvidia touts (and hands-on benchmarks reinforce) that the GeForce 7800 GTX delivers twice the floating-point shader performance of the previous-generation chip. Minor increments in clock speeds, and a ~50% higher number of pipelines, aren't enough to solely account for this achievement. Nvidia fine-tuned its prior-generation texture and pixel (i.e. fragment) shaders' designs, after analyzing over 1300 of the most commonly used shader programs found in today's 3-D applications. Application profiling also revealed, for example, that Nvidia didn't need to boost the number of on-chip ROP (raster operator) units from 16 to 24 to match the increment in pixel shaders, so the company passed on what might have in the past been a knee-jerk transistor count up-tick.

Continued with 'Moore's Law Matures….More'….