- Jul 27, 2002
- 13,312
- 687
- 126
I ran into this short but intriguing idea while reading @Beyond3D and I thought I'd share it with my fellow AT'ers. It's a very sharp thinking, when 'Performance/Watt' seems to become the most important thing in tech scenes for the past couple of years. While I don't disagree that high performance and low power usage is a good thing, but it is always interesting to look deeper to see by who, when, for what this new 'discourse' is being driven.
The original thread can be found here: http://www.beyond3d.com/forum/showthread.php?t=39363
If you want low power consumption, buy a laptop or an old DX5 GPU. Higher performance (for irreversible computations) inevitably requires more power (Landauer's principle). Simply plot the performance vs power requirements for either CPUs or GPUs over the last number of years. So if you want faster and faster GPUs, you need to get comfortable with continually higher power requirements.
The push by chip vendors to make chips more power efficient is simply an economic concern from their part. When power consumption was relatively low compared to other equipment or parts of the system, costs were mainly silicon driven. Now that performance has scaled to the level that the costs of acceptably quiet cooling solutions are becoming a significant part of overall delivery costs, then making chips more power efficient becomes more of a priority, with a tradeoff between more complex silicon design and more die area given toward power efficiency.
When chips were silicon constrained, the priority was given to minimizing die area and maximizing freqency to maximize the performance/die area ratio and thus maximize the performance/cost ratio. As chips become power constained, you maximize the performance/cost ratio by maximizing the performance/power ratio instead. This leads to the rather interesting reversal of emphasis, since increasing the power efficency means reducing the frequency and increasing the die area (more transisters running at lower frequencies). This is because power scales up nonlinearly with frequency but performance only scales linearly with frequency. However, both power and performance scale linearly with transistor count. As a result, in the future you will see more and more chips primarily increase performance by dramatically increasing transitor count rather than frequency.
From a buyers point of view, the main differences are total cost of the equipment as well (the cost of the power differences are small). The other considerations (noise, etc.) all have solutions that simply add to the cost.
So if you had two machines available that had the same quietness, the same features, the same reliability and the same performance, etc. but one was 25% less expensive, most would choose the less expensive one. 12 months later, if they took their machine apart, examined it with a volt meter and noticed a sophisticated heat exchanger mounted on the chip, and realized that it had a higher power consumption than the other machine that they didn't buy, would they care much?
The original thread can be found here: http://www.beyond3d.com/forum/showthread.php?t=39363