vetta project

I was just looking at the famous 1997 article When will computer hardware match the human brain? by Hans Moravec. The article made bold predictions about what the future would be like based on the continued exponential growth in computer performance. As my own interest in AGI was founded on these ideas, and still is, it is interesting to look back to see whether the last 10 years have followed, or diverged from, what was predicted.

Moravec predicts that the world’s biggest super computers in 2007 would achieve 10¹⁴ MIPS and that this would be “more than enough” to mimic the human mind. The first part is definitely true with BlueGene/L rated at 2.8 x 10¹⁴ FLOPS. As for the second part, clearly these machines are not currently mimicking the human mind — though this doesn’t prove that they couldn’t. Maybe with the right algorithm BlueGene/L would have human level intelligence? Who knows.

Let’s focus on his computer performance predictions then. This is summarised in the “Evolution of Computer Power/Cost” graph half way through the article. If we read off the 1995 trend line, which seems reasonable as it was the most up-to-date trend at the time of the article, we see that in 2007 $1000 dollars should buy about 50,000 MIPS. If we equate an instruction with a floating point calculation, this is after all a log plot, then that’s a prediction of about 50 GFLOPS per $1000.

If we look at the folding@home statistics we see that a Playstation 3 gets about 25 GFLOPS, and costs around $600, which is about 40 GFLOPS per $1000. For GPUs the average is about 60 GFLOPS. With a typical GPU graphics card costing around $250, if you plugged this into a $1000 PC you would get about 50 GFLOPS per $1000. For the Windows, OS X and Linux PCs the numbers are significantly worse, however it’s not clear just how old most of these computers are (unlike the PS3s which are very new), and thus it’s hard to estimate their average worth. I estimate that a decent dual core PC should be able to do about 5 GFLOPS per $1000 when not using the graphic card. For a very high end graphic card such as an nVidia 8800 GTX when plugged into a high end PC, the numbers work out at around 100 GFLOPS per $1000.

This is not a detailed analysis, nevertheless from just these few examples it is clear that Moravec’s 1997 prediction of around 50 GFLOPS per $1000 in 2007 is indeed right on target.

So, what if his predictions are correct for another ten years? In 2017 we should see $1000 buying about 10,000 GFLOPS, or 10 TFLOPS. With Intel already showing off wafers with 80 core CPUs that they predict should be capable for 1 TFLOP, this doesn’t sound unrealistic.  If 10 PFLOPS is what is required for a functional simulation of the human brain, as a number of people have estimated, then in 10 years this should be achievable with just a few million dollars of hardware.

Ten years is not exactly a long time.