Einstein and L1 cache

You may be interested that there is a link between on-chip cache & relativity.

The founding postulate of relativity (special or general) is that no signal propagates faster than light.

If you have a 3GHz processor, one cycle takes 333picoseconds, and a light signal can only travel 10cm. That is the there-and-back distance, so for a same-cycle fetch instruction the info must be within 5cm of the chip. If your motherboard cache is more than 5cm (two inches) from the cpu then an on-chip cache is essential to run at this speed.

For each 67 picoseconds it takes the components to react (total reaction time of the cpu and of the memory) the max two-way path is reduced by another 1cm, until the point where most of the reaction time is taken up by the components, and the memory must be actually on the chip.

Realistically it is harder to reduce reaction time than to make the path-lengths shorter, so in fact the overall reaction time of the components is the dominant limit, but path-length is the second most limiting factor on the clock speed of a modern computer, and that limit arises from relativity.

If the on-chip cache is not big enough to hold almost all the code in the main loop of a process, then most of the 'processing' time is not processing at all, but simply time spent waiting for signals to travel back and forth within your box. Hence the dramatic increase in speed when the on-chip cache gets to some critical value, and the only-slight increase in speed beyond that point, are both predicted by relativity.

Again, real world constraints make real-world machines slower than the theoretical limits, but I find it mind boggling that our home computers are now fast enough that relativity sets meaningful limits to their performance, and within a factor of five of the practical limits.

You could say the thought brings me Einstein at Home ;-)