4 ms·
The overblown frequency vs. cost efficiency trade-off
- bsder 11y agoThe problem with higher frequency is the portion of time you spend doing non-work things. The amount of time you spend on things like clock skew and setting up pipeline flip flops is generally fixed per cycle. So, let's say that you need 50ps to cover clock skew and 100ps to cover flip-flop overhead. At 1GHz, you have 850ps to do work over 1ns (one clock cycle in this instance). At 2Ghz, you have 500ps-100ps-50ps or 350ps per cycle which means that you have 700ps to do work in 2 cycles at 2GHz. At 1GHz, you have 850ps of work over the same time that you only have 700ps of work at 2GHz. That's quite significant. There are other advantages at 1GHz as well. You burn far less power trying to keep your clock aligned since you can have a little more slack and still come out ahead. You also don't have to flop as many intermediate results. etc. The problem is that we have only recently started caring about amount of work per (watt/area/cost/etc) instead of simply amount of work (total).
- _yosefk 11y agoWhat you pointed out is significant and I probably should have mentioned it in the article together with the other things; my point, however, is that these problems kick in at a higher frequency than those historically reached by many accelerators and defended as "optimal" by marketing. In your example, 1 GHz is used as a "nice low frequency" which is a sensible order of magnitude for a modern process; people often believe that you have to go as low as 400 or 100 MHz. (Again, the numbers vary with process, this is just to illustrate the extent of the difference between what I think is the reality vs what I think are the popular beliefs.)