0:00:05 | i want my name is matt dispenser inaccurate a today to talk to about replacing |
---|
0:00:10 | transistors with mechanical switches |
---|
0:00:13 | upon hearing this |
---|
0:00:14 | you might rightly think why on earth would you do that |
---|
0:00:18 | i'm going to reply by taking a sort of roundabout route through history and pointing |
---|
0:00:22 | out that there's some poetry to this |
---|
0:00:24 | the first computers word fact mechanical the picture of data just difference engine |
---|
0:00:29 | which one powered by hand crank hoods all fit order polynomials |
---|
0:00:35 | if you best for the hundred years you get any which was the first fully |
---|
0:00:39 | digital electronic computer |
---|
0:00:41 | it way twenty tons consumed hundred fifty kilowatt power and perform the blazing five syllables |
---|
0:00:47 | per second of floating point operations |
---|
0:00:50 | now the difference between these two computers points out a tension in computer design that's |
---|
0:00:55 | been around since the eighteen forties which is between |
---|
0:00:58 | high powered high performance and lower power and lower performance |
---|
0:01:02 | fortunately we broke some of the design tradeoffs in a pretty significant way since the |
---|
0:01:06 | seventies |
---|
0:01:07 | but this tension actually resurfaced in a very significant way around two thousand |
---|
0:01:12 | this plot is a prediction of how computer power would increase from two thousand two |
---|
0:01:17 | thousand and ten and you lotus people expected power would increase a lot |
---|
0:01:23 | this obviously didn't happen as indicated by some of the annotations on the slide |
---|
0:01:29 | we don't have nuclear reactors in our laptops |
---|
0:01:31 | the question is how this happened |
---|
0:01:35 | then is actually prepared is very well for this the idea is that transistors have |
---|
0:01:40 | a property called their threshold voltage |
---|
0:01:42 | and if you just the threshold voltage properly you can trade off between two kinds |
---|
0:01:46 | of energy their dissipated in |
---|
0:01:48 | a very necessary energy called dynamic energy which has to do with running computer and |
---|
0:01:52 | weighted energy called leakage energy |
---|
0:01:55 | and |
---|
0:01:56 | by setting the threshold voltage properly you can actually find the minimum between them and |
---|
0:02:00 | make them operate perfectly |
---|
0:02:02 | this is actually what happened between two thousand two thousand ten in this why we |
---|
0:02:05 | have many course not computers right now uptalk three that based on the cartoon on |
---|
0:02:09 | the right the slide |
---|
0:02:10 | the idea is that if you were operating at the one x point in that |
---|
0:02:13 | cartoon you're consuming lots of dynamic energy above the optimum however you can slow yourself |
---|
0:02:18 | down in order to save energy |
---|
0:02:20 | and then stick to computers next to each other in order to recover your performance |
---|
0:02:23 | and let the software engineers figure out what to do with two computers |
---|
0:02:30 | you can do that again in this part you going from to x the four |
---|
0:02:32 | x parallelism in order to save energy but once you're cores at that point |
---|
0:02:38 | running it's lower won't save you any energy |
---|
0:02:41 | and so as a result |
---|
0:02:42 | you could this ceiling on parallelism which limits our ability to improve computing performance going |
---|
0:02:47 | forward |
---|
0:02:50 | now at this point my group likes it |
---|
0:02:52 | turns sharply in the left field and says |
---|
0:02:54 | the problem here is the transistor |
---|
0:02:56 | if we can replace that with something that doesn't have this wasted leakage energy |
---|
0:03:00 | then we can continue improving computing performance by scaling or voltage forever until we get |
---|
0:03:05 | some other physical them |
---|
0:03:07 | and what significant about this idea is that we've succeeded in building it |
---|
0:03:11 | this is a cartoon of the device that we build the ideas you have a |
---|
0:03:14 | piece of metal thing up near suspended by spring |
---|
0:03:17 | when you put a voltage on that piece of metal at home towards the surface |
---|
0:03:21 | and decorations on the bottom of connect different points on your chip |
---|
0:03:24 | the other significant things |
---|
0:03:26 | after we built that we measured it and found out that it has immeasurably low |
---|
0:03:29 | leakage as near as we can tell it has not |
---|
0:03:32 | so this means that we can replace the car drawing from the previous slide with |
---|
0:03:35 | the drawing in the bottom corner of this one where there's no leakage energy and |
---|
0:03:38 | we can just keep scaling forever |
---|
0:03:41 | now does this mean it's a good idea not necessarily |
---|
0:03:44 | we can call devices are big |
---|
0:03:46 | and their slow compared to electrons |
---|
0:03:48 | so there's a chance that we get a very energy efficient terribly performing computers if |
---|
0:03:53 | we tried use these |
---|
0:03:54 | however we've done a lot of very interesting work with circuit design in order to |
---|
0:03:58 | mitigate that |
---|
0:04:00 | problem in particular |
---|
0:04:02 | by changing our design style from stacking series of gates next to each other making |
---|
0:04:07 | very large distributed gates where all the input sit at the same time therefore all |
---|
0:04:11 | mechanical delay is incurred at the same time |
---|
0:04:14 | we can improve both are performance and advice count to save power energy and delay |
---|
0:04:21 | this is even more significant because we built some of these things we demonstrated that |
---|
0:04:24 | it is possible to get the functionality that we've been talking about and we don't |
---|
0:04:27 | a lot of extensive simulations showing that we can improve performance |
---|
0:04:32 | a result test case wasn't adder which we demonstrated twenty ten and simulations show that |
---|
0:04:35 | you can get ten x m for improvement over the absolute best transistor could do |
---|
0:04:42 | in terms of energy from the a ten x |
---|
0:04:44 | delay penalty |
---|
0:04:46 | we've also built a microprocessor and this other stuff coming the future including optimize memory |
---|
0:04:50 | structures minorities |
---|