Probabilistic Logic Allows Computer Chip to Run Faster - 3 views
-
Dario Izzo on 07 Dec 09Francesco pointed out this research one year ago, we dropped it as noone was really considering it ... but in space a low CPU power consumption is crucial!! Maybe we should look back into this?
- ...6 more comments...
-
Ma Ru on 07 Dec 09Q1: For the time being, for what purposes computers are mainly used on-board?
-
Dario Izzo on 07 Dec 09for navigation, control, data handling and so on .... why?
-
Ma Ru on 07 Dec 09Well, because the point is to identify an application in which such computers would do the job... That could be either an existing application which can be done sufficiently well by such computers or a completely new application which is not already there for instance because of some power consumption constraints... Q2 would be then: for which of these purposes strict determinism of the results is not crucial? As the answer to this may not be obvious, a potential study could address this very issue. For instance one can consider on-board navigation systems with limited accuracy... I may be talking bullshit now, but perhaps in some applications it doesn't matter whether a satellite flies on the exact route but +/-10km to the left/right? ...and so on for the other systems. Another thing is understanding what exactly this probabilistic computing is, and what can be achieved using it (like the result is probabilistic but falls within a defined range of precision), etc. Did they build a complete chip or at least a sub-circiut, or still only logic gates...
-
Dario Izzo on 07 Dec 09Satellites use old CPUs also because with the trend of going for higher power modern CPUs are not very convenient from a system design point of view (TBC)... as a consequence the constraints put on on-board algorithms can be demanding. I agree with you that double precision might just not be necessary for a number of applications (navigation also), but I guess we are not talking about 10km as an absolute value, rather to a relative error that can be tolerated at level of (say) 10^-6. All in all you are right a first study should assess what application this would be useful at all.. and at what precision / power levels
-
Joris _ on 07 Dec 09The interest of this can be a high fault tolerance for some math operations, ... which would have for effect to simplify the job of coders! I don't think this is a good idea regarding power consumption for CPU (strictly speaking). The reason we use old chip is just a matter of qualification for space, not power. For instance a LEON Sparc (e.g. use on some platform for ESA) consumes something like 5mW/MHz so it is definitely not were an engineer will look for some power saving considering a usual 10-15kW spacecraft
-
Dario Izzo on 07 Dec 09What about speed then? Seven time faster could allow some real time navigation at higher speed (e.g. velocity of a terminal guidance for an asteroid impactor is limited to 10 km/s ... would a higher velocity be possible with faster processors?) Another issue is the radiation tolerance of the technology ... if the PCMOS are more tolerant to radiation they could get more easily space qualified.....
-
Joris _ on 08 Dec 09I don't remember what is the speed factor, but I guess this might do it! Although, I remember when using an IMU that you cannot have the data above a given rate (e.g. 20Hz even though the ADC samples the sensor at a little faster rate), so somehow it is not just the CPU that must be re-thought. When I say qualification I also imply the "hardened" phase.
-
Francesco Biscani on 08 Dec 09I don't know if the (promised) one-order-of-magnitude improvements in power efficiency and performance are enough to justify looking into this. For once, it is not clear to me what embracing this technology would mean from an engineering point of view: does this technology need an entirely new software/hardware stack? If that were the case, in my opinion any potential benefit would be nullified. Also, is it realistic to build an entire self-sufficient chip on this technology? While the precision of floating point computations may be degraded and still be useful, how does all this play with integer arithmetic? Keep in mind that, e.g., in the Linux kernel code floating-point calculations are not even allowed/available... It is probably possible to integrate an "accelerated" low-accuracy floating-point unit together with a traditional CPU, but then again you have more implementation overhead creeping in. Finally, recent processors by Intel (e.g., the Atom) and especially ARM boast really low power-consumption levels, at the same time offering performance-boosting features such as multi-core and vectorization capabilities. Don't such efforts have more potential, if anything because of economical/industrial inertia?