• If you had a million core CPU, then a 1thz clock speed could make a mil

    From Treon Verdery@21:1/5 to All on Wed Sep 7 09:39:12 2022
    if the 1 thz clock was 90% (or 7%) data reliable then each 10,000 cores would be 99.9% likely to be running a as-written version of the program, at all million cores it would be 99.999% likely to be running the as written version of the program, some
    programs I perceive are flexible like neural networks and deep learning AI where 99.9% might sometimes be adequate, especially when more learning data could compensate for the .01% variation at neural weights, so a 1thz computer clock speed at a neural
    network could be possible, that is 250 times faster than a 2019 personal computer or server, that makes neural network computing orders of magnitude more affordable than other kinds of computing

    Each core could make a minimized byte hash of the program it has and runs, then when first sending program output the hash is sent out once, along with the program output, that is compared to the hash of the actual program as written, when they are the
    same then that is a core verified as running the program as written, that can then while the ones doing it right, are doing it right, then iteratively move the rest of the cores to running the program as written on all 100 or one million cores, then if
    the core running the program has 64 registers of 128 bytes each then the accuracy of the contents and computing actions at those registers are multiplied at the "9% of copies of a million byte program loaded has integrity", so each 111 one thz clock
    cycles(9% majority), or each 333 one thz clock cycles the 3% majority is utilized, comparing register contents between cores, to say " the majority say thats what the register contents or data actually are, so one between cores register comparison (
    optionally register hash comparison) every 333 cycles, it could be all the other cores register variants hashes identicalize at far less than 1%, so a 1% majority, with about 1k clock cycles between hash comparisons is possible, what about power
    consumption, IC area and affordability, and which applications benefit from being run at 1 thz per core, driverless cars work at 4ghz but might work better at 1thz, some medical imaging like brain and body scans like positron emission tomography noting
    neuron type and tissue structure less than 1 mm area (I may have read decimal millimeters) could process at 99.9% accuracy and do another scan if more accuracy was preferred, this omits the data bandwidth of sending the image to the cloud to be processed
    at a couple hundred computers, anyplace where bandwidth to the cloud is lengthy, like huge multipetabyte databases, where at some versions of this a 99.9% accurate output is sufficient( processing all of Facebook, a social networking site among others to
    bring voluntary content or products to finding children that could easily be made happy, parents that could improve their parenting style or, at the 1 per 10 million error rate children that would benefit from being rescued), enterprise resource planning
    (ERP) data repositories, some large physics experiments,
    GPUs exist now, comparing a hundred or million cores at 1thz to GPUs, other than the one thz processing velocity, the two out of three approach at highly over clocked GPUs has very similar benefits

    40k people, 20k things each, motion processed every 10 milliseconds (see a keyboard keypress, grab each

    Positron sensors, do isotopically pure semiconductors, or even CCDs respond more accurately giving higher resolution

    Memory CPU

    or even to make at each 1 thz clock cycle, and all the other cores out of 100 or one million are availablized to run other programs, the 100 or one million core CPU then loads programs to those other cores until perhaps all (resend the program data
    multiple times until the hash matches at the latter 9 cores), the 100 or one million core CPU is then fully loaded, along with neural networks some applications are running multiple different programs from different people, aggregated on the internet
    cloud, possibly servers, at personal computers and even possibly phones although it is possible to run each program, application, or component of an operating system on a separate core right now for waitless utilization, a separated component software
    form, and

    A better way might, depending on program execution time, might be 9% likeliness of the neural network program being as written at 100 cores, then if there were anything other than 7-9 identical outputs and hashes (the other 89 are likely to differ from
    each other stochastically and be less homogenous than the 9%)then the program would reload, or at a million CPU chip, 10,000 groups of 100 each at 9% likeliness of being the program as written the output of the program as written could be very strongly
    numerically present,

    100ghz and possibly 1 thz test instruments exist, I think using analog ICs so loading the 100 or 1 million cores that fast is likely possible

    --- SoupGate-Win32 v1.05
    * Origin: fsxNet Usenet Gateway (21:1/5)