2015-11-05 23:11 GMT+01:00 Ilya Zakharevich <[email protected]> wrote > > • 128-bit architectures may be at hand (sooner or later).
This is specialation for something that is still not envisioned: a global worldwide working space where users and applications would interoperate transparently in a giant virtualized environment. However, this virtualized environment will be supported by 64-bit OSes that will never need native support of more the 64-bit pointers. Those 128-bit entities needed for adressing will not be used to work on units of data but to address some small selection of remote entities. Softwares that would requiring parsing coompletely chunks of memory data larger than 64-bit would be extremely inefficient, instead this data will be internally structured/paged, and only virtually mapped to some 128 bit global reference (such as GUID/UUIDs) only to select smaller chunks within the structure (and in most cases those chunks will remain in a 32-bit space (even in today's 64-bit OSes, the largest pages are 20-bit wide, but typically 10-bit wide (512-byte sectors) to 12-bit wide (standard VMM and I/O page sizes, networking MTUs), or about 16-bit wide (such as transmission window for TCP). This will not eveolve significantly before a major evolution in the worldwide Internet backbones requiring more than about 1Gigabit/s (a speed not even needed for 4K HD video, but needed only in massive computing grids, still built with a complex mesh of much slower data links). With 64-bit we already reach the physical limits of networking links, and higher speeds using large buses are only for extremely local links whose lengths are largely below a few millimters within chips themselves. 128 bit however is possible not for the working spaces (or document sizes) it will be very unlikely that ANSI C/C++ "size_t" type will be more than 64-bit (ecept for a few experimentations which will fail to be more efficient). What is more realist is that internal buses and caches will be 128 bits or even larger (this is already true for GPU memory), only to support more parallelism or massive parallelism (and typically by using vectored instructions working on sets of smaller values). And some data need 128-bit values for their numerical ranges (ALUs in CPU/GPU/APU are already 128-bit, as well as common floating point types) where extra precision is necessary. I doubt we'll ever see any true native 128-bit architecture in any time of our remaining life. We are still very far from the limit of the 64-bit architecture and it won't happend before the next century (if the current sequential binary model for computing is still used at that time, may be computing will use predictive technologies returning only heuristic results with a very high probability of giving a good solution to the problems we'll need to solve extremely rapidly, and those solutions will then be validated using today's binary logic with 64-bit computing). Even in the case where a global 128-bit networking space would appear, users will never be exposed to all that, msot of this content will be unacessible to them (restricted by secuiry concerns or privacy) and simply unmanageable by them : no one on earth is able to have any idea of what 2^64 bits of global data represents, no one will ever need it in their whole life. That amount of data will only be partly implemented by large organisations trying to build a giant cloud and whiching to interoperate by coordinating their addressing spaces (for that we have now IPv6). So your "sooner or later" is very optimistic. IMHO we'll stay with 64-bit architectures for very long, up to the time where our seuqnetial computing model will be deprecated and the concept of native integer sizes will be obsoleted and replaced by other kinds of computing "units" (notably parallel vectors, distributed computing, and heuristic computing, or may be optical computing based on Fourier transforms on analog signals or quantum computing, where our simple notion of "integers" or even "bits" will not even be placeable into individual physically placed units; their persistence will not even be localized, and there will be redundant/fault-tolerant placements). In fact our computing limits wil no longer be in terms of storage space, but in terms of access time, distance and predictability of results. The next technologies for faster computing will be certainly predictive/probabilistic rather than affirmative (with today's Turing/Von Neumann machines). "Algorithms" for working with it will be completely different. Fuzzy logic will be everywhere and we'll even need less the binary logic except for small problems. We'll have to live with the possibility of errors but anyway we already have to live with them evne with our binary logic (due to human bugs, haardware faults, accidents, and so on...) In most problems we don't even need to have 100% proven solutions (e.g. viewing a high-quality video, we already accept the possibility of some "quirks" occuring, and we already accept some minor alterationj of the exact pixel colors in which we can't even note any visible difference from the original ; another example is in what we call a "scientific proof" which is in fact only a solution with the highest probability of being correct in almost all known contexts, because we can never reproduce exactly the same exprimental environment: even a basic binary NAND gate cannot be warrantied at 100% of always returning a "0" state after a defined delay when its inputs are all "1"). We can certainly produce results with the same (or better) probability of giving the expected result using fuzzy logic (or quantum logic) rather then existing binary logic, and certainly with smaller computing delays (and better throughputs and better fault torlerance, incliuding after hardware faults or damages, and even with better security).
