3 ms·
Best of times, worst of times. I asked a friend that if they could "fix" or "improve" any feature of a modern microprocessor what would it be, and they struggle
by ChuckMcM 10y ago
Best of times, worst of times. I asked a friend that if they could "fix" or "improve" any feature of a modern microprocessor what would it be, and they struggled to come up with anything. It it interesting though that a "nominal" desktop computer these days consists of a nominally sequential processing unit (albeit with multiple cores) and a large parallel processing unit in the GPU with lighter weight cores.
The "painful" part isn't the computation any more its getting multi-gigabyte data sets into and out of DRAM or moving them around in DRAM. You can finesse some of that by adding more and more cache on chip but you end up with the 'buffer bloat' problem where you're caches are fighting each other. If you pick the 16 bytes per computation figure that is sometime bandied about then your 4.2Ghz average computation bandwidth wants something like 67 GB/second of main memory bandwidth.
I suspect that is why IBM spent a lot of time on upping the memory bandwidth in Power8, it does no good to have CPU cycles stalled waiting for memory.
Looking at the existing Intel line up, they have specialized instructions for some pretty oblique programming tasks, they have tweaked their pipeline and register pool to minimize bubbles in the pipeline path pretty effectively, and they have I/O channels that can out perform much of the gear out there. So what do you improve go after next?
- kev009 10y agoI was hoping someone else would comment about the opera playing out as it were. IBM has tooled POWER8 and POWER9 to where the industry is heading.. all flash, post-NAND, accelerators, it's geared for buzzwords like big data, machine learning but also traditional RDBMS and high throughput computing. Contrasting POWER to Xeon is pretty interesting, and I think IBM will have a good few years. Historically, that kind of thing reengages Intel who come back with vengeance.
- TheOtherHobbes 10y agoThat still assumes you have the usual Von Neumann architecture. But that's not the only option, and for some tasks it's an active handicap. Entire application classes are still CPU/RAM bound. What we have no is one of those plateaus the industry gets stuck at every so often. It's very similar to the 8-bit days, in that it seems to be enough for most of the things user do already, if you don't mind grinding along. But it's very much not enough for many things users might be doing, but can't - including domestic robotics, non-trivial NLP, VR (although I have doubts about that, unless someone invents less clunky hardware), and pretty much any AI-enhanced application. Current versions of many of those run on the modern equivalent of a mainframe - i.e. one or more servers in a remote data center. There are huge gains in usefulness to be made with AI that is personal, domestic and/or office-aware, always-on, and local.
- userbinator 10y agoLooking at the existing Intel line up, they have specialized instructions for some pretty oblique programming tasks, they have tweaked their pipeline and register pool to minimize bubbles in the pipeline path pretty effectively, and they have I/O channels that can out perform much of the gear out there. So what do you improve go after next? How about moving the computation to the data? I.e. add ALUs and similar structures to DRAM, so there's nearly no need to move the data itself. Here's an interesting paper a few years ago that dealt with memcpy() and memset() effectively without using any memory bandwidth: https://users.ece.cmu.edu/~omutlu/pub/rowclone_micro13.pdf https://users.ece.cmu.edu/~omutlu/pub/rowclone_micro13.pdf The relevance here being that memcpy() and memset() are already specialised in x86 by the REP MOVS and REP STOS instructions, respectively.
- B1FF_PSUVM 10y ago> add ALUs and similar structures to DRAM CDC-6600 all the things! (The 'talk to smart peripherals' approach, if memory serves.)
- vardump 10y ago> I asked a friend that if they could "fix" or "improve" any feature of a modern microprocessor what would it be, and they struggled to come up with anything. Ever tried to compute histogram efficiently on a modern CPU? Even if it's simple 256 bin. As far as I know, it's impossible to vectorize, things will need to go slow scalar path. Or anything that needs to access memory at multiple (semi) random addresses inside one SIMD register (128/256/512 bits wide). It might help if Intel could also add scatter vector instructions. Well, assuming they can also improve gather/scatter operations on the memory controller...
- jburgess777 10y agoIntel added gather operations in AVX2 and scatter operations in AVX512. Here is a presentation from 2014 describing the features and the initial compiler support: https://gcc.gnu.org/wiki/cauldron2014?action=AttachFile&do=get&target=Cauldron14_AVX-512_Vector_ISA_Kirill_Yukhin_20140711.pdf https://gcc.gnu.org/wiki/cauldron2014?action=AttachFile&do=g... In October 2016 Intel started talking about adding more 512 bit instructions: http://sites.utexas.edu/jdm4372/2016/11/05/intel-discloses-vectorsimd-instructions-for-future-processors/ http://sites.utexas.edu/jdm4372/2016/11/05/intel-discloses-v...