FlexLogic Minimizes DRAM For AI

Discussion in 'Science and Technology' started by wooleeheron, Nov 2, 2018.

  1. wooleeheron

    wooleeheron Brain Damaged Lifetime Supporter HipForums Supporter

    Flex Logix Says It Has Solved Deep Learning’s DRAM Problem

    This is one of those minor technical leaps that often goes under the radar. The entire computer industry is attempting to redesign the modern PC for distributed computing, with every chip now coming out with AI circuitry embedded in memory. The emphasis on distributed computing for consumer products can actually be traced back to Intel's rude surprise when they discovered that, past a certain point, they could not make a chip go faster by merely adding more parts, due to a physical limitation in silicon. Only in the last few years has the industry and academics managed to work out most of the mathematics for designing multicore processors that are really efficient. For example, someone recently figured out how to cut down the amount of time it takes for two or more processors to communicate by at least a third.

    What these people have done is significantly cut down the amount of memory required to operate fgpa circuitry, making it possible to seriously cut prices on embedding such circuitry in the system ram. By embedding AI circuitry in system ram it can serve as not only ram, but a third "Goldilocks" processor that only crunches numbers of a specific size for repeated tasks. The idea is to have the ram do more of the stupid work so the CPU and GPU can focus on more important things. Cutting the amount of ram required means the cost of doing the stupid calculations goes down and you can send data back and forth to memory less often.

    Assuming the research is applicable to something like Nvidia's tensor cores, it could save on the amount of vram required on graphics cards, which could bring down the cost by as much as a third. Memory like HBM2 is outrageously expensive because the demand far exceeds their ability to manufacture it and any means of potentially cutting the costs of high speed memory are being perused. While the prices on high performance processors and graphics cards is suddenly within the reach of consumers, all the cellphones in particular have caused the price of memory to skyrocket despite everyone's efforts.
    Last edited: Nov 2, 2018
  2. Friar Turk

    Friar Turk Dankin' and Tankin'

    "every chip now coming out with AI circuitry embedded in memory"........................................CAN THIS BE VALID!

Share This Page