In memory or near memory compute?

Cisco
vif

Go to company page Cisco

vif
Sep 7, 2020 2 Comments

I read many papers that talks about bottle neck in Van-Neuman architecture. There is a lot of talk about in or near memory computing. Any chip makers looking at it? I know GSI technology released their Gemini-I chip last quarter. Benchmarks and initial studies validates the chip.

https://www.gsitechnology.com/APU

Anyone else doing the similar work? #semiconductor #in-memoryCompute #AIchips

2

comments

Want to comment? LOG IN or SIGN UP
  1. TOP
  2. RECENT
  3. OLD
 TOP 2 Comments
  • Western Digital / Eng
    Troll Food

    Go to company page Western Digital Eng

     Troll Food
    The problem is, it is hard to make large memory with fast processing. One tries to optimize for high capacitance while the other low capacitance. Doing both means both don't do well.

    The best approach would be decentralized tightly coupled memory. Keeping processing unit simple and memory small (KB) and near by (SRAM), then massive replication. This is not new and is how the TCM on arm works.

    The problem is you cannot make memory too big or decoding time and transfer time slow down. this is the reason you don't see Intel or AMD with GB of memory do this on PC.
    Sep 7, 2020 1