Hacker News new | comments | ask | show | jobs | submit login

How would someone compare the Edge TPU Accelerator with a Movidius Neural Compute Stick? https://software.intel.com/en-us/movidius-ncs





Movidius is not a TPU. It's more like a GPU, but with SIMD, DSP and even VLIW capabilities and with a _very_ wide memory bus (and massive throughput). It's rather impressive actually, but probably serious overkill for what really needs to be done during inference: https://en.wikichip.org/wiki/movidius/microarchitectures/sha.... Whereas a TPU is highly specialized for just, you guessed it, processing tensors, which basically means matrix and vector multiply. It's a systolic architecture, so it also (purportedly, since I don't have insider knowledge) stores the weights for the computation for the duration of the computation.

As far as I understand "TPU" is Google's brand name, so of course competing products are not TPUs. There is an overlap in what you can do with the devices, so a comparison of their strengths would be useful.

We hardly know anything about Edge TPU. Yes, TPU uses systolic array, but there is really no reason to believe Edge TPU will use it. Edge TPU is not TPU.

It has "TPU" right its name. If that's not a dead giveaway, I don't know what is. :-)



Applications are open for YC Summer 2019

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact

Search: