Can Myriad X be stacked into 8 16 or 32 TOPs inference processor?

[idata]

Let me start with all I know about is that Myriad X is its a chip that can be used with TensorFlow.

So that got me thinking, can this chip somehow be used with trained models and their inference results be sped up be combining Myriad X chips?One Myriad X chip is 4 TOPS is there a way to serialize the information across the chips so you can increase performance based on the number of chips?Just a question.

[idata]

"I must have not had any coffee when I wrote that first post so let me try again."

Let me start with all I know about Myriad X is that its a chip that can be used with TensorFlow to speed up getting inference results from trained data set.

One Myriad X chip gets 4 TOPS so is there a way to combine multiple Myriad X chips to

get higher performance? 2 Myriad X chips gets you 8 TOPS 4 gets you 16 etc.

Just a question.

[idata]

Hi @chicagobob123, I can't speak to Myriad X specifically, but as you are already aware, NCS (with Myriad 2) can be chained/combined to dynamically scale inference performance. Here's some implementation details about chaining multiple NCS - https://ncsforum.movidius.com/discussion/comment/395/#Comment_395. Conceptually, this would be no different for Myriad X.

[idata]

If anyone can buy these NCS sticks to have more than one let me know.

Impossible to get.

[idata]

@chicagobob123 As you may already know the NCS are now available. As I read in the Intel Announcement previous batch of NCS had run out very quickly because of high demand. So, in case you feel that you will require multiple chips, it may be better to stock up now when these are still available. I myself bought a couple, in case one get damaged for some reason. Also, as you may know MyriadX will have 16 cores, instead of 12 cores on Myriad v2. Intel claims MyriadX to be 30% faster than Myriad v2.

[idata]

Yes my stick is finally plugged into a laptop trying to understand all the new code and readme's.

This will take a while :)

[idata]

can be chained/combined to dynamically scale inference performance<<

This seems a bit misleading. In reading the comment you made it seem like they would accelerate a single

network. According to the comments in the thread it can only be used as 4 independent devices.

I see nowhere a mention of them working as one unit.