Why does Myriad X Support only FP16 Models?

[sovit]

I was going through this link and noticed that the Vision Processing Unit (MYRIAD) supports only FP16 models. I tried searching on the internet for any particular reason/explanation but could not find any. It will be very helpful if anyone can clear this up or point to any resources explaining the reason for the same.

[Wan_Intel · Accepted solution]

Hi Sovit Rath,

Thank you for reaching out to us.

Advantages of FP16 are improved speed (TFLOPS) and performance, reduced memory usage of a neural network, and faster data transfer. Size of FP16 is half of FP32, thus it provides 50% saving in terms of space requirements of memory access, cache, bandwidth, and storage and disk IO.

Myriad-X was designed from the ground up to be a vision and AI accelerator.

It is a small memory, low power device, and its native data format is FP16 like the Intel integrated GPU. Myriad-X uses 16 SHAVE vector processors, an accelerator microarchitecture designed by Movidius for vision processors. The details of Streaming Hybrid Architecture Vector Engine (SHAVE) are available at the following page:

https://en.wikichip.org/wiki/movidius/microarchitectures/shave_v2.0

You can have a look at this publication:

https://www.researchgate.net/publication/271722951_The_Movidius_Myriad_Architecture%27s_Potential_for_Scientific_Computing

Regards,

Wan

[Wan_Intel]

Hi Sovit Rath,

This thread will no longer be monitored since this issue has been resolved. If you need any additional information from Intel, please submit a new question.

Regards,

Wan