Raspberry Pi 3 + Google TPU?

PINTO0309 / TPU-MobilenetSSD

Edge TPU Accelerator / Multi-TPU + MobileNet-SSD v2 + Python + Async + LattePandaAlpha/RaspberryPi3/LaptopPC

https://qiita.com/PINTO

MIT License

82 stars 22 forks source link

Raspberry Pi 3 + Google TPU? #1

Open reanlee opened 5 years ago

reanlee commented 5 years ago

Hi,

Thanks for your sharing!

The Core m3 seems more powerful than other SoCs, Do you have plan to try TPU with RP3?

Thanks.

BRs, Lee

PINTO0309 commented 5 years ago

@reanlee I already tried the Rpi3 + TPU combination. There are only Japanese articles, please refer to the following. https://qiita.com/PINTO/items/dd6ba67643bdd3a0e595

reanlee commented 5 years ago

Hi,

Thank you for your article.

Originally I expected Edge TPU to be better than NCS2's performance in RPi3, the bottleneck seems in the USB bandwidth?

If so, there is a TPU Dev. board, maybe it wil have better bandwidth balance and performance.

Thanks again.

BRs, Lee

PINTO0309 commented 5 years ago

Originally I expected Edge TPU to be better than NCS2's performance in RPi3, the bottleneck seems in the USB bandwidth?

I can not say yet so in an assertive way. I have not been able to determine where the bottleneck is.

Image size 320 x 240 x 8 x 3 = 1843200 bit 1843200 / 8 = 230400 byte 230400 / 1024 /1024 = 0.2197265625 Mbyte
USB2.0 Effective speed 300 Mbit/sec 37.5 Mbyte/sec
Image transfer time 0.2197265625 Mbyte / 37.5 Mbyte = 0.005859375 sec 0.005859375 sec = 5.859375 ms

reanlee commented 5 years ago

@PINTO0309 Hi,

Thanks for the reply!

I have no idea at the moment, I can't buy the edge TPU in my region for now. I think I can try the dev. board to compare the performance with USB TPU after I get one.

Thanks & BRs, Lee