TDA2EXEVM: Running Multiple Models on different EVE Cores

SagarK

Part Number: TDA2EXEVM

Hi Kumar,

I have 4 different models which needs to be run on TDA2X. And to get real time performance of 30fps, I have done few calculations based on what each EVE Core computation power can provide in terms of Number of MACs?

As per the datasheet, EVE core at frequency of 530 MHz can provide 8 GMACs.

8 Giga MACs @ 1 fps

Aprox 500 Million MACs @ 15 fps

Aprox 250 Million MACs @ 30 fps

From the above calculation, If i design each deep learning model with less than 250 million MACS, Will all the 4 models can run at 30 fps concurrently on 4 EVE cores ?

Regards,

Sagar Kale

over 6 years ago

0 kumar.desappan over 6 years ago

TI__Mastermind 22085 points

Hi Sagar,

Your calculations are right. These MAC numbers are considering 100% utilization of VCOP MACS. Instead, you need to consider 80% utilization. Please refer data sheet for details on MAC utilization of various convolution configurations. If your network has small feature maps (less than 16x16) then the utilization will be lesser than this for those layers.

If you can use caffe-jacinto to train with sparsity, your model can have higher MACs. Please refer below for more details on sparse training.

https://github.com/tidsp/caffe-jacinto

https://github.com/tidsp/caffe-jacinto-models

http://openaccess.thecvf.com/content_cvpr_2017_workshops/w4/papers/Mathew_Sparse_Quantized_Full_CVPR_2017_paper.pdf

Regards,

Kumar.D

0 Eric Lai over 6 years ago in reply to kumar.desappan

Expert 1140 points

Hello Kumar,

I face the similar problem about evaluating the MACS of models.

From the TIDL_model_import.exe , I can get the MACS of a model from the log.

i try the same model prototxt with different caffemodel ,one model is sparse and the other one is not.

but get the same MACS .

From the paper ,I know we can reduce the MACS by sparse property(due to huge number of zero).

Is there any way to evaluate the total MACS based on the sparsity?

Because I am new to the EVE core,i thought EVE core can make use of the sparsity to get the better performance.

if there is any mistake , please tell me.

Thank you so much.

Sincerely,
Eric Lai

0 kumar.desappan over 6 years ago in reply to Eric Lai

TI__Mastermind 22085 points

Eric, Your had simialr question in below thred also.
e2e.ti.com/.../2480884

0 Eric Lai over 6 years ago in reply to kumar.desappan

Expert 1140 points

Yes,thank you .

Sincerely,
Eric Lai

Processors

Processors forum

TDA2EXEVM: Running Multiple Models on different EVE Cores