TDA4VEN-Q1: Please help confirm tda4ven performance confirmation

Daniel Cheng1

Mastermind 6415 points

Part Number: TDA4VEN-Q1

Hi Expert,

1.What is the maximum TOPS that TDA4VEN inference INT8 can run? Is 4 TOPS possible?

2.If tda4ven is not efficient enough, is there any recommended soc for 4tops?

3.How many modes can TDA4VEN run at the same time?

4.When the model is not shared - how long does it take to unload and reload the model?

5.How many channels can the SOC's built-in ISP support?

6.How many exposure combinations can ISP HDR support?

Thanks

Daniel

11 months ago

0 Daniel Cheng1 11 months ago

Mastermind 6415 points

Hi Expert,

Most of the problems have been solved by local EP FAE.

We still need your help to confirm the following questions:

4.When the model is not shared - how long does it take to unload and reload the model?

6.How many exposure combinations can ISP HDR support?

Thanks

Daniel

0 Brijesh Jadav 11 months ago in reply to Daniel Cheng1

TI__Guru**** 450265 points

Hi Daniel,

I can help with #6 question. ISP can support merging at max 3 exposures. I will recommend referring to ISP documentation for these questions.

Regards,

Brijesh

0 Pratik Kedar 11 months ago in reply to Brijesh Jadav

TI__Mastermind 24041 points

Daniel Cheng1 said:
4.When the model is not shared - how long does it take to unload and reload the model?

Could you please be more specific ? Do you mean upload on forum ?

0 Daniel Cheng1 11 months ago in reply to Brijesh Jadav

Mastermind 6415 points

Hi Brijesh

Are you referring to the following files? Are there any other suggested files?

TDA4VMid VPAC ISP Tuning Overview (Rev. A)

Thanks

Daniel

0 Brijesh Jadav 11 months ago in reply to Daniel Cheng1

TI__Guru**** 450265 points

Hi Daniel,

No, this is an application report. Please check in the TRM of the device.

Regards,

Brijesh

0 Daniel Cheng1 11 months ago in reply to Pratik Kedar

Mastermind 6415 points

Hi Pratik

Assuming we have already loaded the weights of multiple AI models into memory space, how are these weights loaded by the AI engine when using it?

Does switching the weights of different models during the loading and unloading process require additional data transfer time?

Thanks

Daniel

0 Pratik Kedar 11 months ago in reply to Daniel Cheng1

TI__Mastermind 24041 points

Hi Daniel,

The loading and storing (load - upload in your terminology) is something which is internal to ti, as this includes software IP based decision to utilize on chip memories(SRAMs).

In public forum i can share few things they are,

The decision making of loading and storing intermediate tensors during inference of model is done during network compilation process, NC tool does this optimally to have less memory latencies.

Hope this will add clarity.

Thank You

Processors

Processors forum

TDA4VEN-Q1: Please help confirm tda4ven performance confirmation