DRA74P: ADAS and Cluster Details for DRAX digital SOC platforms

Ekrem Sahin

Part Number: DRA74P
Other Parts Discussed in Thread: DRA742, DRA756

Hello,

We want to prepare an In-Vehicle Infotaintment System(IVI) that supports media,navigation, bluetooth,Voice Control, AM/FM Radio, Telephony Features, Hand Gesture Recognition and Can bus functionalities that is based on Android and We want to use Advanced Driver Assistance Systems (ADAS) functionalities that supports Deep Learning.We also willing to add cluster functionality later.We found that DRAx digital soc platforms fits for our requirements but we have some questions as the following:

Could you please clarify for the followings:

-It seems to us not all of the DRAx digital soc platforms fits for our requirements. For instance, for ADAS functionality, We need to buy one of DRA75x or DRA77x. Rest of the platforms, DRA74x and DRA76x not supporting ADAS functionality. Does DRA74x and DRA76x supporting ADAS functionality or not?

-Also, From the link "www.ti.com/.../digital-instrument-cluster.html" , We did not see that DRA75x or DRA77x is supporting reconfigurable digital instrument cluster. Does DRA75x or DRA77x supporting reconfigurable digital instrument cluster or not ?

For ADAS:

-Our ADAS team has experienced with running TIDL/Vision SDK on TDA2xx SoCs in custom boards (Alphaboard), however they are saying that the supported models are limited (TFLite 1.0, Caffe<=1.7). caffe-jacinto-models repository only references JacintoNetV2 and MobileNet, the latter is not complete regarding the training environment. They also say that they were only able to train JacintoNetV2 with custom dataset, and not any other model due to lack of utility provided in the repository for other models. Our end goal is to say, take YolooV3 or any other high performance model, and train it with a completely custom dataset, then run it on the board seamlesly. However, for such inquiries TIDL has constant limitations (e.g.: e2e.ti.com/.../687877) Another issue that they report is the lack of layer support (e.g. Norm) in TIDL import tool, preventing them to import well-known models to the embedded platform.Our team has been able to run TIDL/Vision SDK with 3.03 and on SYS-BIOS. Do you think running it on a Linux and Jacinto ES 1.1 will give the flexibility to achieve the aforementioned goals? Is there a complete guide in importing generic models, from code to training to importing, and running on the board?

-We have following platforms and trying to understand if we can manage our requirements with the platforms we have.So Could you please check our platform details and tell us if we can manage our requirements with our platform or If it is not possible , Could you please suggest us a platform so that we can buy ?

CPU : DRA752-GP ES1.1
Model: TI DRA742
Board: DRA74x EVM REV G.0
DRAM: 1.5 GiB

Best Regards

Ekrem Sahin
Senior Software Engineer

over 5 years ago

0 Karthik Ramanan over 5 years ago

TI__Guru** 113810 points

Hi Ekrem,

I will ask my colleague to get in touch with you shortly.

Regards

Karthik

0 Ekrem Sahin over 5 years ago in reply to Karthik Ramanan

Prodigy 130 points

Hello Karthik,

Do you have any update for my question ?

Regards

0 Steve Reis over 5 years ago in reply to Ekrem Sahin

TI__Genius 17625 points

The Jacinto 6 EX family of devices is more suitable for combined IVI + ADAS function as this variant includes both dual C66x DSP and dual EVE (Embedded Vision Engine).

Jacinto 6 EX (DRA756 / 755 / 754)
Jacinto 6 EX Plus (DRA777 / 776 / 775 / 774)

These devices can also support digital cluster, even though only the standard Jacinto 6 devices are listed on the link you noted.

The level of ADAS support these devices can provide in a digital cockpit (IVI+cluster) configuration depends on the usecase load of the cockpit usecases, especially the display sizes

Below find a link to a CES 2018 video showing a Digital Cockpit and Surround View demo on Jacinto 6 Plus. https://www.youtube.com/watch?v=e1SdVrClqY0&index=11&list=PLISmVLHAZbTSaEdXXsB3rW3dtuFta80od

I have asked a colleague to respond on the Deep Learning portion of your query.

0 Ekrem Sahin over 5 years ago in reply to Steve Reis

Prodigy 130 points

Hello Steve,

Thanks for the information.We are waiting Deep Learning portion of our query.

Best Regards

0 kumar.desappan over 5 years ago in reply to Ekrem Sahin

TI__Mastermind 22145 points

Hi,

J6 family of devices support model formats like Jacintinet11v2, Mobilenet v1/v2, resnet18, inception net v1, squeeze 1.1 etc.

Please refers to the datasheet for more details.

We do not support the layers required for yolo v3 model architecture.

Regards,

Kumar.D

0 Ekrem Sahin over 5 years ago in reply to kumar.desappan

Prodigy 130 points

Hi Kumar,

Thanks for the feedback.

Could you give us links as to where these officially supported models are located and their documentation?

caffee-jacinto-models repository only supports jacintoNet and mobileNet, however l1reg and sparse training steps for mobileNet is missing, and therefore the output is very fuzzy.

Therefore, using TIDL_OD usecase example, only jacintoNet works. We want to be able to run other models with the TIDL_OD example.

Thank you

Best Regards

0 kumar.desappan over 4 years ago in reply to Ekrem Sahin

TI__Mastermind 22145 points

HI TIDL documentation is available in TIDL SW folder in SDK release package.

Validated models and their performance details are available in "TIDeepLearningLibrary_DataSheet.pdf"

0 Ekrem Sahin over 4 years ago in reply to kumar.desappan

Prodigy 130 points

Hi Kumar,

Thanks for the information.

Best Regards

Processors

Processors forum

DRA74P: ADAS and Cluster Details for DRAX digital SOC platforms