The emulating mode setting of Jetson AGX Orin Developer Kit

Autonomous Machines Jetson & Embedded Systems Jetson AGX Orin
1

I’m using Jetson AGX Orin Developer Kit.

And I have two questions about Emulating mode.
When I used the following command for emulating Jetson AGX Orin 32GB.

$ sudo ./flash.sh jetson-agx-orin-devkit-as-jao-32gb mmcblk0p

Develop for All Six NVIDIA Jetson Orin Modules with the Power of One Developer Kit | NVIDIA Technical Blog

The GPU clock is not 940MHz but 816MHz as follows.

image
image1040×405 11.1 KB

image
image365×1731 40.2 KB

Question 1)
Does this mean Jetson AGX Orin 32GB emulating mode setting is as follows.

Jetson AGX Orin 32GHz Spec(Emulating Mode)
Real GPU Clocks: 816MHz
Cuda Core: 2048 Core
Tensor Core: 64 Core

Jetson AGX Orin 32GHz Spec:
Real GPU Clocks: 930MHz
Cuda Core: 1792 Core
Tensor Core: 56 Core

Question 2)
And total work loads are similar as follows?

Cuda Core Work loads(Real) : 930MHz * 1792 = 1,666,560
Cuda Core Work loads(Emulating) : 816MHz * 2048 = 1,671,168
Cuda Core Work loads(Real) : 930MHz * 56 = 52,080
Cuda Core Work loads(Emulating) : 816MHz * 64 = 52244

3

That seemes due to the NVPModel not set to MAXN.
See also Jetson Orin Nano Series, Jetson Orin NX Series and Jetson AGX Orin Series — Jetson Linux Developer Guide documentation (nvidia.com)

4

Dear @kayccc,

Thank you for your information.
I understood that the GPU clock is set to 940MHz by setting NVPModel to MAXN.

I want to ask two more questions?

Question3)
If so, is the number of cuda cores and tensor cores are decreased when jetson-agx-orin-devkit-as-jao-32gb is set?

Cuda Core: 2048 Core → 1792 Core
Tensor Core: 64 Core → 56 Core

Question 4)
Is there any way to confirm the number of Cuda core and Tensor Core?

6

Hi,

Yes, CUDA core and Tensor Core will change when emulating.
You can run deviceQuery sample to check the current hardware status.

Thanks.

7

Dear @AastaLLL,

Thank you for your information.

Can I get Tensor Core Number by deviceQuery sample?
I confirmed there is a code for Cuda core as follows.

    printf("  (%03d) Multiprocessors, (%03d) CUDA Cores/MP:    %d CUDA Cores\n",
           deviceProp.multiProcessorCount,
           _ConvertSMVer2Cores(deviceProp.major, deviceProp.minor),
           _ConvertSMVer2Cores(deviceProp.major, deviceProp.minor) *
               deviceProp.multiProcessorCount);

However, I cannot find the code for Tensor Core.

Regards,
hiro

8

Hi,

You can calculate it based on the CUDA cores amount.

On Jetson, GPU is composed of TPCs.
One TPCs have 256 CUDA cores + 8 Tensor Cores.

So for example, if 1792 cuda cores are reported.
This indicates that 7TPCs are enabled in the devices and it should have total of 56 Tensor Cores.

Thanks.

1 Like
9

Dear @AastaLLL ,

Thank you for your information.
I understood.

Regards,
hiro

11

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.