So what's up with Orin AGX clock speeds at 15W? And other questions

Hello, just asking after watching this video from JetsonHacks about Orin AGX and reading some of the replies.

He reports back that Orin AGX at 15W drops the CPU down to 4 Cores at 1Ghz and the GPU down to only 420MHz?

On that profile, how many CUDA Cores/SMs are active? As that is a massive drop in clocks across both and at least core count in the case of the CPU.

Are the 2 DLA’s and the PVA drawing back that much power to prevent all CPUs running at 1Ghz and the GPU from even hitting half of its potential clock max? (At whatever CUDA/SM count it’s set to at 15W)

Or is it just a limitation of the Process node (Samsung 8nm like Desktop Ampere???) that prevents it from hitting Xaiver-level clocks at the config it is at at 15W (Not saying it’s weaker than 15W Xaiver, but surprised it is seemingly clocked lower by a notable amount if my memory serves me right)

Either way though, my other question is can anyone run something like 3DMark or other standardized benchmarks for the GPU? I am very curious as to what benefits over what should be directly relative to Orin’s 16SM GPU (The RTX 3050 Laptop) the 1.5x increase in L1 Cache and the doubling of L2 Cache per-GPC would result as mentioned in the documentation for Orin’s specifications.

I feel this should be looked into because AMD added extra cache to RDNA1 and got most of their performance increase with RDNA2 out of it so wondering if something similar can happen with Ampere having that extra cache added onto it here in Orin.

The clock’s for that test I’m curious for are 768MHz, 1GHz, and 1.3GHz, but that’s just me

We did something similar for previous generation products like Xavier. Where the lower preconfigured power modes have different GPU and CPU Frequencies and core counts. We choose the power modes to optimize for generic usecases that can leverage various components on the module within the specified power budget. See the nvpmodel data for Xavier here: NVIDIA Jetson Linux Developer Guide : Clock Frequency and Power Management | NVIDIA Docs

These nvpmodel docs will be updated with Orin details with the upcoming JP 5.0 DP. Customers can customize their own usecases between 15W and 60W on Orin depending on what there end application is using our nvpmodel to optimize the various components like the GPU and CPU for their usecase and power budget.

We will have some benchmark data that we have performed on Orin using our pretrained models to share in coming future.
In terms of other GPU benchmarks, this is something we will look to provide more information afterward.


Thank you Kay

Are there any incidations however on what the full details are for the Orin AGX kit at 15W like in the video linked?

What I can tell is it’s 4 CPU cores enabled at 1Ghz and the GPU with however many CUDA cores enabled are paired down to 420MHz.

But what is the actual SM/CUDA count at 15Ws, and are the DLAs, and the PVA.etc active at 15W? And if so what clocks could be expected when the DLAs and PVAs are off at 15W?

My main end-question is would it be possible to have Orin AGX run with 8 Cores at 1Ghz with 8-14SMs (CUDA, RT, and Tensor cores in them) at 1Ghz as well at 15-20W even if it required turning off the DLA’s and the PVA’s

As the workload I would be interested into looking for a AGX kit would be more GPU-reliant.

Please take a look at the nvpmodel for the various power modes.

It has the usecase we have defined for 15W:

For the customized more gpu centric usecase that you describe, you can create your own nvpmodel, and turn off the components that you don’t need, and identify what the power looks like with our Jetson AGX Orin Developer Kit. Please try out our new Jetson Power GUI: Jetson Orin NX Series and Jetson AGX Orin Series — Jetson Linux<br/>Developer Guide 34.1 documentation

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