About Orin CPU performance vs Xavier CPU

Xavier AGX CPU:8 NVIDIA Carmel processor cores @2.26G;
Orin AGX CPU:12× Arm Cortex-A78AE cores @2G

How to quantitatively evaluation the CPU performance improvement :Orin vs Xavier ?

1 Like

I’m checking with internal team, will update back once clarified.

Please see the comparison chart in this tech brief that shows Spec2K6 benchmarks with the new 2.2 GHz frequency: https://www.nvidia.com/content/dam/en-zz/Solutions/gtcf21/jetson-orin/nvidia-jetson-agx-orin-technical-brief.pdf

What’s the Xavier version in the test : JAX@2.26Ghz or JAXi@2.03GHz ?

In the tech brief:

8x DL/AI Performance: Xavier 32 Dense TOPS vs Orin 275 Sparse Tops ?
9x DLA Performance:Xavier 11.4 Dense TOPS vs Orin 105 Sparse Tops ?

Can Nvidia make a clear about Orin Sparse Tops,Dense Tops ?

from above Orin 64GB 275 Sparse TOPS = 138 Dense TOPS;

Sparse TOPs is double the Dense TOPs. Sparsity is only supported in Ampere, and this is a new feature with Orin. It is described in the tech brief in section 3rd Generation Tensor Cores and Sparsity on page 5. A further breakdown can be found in our Jetson AGX Orin Series Data Sheet of both Sparse and Dense for the total AI Performance, the GPU alone, and the DLA alone.

The 8x performance increase represents the performance comparison between the overall GPU and DLA performance combined on Orin vs. Xavier. The 9x performance increase represents the performance increase on the DLA. With sparsity you could get up to 2x the dense performance.

It brings on average about 3.4x more performance across the full breadth of usages tested;
resnet50 performance 2X improve;

From above,we can get future performance 4.9X;

It’s much more gap to 8x/9x performance impove;

The benchmarks listed in the above table and charts are Dense Benchmarks. For more details on the sparsity feature please refer to this Accelerating Inference with Sparsity Using the NVIDIA Ampere Architecture and NVIDIA TensorRT | NVIDIA Technical Blog. Xavier is based on older Volta GPU architecture that does not support sparsity and will not deliver this perf boost for sparse networks. Sparse networks are becoming more common and AI applications that use sparse networks can harness up to 275 TOPS on Orin. On Xavier, they will be limited to 32 TOPS max. For dense AI networks, Orin offers 138 TOPS while Xavier is limited to 32 TOPS.

With the increasing adoption of Sparse neural networks in AI applications, we believe comparing Orin’s INT8+Sparsity compute of 275 TOPs to Xavier’s 32 TOPs is a fair comparison and it highlights the performance ceiling that each platform offers when all the features of each platform is enabled. Even on neural networks in the charts, Orin delivers 3.3X the performance of Xavier and this speedup will continue to increase as these neural networks are further tuned for Orin and adopt features like Sparsity. The chart highlights that we estimate this speedup to improve to 5X. Of course, there are other system bottlenecks that may not allow an ideal 8X speedup but we are continually optimizing our JetPack components towards that goal.

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