TE0950, Vitis 2023.2.1, AI 3.5 comparison of AMD DPUCV2DX8G with AMD DPUCZDX8G

Description

This application note evaluates performance of the AMD Vitis AI 3.5 image classification examples accelerated by the AMD DPUCV2DX8G DPU.

The AMD Versal Deep Learning Processing Unit (DPUCV2DX8G) is a configurable computation engine optimized for convolutional neural networks in AMD Versal Adaptive SoCs devices with acceleration by the AI Engine-ML tiles.

The performance of AMD Vitis AI 3.5 image classification examples accelerated by the AMD DPUCV2DX8G is compared with performance of the same AMD Vitis AI 3.5 image classification examples accelerated by the AMD DPUCZDX8G HW IP implemented in the PL logic of the device without the acceleration by the AI Engine-ML tiles.

The DPUCZDX8G is the Deep learning Processor Unit (DPU) designed by AMD for the Zynq UltraScale+ MPSoC devices. However, the DPUCZDX8G can be also compiled from the RTL source code to PL part of the AMD Versal XCVE2302 device.

Both compared AMD DPUs (DPUCV2DX8G and DPUCZDX8G) support set of highly optimized instructions, and supports most convolutional neural networks, such as VGG, ResNet, GoogLeNet, YOLO, SSD, MobileNet, FPN, and others.

The AMD DPUCV2DX8G (1 GHz) with 20 AI-ML tiles provides (in average) eight times shorter inference times in comparison to the AMD DPUCZDX8G with configuration B2304 (150 and 300 MHz) on the same Versal XCVE2302 device.

Both DPUs can accelerate wide range of quantized AI classification models.

Both DPUs can use identical run-time API and evaluation applications.

Both DPUs support only inference in Int8. Both DPUs are designed for processing of images with batch size equal to 1.

Package Summary

Title TE0950, Vitis 2023.2.1, AI 3.5 comparison of AMD DPUCV2DX8G with AMD DPUCZDX8G
Filenames TE0950_2023_2_1_dpucv2dx8g_AI_3_5_v1_0.pdf (880 kB)
License See application note for licensing conditions.
Size PDF file: 902 002 Bytes
Required tools
& platform
AMD Vitis 2023.2.1

Result Category

Project number Project acronym Year RIV category Comment
9A25004 NeAIxt 2026 Gfunk Functional sample (demo)

Contact Person

Please, do not hesitate to contact Jiri Kadlec to obtain more information.