Comparison of ResNet50v2 on Vitis 2025.2 AI 6.2 Versal VE2302 NPU and on ST STM32N6-DK NPU

Description

Main objective of this application note is to compare inference performance and power consumption of the new NPU IP for the AMD VersalTM XCVE2302 device on Trenz Electronic TE0950 evaluation board with the inference performance and power consumption of the ST STM32N6 NPU on the STM32N6-DK for the Resnet50v2 model.

Reference data for ST STM32N6 NPU on the STM32N6-DK for the Resnet50v2 int8 model, Imagenet datset with image resolution 224x224x3 can be found in the ST AI Model Zoo.

The ST STM32N6 1 GHz NPU has 300 MAC units and maximum theoretical performance is 0.6 TOPs (int8). The maximal theoretical (NPU only) power efficiency is 3.0 TOPs/W.

The ResNet50v2 int8 inference model has been quantized in int8 using tensorflow lite converter and implemented on the STM32N6 NPU. The dataset is Imagenet and image resolution is 224x224x3. Model has 1000 classes. The Top 1 Accuracy is 67.99 %.

TE0950-03 fits into the 10-15 W system power range. STM32N6-DK fits into the 2-3 W system power range.

TE0950-03 current cost is 1204 euro. STM32N6-DK current cost is 199 euro.

TE0950-03 has 237x higher img/s performance for the int8 Restnet50v2 inference.

Package Summary

Title Comparison of ResNet50v2 on Vitis 2025.2 AI 6.2 Versal VE2302 NPU and on ST STM32N6-DK NPU
Filenames VE2302_NPU_IP_O00_A032_M1_sd_card.zip (869 MB)
TE0950_2025_2_AI_6_2_A032_and_STM32N6_Resnet50v2_v1_0.pdf (660 kB)
License See application note for licensing conditions.
Package content ZIP archive with support files as described in application note.
Size ZIP file: 911 754 801 Bytes
PDF file: 676 322 Bytes
Required tools
& platform
AMD Vitis 2025.2
Installation notes See application note

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.