White Papers

12 Deep Learning Inferencing with Mipsology using Xilinx ALVEO™ on Dell EMC Infrastructure
Evaluation Results
Ease of use
Switching from CPU/GPU to Zebra running on an ALVEO™ board deployed for inference was
surprisingly simple, carried out via a single Linux command. No FPGA tools or knowledge was
necessary.
During the evaluation, eight neural networks were executed without any changes, proving that
the Zebra/ALVEO U200 is the most versatile FPGA solution for neural networks.
The application was based on TensorFlow workflow. No change was required to the application,
nor to the neural network, nor to the training, making the transition effortless and free from
engineering resources.
The quantization used did not require any user intervention, which emphasizes the ease-of-use
of Zebra.
Performance
Eight popular neural networks were processed during the evaluation. The performance of each
of them when run on the Zebra mapped on an ALVEO U200 installed in a Dell PowerEdge
R740/R740xd servers is charted in figure 8.
Figure 8. Comparing deep learning models using Mipsology Zebra stack.
Accuracy:
Running the complete set of images from ImageNet, the evaluation showed a difference of
accuracy of less than 1%. For some of the NN the results were better than using FP32.
However, it should be noted that the purpose of the evaluation was not to reproduce the best
accuracy results for each network, rather to compare Zebra against a CPU/GPU solution
running FP32 with no changes to the networks.