Solution Performance

Discover the Benefits

  • Significantly reduced face recognition inference latency

  • Maintain SLAs for accuracy loss



AsiaInfo provides big data and AI solutions to all three telecom carriers in China.  As the key AI product, Asiainfo Aura is a Machine Learning (ML) and Deep Learning (DL) development platform that is driven by telecom sector data.  Within this platform, the function of Face Recognition uses the Keras VGG16 model for image recognition to identify the human face online.

Visit ›


Deploying facial recognition solutions often faces two bottlenecks: network bandwidth and computing capabilities, that negatively impact deep learning inference throughput and latency, thereby resulting in less than optimal user experiences.


For facial recognition, the VGG16 model was optimized on Intel Caffe. Compared to FP32, Intel® DL Boost (delivered by Vector Neural Network Instructions (VNNI)/INT8) optimizations helped achieve a 2.08X speedup in inference latency, for the same batch size and same instance (see chart)123, while keeping to the desired accuracy requirements.



Significantly reduced face recognition inference latency, delivering better user experience – AsiaInfo customers will benefit from improved performance (i.e., lower latency), while maintaining SLAs for accuracy loss.

Product and Performance Information


Performance results are based on testing by Intel and AsiaInfo on 1/3/19 and may not reflect all publicly available security updates. No product or component can be absolutely secure.


Asiainfo Face Recognition* (self-defined workload); OS: Red Hat Enterprise Linux* 7.4 Kernel 4.19.3-1.el7.elrepo.x86_64. Testing by Intel and Asiainfo completed on Jan 3, 2019. Security Mitigations for Variants 1, 2, 3 and L1TF in place. 

TEST SYSTEM CONFIG: 2nd Gen Intel® Xeon® Platinum processor 8260L, 2.3 GHz, 24 cores, turbo, and HT on, BIOS 1.0180, 384GB total memory, 24 slots / 16GB / 2666 MT/s / DDR4 LRDIMM, 1 x 800GB / Intel® SSD DC S4500; Intel® Optimization for Caffe*.


Performance varies by use, configuration and other factors. Learn more at