Collaboration delivers competitive products faster
Posted: 21 Jun 2019 | 10:32
Optos is a leading provider of medical products and services to eyecare professionals. A collaboration between Optos, EPCC, and Intel has led to significant improvement in the performance of a software algorithm used in a product under development by the company.
The project set out to identify a suitably priced Intel multicore processor on which to run the company’s code, and to accelerate an important Optos machine-learning code. The project baseline was Optos code running on a specific Intel chip and the target was to reduce its runtime to below 12 seconds on the preferred chip.
EPCC, aided by Intel, achieved a runtime of 6.6 seconds on the preferred Intel chip. That’s almost two times faster than the target, over 15 times faster than the baseline software-hardware combination, and five times faster than the baseline software on the new chip. Importantly for Optos, the preferred Intel chip is 2.5 times faster than the original processor when running the optimised codes, and 25% cheaper than the target price. We also accelerated a second Optos code to run in just over five minutes on the preferred hardware, matching the acceleration figures of the main code.
All three partners benefited from working closely together on this project. The accelerated Optos codes and more affordable yet faster processor can be deployed on a production device, allowing Optos to cut costs while enhancing the performance of its product. This will further boost Optos’ market share as a whole, helping promote sales of its other related products.
Optos has also gained insight into its code and ideas for how to further improve it. The company has better understanding of advanced techniques to benchmark and optimise multi-core codes, as used by EPCC for this project, and will apply this new knowledge to its software for years to come.
The Scottish economy also stands to benefit. It is expected that this new opportunity will support ongoing product development activities at Optos’ headquarters in Dunfermline, supporting jobs in areas from research and development through to manufacturing, so enhancing Scotland’s reputation in the areas of retinal imaging and machine learning.
Intel gained insights into the performance of the MKL-DNN mathematics library employed by the popular Deep Learning library TensorFlow. In the seven months that it was active, the project established that MKL-DNN has been accelerated by up to 4.5 times in our use cases, information which is important for the uptake of the library. EPCC also provided profiling information gleaned from the Optos use cases and this will help Intel further improve the performance of MKL-DNN. Prompted by the project findings, Intel improved Windows documentation in particular, which will increase the uptake of TensorFlow and MKL-DNN.
EPCC has increased its exposure to machine-learning codes for cutting-edge industrial applications and developed in-depth understanding of the industry-leading TensorFlow library. We will exploit this new knowledge in future collaborative projects, notably in our leading role in the Edinburgh and South East Scotland City Regional Deal.
“This project has brought direct benefits to our business, both by enabling us to develop competitive products faster and by building essential knowledge within our organisation. We look forward to a continued productive relationship with EPCC.” Jano van Hemert, Optos Research Director
“We are excited to partner with Optos to identify key computing capabilities needed to accelerate their critical machine-learning codes. Intel processors demonstrated a significant boost in the performance for their codes, outperforming the target time of the currently deployed technology.”
Trish Damkroger, Vice President and General Manager, Intel’s Data Center Group
“The Optos CodeOptScotland project showcases the impact of engaging Intel for the benefit of Scotland’s economy, Scotland-based companies and Scottish academia.”
David Smith, Scottish Enterprise Director of Technology
“The success of this project highlights EPCC’s capability to translate academic excellence to long-lasting benefit for our commercial partners.”
Mark Parsons, EPCC Director
The project was part-funded by the CodeOpt Scotland programme, which was designed to increase company competitiveness by optimising the performance of business-critical software. CodeOpt is a partnership of EPCC, Scottish Enterprise and Intel.