This milestone objective was to deliver an initial report on the applications and micro-benchmarks selected for the project as well as the measured performance on state-of-the-art architectures.
At least 9 different sub-algorithms reported in deliverable D7.1 have been validated.
MS2: Characterization of micro-benchmarks's resilience to error
Milestone MS2 has been delivered on M18.
The goal of this milestone was to provide a characterization, in terms of error resilience, of the algorithmic kernels to be targeted by the transprecision computing techniques on the five problem domains considered in the context of the OPRECOMP project.
A qualitative analysis of error resilience for all micro-benchmarks considered in the project has been elaborated, distinguishing algorithms which are very sensitive to errors from others that can potentially recover for many errors.
MS3: Transprecision Software stack
The aim of this milestone was to create an initial version of software stack for use by the partners that contain multiple modules for static and dynamic precision tuning libraries. The initial version supports incorporated transprecision capabilities based on the new hardware being developed in the project and abstracts these to provide application developers with the capability to explore realisations.
Future work will look to provide developers with the capability to automate the exploration process by abstracting the underlying transprecision techniques. In addition, various modules will be integrated to provide one cohesive framework.
MS4: Initial version of system architectures completed
Milestone MS4 has been delivered on M21
The aim of this milestone was to deliver a hardware architecture that allows fine-grained control of both temporal and spatial precision, in order to achieve energy-precision tradeoffs over a wide range. Investigations have been carried out with respect to disruptive technologies, such as heterogeneous 3D memory stacks (D emerging new memory technologies and their interfaces and QoS-aware adaptive memory controllers. Special focus has been put on approximate storage concepts to increase energy efficiency.
MS5: Prototype version of transprecision algorithms completed
MS5 has been delivered at M24.
The goal of this milestone was to deliver legacy software able to run transprecision kernels. First prototypes of innovative algorithms have been developed, covering five problem domains: data assimilation, sparse linear algebra, deep learning and optimization, big data storage and compression, and graph analytics, aligned with the three applications selected for the project.
MS9: Transprecision Application Benchmarks
MS9 has been delivered at M48.
The goal of this milestone was to complete three major benchmarks with one order of magnitude reduction in time and/or energy-to-solution.We completed the following:
A demonstrator processing transprecision bio-signals at the edge has been simulated and showed results at least two orders of magnitude improvements in performance and energy efficiency(D8.3)
A transprecision Brain Machine Interface (BMI) speller that detects brain response to a given visual stimulus: we obtained a 4.3 processing speed-up with respect to the single-precision version.
A transprecisioncommincation mechanism for kw-pilot. Evaluation results show that it can increase the effective throughput of OpenCAPI by up to 7.4× for various transprecision data-types.