Efficient computing

[ Home ]

One of the main reasons why we have developed Collective Knowledge Framework (and previous cTuning technology) is to collaboratively enable faster, more energy efficient, smaller, cheaper and more reliable self-tuning adaptive computer systems. Such systems, in turn, should be powerful enough to help us get back to our original research on AI via efficient and compact neural networks as described in our motivation paper.

CK provided to be useful and we now partner with many academic and industrial partners to make it happen:

• ARM TechCon'16 presentation, CK vision paper at DATE'16
• dividiti (UK)
• ARM (see testimonial at page 17)
• Imperial College (UK)
• leading conferences including PPoPP,PACT,CGO
• TETRACOM (EU)
• General Motors (USA)
• non-profit cTuning foundation (France)

The community continuously shares various workloads (programs, benchmarks, kernels, codelets, data sets) with JSON meta information and JSON API, SW/HW co-design exploration plugins, compiler descriptions, and related packages in the CK format to help researchers quickly prototype their ideas from the shared components, compare with existing techniques, and reuse shared statistical analysis plugins, rather than wasting enormous amount of time on rebuilding their own experimental setups.

We hope that such approach will also help eventually make software/hardware benchmarking, autotuning and co-design practical and reproducible. In fact, we hope that CK will help move from ad-hoc and non-representative benchmarks to continuously evolving set of shared and realistic applications, kernels and data sets.

At the same time, the community can take advantage of powerful predictive analytics accessible locally or via unified CK web service to automatically find representative sets for a given user task as describe in our vision publications.

Feel free to join this community effort or tell your colleagues. You can always get in touch with the community via this public mailing list or LinkedIn group.