OPRECOMP · Project

Cut Computing Energy Bills by Using Only the Precision You Actually Need

digitalTestedTRL 4

Imagine you're doing a rough sketch before painting — you don't need pixel-perfect detail at every step. Computers today waste enormous energy by calculating everything to maximum precision, even when a rough answer would do just fine for intermediate steps. OPRECOMP built hardware and software that automatically dials precision up or down depending on what each calculation actually needs, slashing energy use from tiny IoT sensors all the way up to massive supercomputers. The result: the same quality answers at a fraction of the electricity cost.

By the numbers

9 orders of magnitude

Power range covered (sub-milliWatt to MegaWatt)

EUR 5,990,510

EU research investment

10 partners

Consortium size across 6 countries

Total project deliverables produced

sub-picoJoule/operation

Target energy efficiency level

Physical demo platforms built (2 initial + 2 final versions)

The business problem

What needed solving

Computing systems today waste massive amounts of energy by performing every single calculation at maximum precision, even when intermediate steps don't need it. For data centers spending millions on electricity and IoT devices constrained by tiny batteries, this unnecessary precision translates directly into higher costs, shorter battery life, and larger carbon footprints. The problem spans from the smallest sensors to the largest supercomputers.

The solution

What was built

The project built two physical pilot platforms: a kW-range platform for HPC/data center applications and a mW-range platform for IoT/embedded devices, each delivered in initial and final versions. Alongside the hardware, the team developed software tools, algorithms, and a mathematical theory providing guaranteed error bounds for transprecision computing across 52 total deliverables.

Audience

Who needs this

Data center operators looking to reduce electricity and cooling costsIoT device manufacturers needing longer battery life without sacrificing data qualityCompanies running large-scale deep learning training workloadsHPC service providers running climate or scientific simulationsEdge computing companies deploying AI inference on low-power hardware

Business applications

Who can put this to work

Data Center Operations

enterprise

Target: Cloud and HPC data center operators

If you are a data center operator dealing with skyrocketing electricity bills and cooling costs — this project developed a kW-range pilot platform that demonstrates how transprecision computing can dramatically cut energy per operation to sub-picoJoule levels, while maintaining the same quality of results. With 10 consortium partners across 6 countries validating the approach across nine orders of magnitude (sub-milliWatt to MegaWatt), the energy savings potential applies directly to your most power-hungry workloads.

IoT and Edge Computing

any

Target: IoT device manufacturers and edge computing providers

If you are an IoT device maker struggling with battery life and heat constraints on your sensors — this project built a mW-range pilot platform proving that intelligent precision management extends device operation without sacrificing data quality. The technology covers the sub-milliWatt range specifically relevant to embedded sensors, letting your products run longer on smaller batteries while still delivering reliable deep learning and big data analytics at the edge.

AI and Machine Learning

mid-size

Target: Companies running large-scale deep learning training and inference

If you are running deep learning workloads and your GPU electricity costs keep climbing — this project proved across 52 deliverables that many intermediate calculations in neural network training don't need full 64-bit precision. The IBM-led consortium built working pilot platforms demonstrating that transprecision techniques deliver the same end-to-end application quality while consuming far less power, directly reducing your training and inference costs.

Frequently asked

Quick answers

How much would it cost to implement transprecision computing in our systems?

The project does not publish licensing fees or implementation costs. However, with EUR 5,990,510 in EU funding across 10 partners over 4 years, the R&D investment was substantial. Contact the consortium to discuss technology transfer terms and pilot pricing.

Can this scale to our production workloads?

OPRECOMP demonstrated the technology across nine orders of magnitude — from sub-milliWatt IoT devices to MegaWatt-class HPC systems. Both a kW pilot platform (for HPC-scale applications) and a mW pilot platform (for embedded/IoT applications) were built and tested. This range suggests the approach is designed for production-scale deployment.

What about intellectual property and licensing?

The project was coordinated by IBM Research and included 2 industry partners among 10 total consortium members. As a publicly funded RIA project, results are typically available for licensing. Specific IP terms should be discussed directly with IBM Research or the relevant partner holding the specific technology you need.

Will our end results lose accuracy?

A core principle of the project is maintaining end-to-end application quality. The mathematical theory developed provides error bounds with respect to full precision results, meaning your final outputs stay reliable even when intermediate steps use reduced precision. This is not approximate computing — it is precision management with guarantees.

What application domains has this been validated for?

Based on the project objective, the technology was demonstrated in IoT, Big Data Analytics, Deep Learning, and HPC simulations. The EuroSciVoc classifications also include climatic change mitigation, confirming applicability to climate modeling workloads.

How long would integration take?

The project ran from 2017 to 2020 and produced 52 deliverables including software tools, algorithms, and working pilot platforms. Based on available project data, integration timelines would depend on your specific hardware and software stack. The open nature of the project (implied by 'Open' in the title) suggests accessible tooling.

Consortium

Who built it

The consortium is led by IBM Research (Switzerland), giving it strong industrial credibility in computing hardware and software. With 10 partners across 6 countries (CH, DE, ES, FR, IT, UK), it has solid European coverage. However, the 20% industry ratio (only 2 industry partners out of 10) and just 1 SME indicate this was primarily a research-driven effort with 6 universities and 2 research organizations. For a business buyer, this means the technology has been rigorously validated academically but may need additional engineering partnership to reach a commercial product. IBM's involvement as coordinator significantly de-risks the technology transfer path.

IBM RESEARCH GMBHCoordinator · CH
UNIVERSITAT JAUME I DE CASTELLONparticipant · ES
GREENWAVES TECHNOLOGIESparticipant · FR
COMMISSARIAT A L ENERGIE ATOMIQUE ET AUX ENERGIES ALTERNATIVESparticipant · FR
THE QUEEN'S UNIVERSITY OF BELFASTparticipant · UK
RHEINLAND-PFALZISCHE TECHNISCHE UNIVERSITATparticipant · DE
EIDGENOESSISCHE TECHNISCHE HOCHSCHULE ZUERICHparticipant · CH
ALMA MATER STUDIORUM - UNIVERSITA DI BOLOGNAparticipant · IT
UNIVERSITA DEGLI STUDI DI PERUGIAparticipant · IT
CINECA CONSORZIO INTERUNIVERSITARIOparticipant · IT

How to reach the team

IBM Research GmbH in Switzerland coordinated this project. Look for the OPRECOMP project lead at IBM Research Zurich.

Order a report on this consortium See IBM RESEARCH GMBH profile →

Next steps

Talk to the team behind this work.

Want to explore how transprecision computing could cut your data center or IoT energy costs? SciTransfer can connect you directly with the OPRECOMP team and help evaluate the fit for your specific workloads.

Order a report on this topic More in Digital & ICT