WP2 – Mathematical Theory and Physical Foundations
UNIPG , IBM, UJI
set the mathematical and physical backgrounds where the fundamental limits in energy efficiency are explored and assessed, and where the role of the unavoidable fluctuations (noise), instead of being detrimental, is put into work for producing efficient and reliable computation results.
Modeling of the impact of transprecision on energy consumption, considering fundamental and technological limitations.
Laying out the Mathematical foundations of error propagation in variable precision to be use in algorithms of WP5.
WP3 – Disruptive Technology Modeling and Exploration
TUKL, ETH Zürich, UNIPG
carry out investigations with respect to disruptive technologies, such as heterogeneous 3D memory stacks (DRAM+NVM), emerging new memory technologies (e.g. ReRAM and non-conventional NEMS memory devices), their interfaces (wide I/O, serial, optical), and QoS-aware adaptive memory controllers.
Special focus on approximate storage concepts to increase energy efficiency.
Developing Models on different abstraction-level to permit design exploration for transprecision memory architectures. Outcome will be the best memory system for each selected computing architecture.
Providing Memory subsystems for the architecture and for demonstrator platforms (WP8) for mW and kW systems.
WP4 – Architecture and Circuits
ETH Zürich , IBM, CEA, UNIBO, QUB, GWT, TUKL
develop 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. This hardware architecture will be implemented and optimized for two different workload conditions, one for mW range of operations in the extreme low-power operation, and the other targeting a single node for HPC applications in the kW range.
define a hardware abstraction layer that enables access to the precision control through a software runtime developed as part of WP6 allowing closed loop control of the precision vs operating cost during runtime. Both the mW and kW architectures will rely on computation, storage and a communication interface with controllable precision.
develop synergies between heterogeneous computing elements, storage hierarchy and interconnect so that overall precision constraints can be balanced.
WP6 – Software Environments and Tools
QUB, IBM, ETH Zürich, CEA, UNIBO, CINECA, UJI, GWT, TUKL
build a software stack for developing, configuring and executing trans-precise programs.
provide input to and build on the definition of hardware platforms in the mW and kW range (WP4) and the definition of disruptive memory technologies (WP3).
define a programming interface (API) that is used by and based on the properties of novel algorithms (WP5) and applications (WP7).
WP5 – Algorithms
UJI, IBM, UNIPG, CINECA, QUB
develop innovative algorithms that, in combination with the hardware architectures produced and the run-time support, yield reduction of time- and energy-to-solution for the applications targeted. The algorithms will cover 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.
leverage models and emulators To explore further directions.
strong connection with WP2 and the mathematical analysis of error propagation in the algorithms.
WP7 – Applications
IBM,ETH Zürich , UNIBO, CINECA, UJI, QUB
select relevant real-world applications from Data analytics (e.g., knowledge graphs, search and clustering), brain-inspired machine learning (e.g., neural network) and complex HPC simulations (e.g., molecular dynamics and material science simulation) domain
reach order of magnitudes in reduction of time- and energy-to-solution by exploiting the full transprecision framework developed in OPRECOMP project.
WP8 – System Integration and demonstration
IBM, ETH Zürich , UNIPG, UNIBO, QUB, GWT, TUKL
develop architecture and circuits for transprecision computing, targeting an IoT type platform (mW range) and a HPC type platform (kW range).
demonstrate impact on the applications developed in WP7, while using the technology, software layers and methodology developed in WP2-3-4-5-6.