Vous êtes ici : Accueil > Projet TEMPO

Europe | Outils ＆ instruments de recherche


TEMPO

When deploying Artificial Intelligence (AI), specific hardware solutions capable of assimilating multiple data from different sensors (radar, lidar, camera) must be developed within the specified constraints of reduced time, low latency, low power consumption and high resolution. The aims of the TEMPO project were to design and combine performing transistors and new emerging memories to build novel neural networks capable of analysing complex situations such as object detection in real time (automotive, space, health and multiple applications).


Publié le 22 avril 2021

​

TEMPO: Technology and hardware for nEuromorphic coMPuting

When deploying Artificial Intelligence (AI), specific hardware solutions capable of assimilating multiple data from different sensors (radar, lidar, camera) must be developed within the specified constraints of reduced time, low latency, low power consumption and high resolution. The aims of the TEMPO project were to design and combine performing transistors and new emerging memories to build novel neural networks capable of analysing complex situations such as object detection in real time (automotive, space, health and multiple applications).





 

Starting date: May 2019 > Apr.. 2022  Lifetime:42 months

Program in support : ECSEL 2018

 

Status project: in progress

CEA-Leti's contact:

 

Project Coordinator: Imec (BE)


Partners:  

  • AiCortex, University of Zurich (CH)
  • Bosch, Fraunhofer, Infineon, Innosent,
    University of Dresden, Valeo, Videantis (DE)
  • ST Microelectronics, Thales Alenia Space, Valeo (FR)
  • Atogear, Imec-NL, Philips Health,
    Philips Research (NL)


​

Publications 


  • The TEMPO project just started in May 2019.


​

Investment:  € 33 m.

EC Contribution: € 10.5 m.


Stakes

  • During the TEMPO project, CEA institutes Leti and List contributed right from novel design to hardware implementation, in CEA-Leti’s 300mm cleanroom dedicated to Spiking Neural Network (SNN) applications, mainly in support of a VALEO ambitious user case implementing a LIDAR sensor.

  • CEA-Leti designed new “neuromorphic” structures to evaluate 1T-1R (one transistor – one resistive element) on top of advanced CMOS logic; their production, based on 28 nm FDSOI technology, being supported by ST Microelectronics (Crolles, France). 

  • Following development and implementation of the specific memory module in the CEA-Leti 300mm cleanroom, we finalised the interconnections to allow testing, part of which was supported by CEA-List (Saclay, near Paris) using so-called “N2D2” hardware. The TEMPO project enabled benchmarking of different memories, so CEA-Leti can now evaluate PCRAM and OxRAM resistive memories for a suitable design targeting SNN applications.

OBJECTIVES

The TEMPO project brought together the three premier European RTOs, industrial fabrication facilities and leading application partners in the field of neuromorphic computing, which is subject to intense global competition in view of upcomingintelligent machines. The underlying concrete aims of the TEMPO project were to:

  • Enable joint development of participating European RTOs, foundries and leading application development companies in identifying emerging semiconductor technologies that best fit neuromorphic hardware and address relevant applications

  • Evaluate current concepts for implementing neuromorphic hardware based on Key Performance Indicators (KPIs) at
    device, architecture and application levels, including power consumption, silicon area/cost, latency, energy for a given
    application task, memory bottlenecks, manufacturing challenges and operating frameworks

  • Extend the technology roadmap driven by Integrated Circuits (ICs) designed specifically for AI and Machine Learning (ML) applications by evaluating and demonstrating the applicability of emerging technologies able to provide scalable power, performance and area benefits

  • Exchange wafers between foundries and participating RTOs to facilitate demonstration of functional neuromorphic
    chips, thereby allowing the use of the extensive know-how of European R&D organizations for future products, while
    maintaining contamination-free high volume production

  • Quantify the capability of the most prevalent neuromorphic hardware implementations by targeting a broad algorithmic
    spectrum and isolating the critical sections of each algorithm; this would include Deep Learning (DL) inference (e.g. CNNs) and SNNs

  • Enable the European semiconductor industry to maintain its position at the cutting edge of neuromorphic chip
    development.

    IMPACT

  • The TEMPO project has had a strong impact because it provided hardware suitable for supporting complex situations specified by industrial partners in different fields (automotive, space, health, food, industrial, etc.). Different types of memory (magnetic, ferroelectric, resistive) can be
    benchmarked not only for demonstrators, but also for providing a scorecard for the ecosystem and for paving the way towards facilitating future recommendations for new applications.

  • These actions will associate the best memory hardware solution to a given requirement based on different criteria in terms of power consumption, latency, resolution, etc.

  • The TEMPO project partners also aligned their different assets, in terms of hardware equipment and processes, to their mutual benefit in long term cooperation especially by the three prime Research, Technology and Organisation(RTO) stakeholders, namely CEA-Leti and the Imec and Fraunhofer institutes.

  • From a long-term perspective, the TEMPO project opened up affordable, fast-track access for European SMEs and systems houses, which will have a key impact on European ecosystem sovereignty.