Periodic Reporting for period 3 - TEMPO (Technology and hardware for neuromorphic computing)
Periodo di rendicontazione: 2021-05-01 al 2023-01-31
TEMPO (Technologies and hardware for neuromorphic computing) developed technologies and hardware solutions to broaden the applicability and strengthen the European ecosystem of integrated neuromorphic computing, enabling Artificial Intelligence in Edge devices and applications.
To achieve this, TEMPO worked according to 4 project objectives.
(1) Identify and enhance emerging semiconductor technologies that best fit neuromorphic systems
(2) Optimize fab infrastructure and wafer exchange between foundries and RTOs to strengthen the European supply chain
(3) Validate and compare neuromorphic concepts based on KPIs on all levels between technologies and applications
(4) Extend the technology roadmap towards future neuromorphic AI hardware and applications and to strengthen the European ecosystem
To achieve this, TEMPO worked according to 4 project objectives.
(1) Identify and enhance emerging semiconductor technologies that best fit neuromorphic systems
(2) Optimize fab infrastructure and wafer exchange between foundries and RTOs to strengthen the European supply chain
(3) Validate and compare neuromorphic concepts based on KPIs on all levels between technologies and applications
(4) Extend the technology roadmap towards future neuromorphic AI hardware and applications and to strengthen the European ecosystem
- Gained fundamental understanding quantifiable correlation between memory technology features and their neuromorphic Implementations. This was based on designed test vehicles and process flows which are leveraged to embed Non-Volatile Memories in neuromorphic and AI applications. Memories on top of CMOS base wafers have been fabricated, including the integration of STT-MRAM arrays, the fabrication of OxRAM memory based on FD28 BW from STM and the integration of FeRAM on GF 28nm.
- Enhanced 3D interconnection technology insights, design flows and validated technologies.
- Explored the design space of the neuromorphic technologies and designed several DNN & SNN core building blocks, accelerator architectures, and fabricated related ASICs by leveraging the memory technologies towards the application demonstrators.
- Explored, designed and implemented DNN & SNN algorithms in coordination with the hardware developments.
- Basewafer exchange procedures and data protocols defined and successful exchange trials completed allowing smooth collaborative processing paving the way to the future Hardware TEF (Test and Experiment Facilities) for AI at EU level.
- 11 demonstrators comprising TEMPO technologies were developed and validated. These demonstrators operate on 4 application domains in which EU has a major stronghold.
- Holistic validation/benchmark developed across reference platforms and TEMPO technologies.
- Enhanced 3D interconnection technology insights, design flows and validated technologies.
- Explored the design space of the neuromorphic technologies and designed several DNN & SNN core building blocks, accelerator architectures, and fabricated related ASICs by leveraging the memory technologies towards the application demonstrators.
- Explored, designed and implemented DNN & SNN algorithms in coordination with the hardware developments.
- Basewafer exchange procedures and data protocols defined and successful exchange trials completed allowing smooth collaborative processing paving the way to the future Hardware TEF (Test and Experiment Facilities) for AI at EU level.
- 11 demonstrators comprising TEMPO technologies were developed and validated. These demonstrators operate on 4 application domains in which EU has a major stronghold.
- Holistic validation/benchmark developed across reference platforms and TEMPO technologies.
TEMPO reached all its objectives, aiming to broaden the applicability of integrated neuromorphic hardware by improving energy efficiency with emerging memory technologies, architectures and algorithms in novel neuromorphic hardware implementations.
(1) After initially identifying the most promising NVM technology options, important initial steps towards mature NVM technology for neuromorphic AI acceleration were taken and are now being taken up beyond the project. The drawbacks and benefits of each technology and how they relate to the requirements imposed by the different markets and the use cases that were defined in each of them are now more clear.
(2) Smooth exchange of base wafers between commercial foundries and the RTO’s for memory processing has been enabled through establishing acceptance criteria for base wafer delivery from commercial foundries. This needed insight into how the RTO cleanrooms are managed in terms of metallic contamination. An extensive study was performed on general contamination protocols and on Lithium in particular.
(3) TEMPO validated and compared neuromorphic concepts based on KPIs on all levels between technologies and applications. To understand how technology improvements translate into system-level performance and vice-versa, TEMPO combined a bottom-up “technology-to-application” approach with a top-down “application-to-technology” approach. This translated into a more holistic approach/benchmark which supports the European ecosystem in their development and market introduction of neuromorphic technologies.
(4) TEMPO has identified emerging neuromorphic technology and application roadmaps from the cross-section of the bottom-up and top-down perspectives and in alignment with other related initiatives. These roadmaps are paramount for the European Commission and the global ecosystem to support the development of advanced compute technologies in Europe. The project built strategic alliances with other initiatives and co-organized four workshops, each focusing on different audiences. TEMPO published several books and published more than 100 articles, papers, abstracts, press releases, etc. The TEMPO concepts were broadly communicated through direct and broadcasting channels (e.g. social media).
(1) After initially identifying the most promising NVM technology options, important initial steps towards mature NVM technology for neuromorphic AI acceleration were taken and are now being taken up beyond the project. The drawbacks and benefits of each technology and how they relate to the requirements imposed by the different markets and the use cases that were defined in each of them are now more clear.
(2) Smooth exchange of base wafers between commercial foundries and the RTO’s for memory processing has been enabled through establishing acceptance criteria for base wafer delivery from commercial foundries. This needed insight into how the RTO cleanrooms are managed in terms of metallic contamination. An extensive study was performed on general contamination protocols and on Lithium in particular.
(3) TEMPO validated and compared neuromorphic concepts based on KPIs on all levels between technologies and applications. To understand how technology improvements translate into system-level performance and vice-versa, TEMPO combined a bottom-up “technology-to-application” approach with a top-down “application-to-technology” approach. This translated into a more holistic approach/benchmark which supports the European ecosystem in their development and market introduction of neuromorphic technologies.
(4) TEMPO has identified emerging neuromorphic technology and application roadmaps from the cross-section of the bottom-up and top-down perspectives and in alignment with other related initiatives. These roadmaps are paramount for the European Commission and the global ecosystem to support the development of advanced compute technologies in Europe. The project built strategic alliances with other initiatives and co-organized four workshops, each focusing on different audiences. TEMPO published several books and published more than 100 articles, papers, abstracts, press releases, etc. The TEMPO concepts were broadly communicated through direct and broadcasting channels (e.g. social media).