October 23: Japan’s Preferred Networks and US-based Tenstorrent have been selected for a Japanese government-backed research programme on advanced AI processors. The project brings together chiplet design, photonics-electronics integration and advanced packaging, with research planned over five years. It is a development programme rather than the launch of a finished commercial chip.
Key highlights
- NEDO published the implementation decision on October 20 under its post-5G semiconductor programme.
- PFN will work on AI chiplets using three-dimensional memory technology.
- Tenstorrent will develop open-architecture CPU chiplets supporting confidential computing.
- The announcement does not establish a commercial release date or independently measured performance advantage.
What NEDO has selected
NEDO’s notice identifies the selected theme as xPU design and development incorporating photonics-electronics integration. It places the work within Japan’s research programme for stronger post-5G information and communications infrastructure and advanced semiconductor manufacturing technologies.
The agency specifies a research period normally within five years, or 60 months, from the start. Its notice describes the programme’s objective of securing technologies for domestic design and manufacture of advanced semiconductors. Selection establishes the planned research structure; it is not proof that a manufacturable system has already been demonstrated.
How the companies divide the work
Preferred Networks’ announcement assigns it the AI-chiplet work, including high-bandwidth designs using three-dimensional memory. Tenstorrent’s announcement describes its responsibility for CPU chiplets built on an open architecture and supporting confidential computing. Both companies will work on system integration using photonics-electronics and packaging technologies.
The companies list universities and public research organizations as subcontractors, including the University of Tokyo, Kobe University, Toyo University, the Japan Agency for Marine-Earth Science and Technology and the Japan Atomic Energy Agency. Their participation supplies potential research expertise and applications; it should not be read as a commitment to purchase a finished system.
At a glance: the architecture in plain English
| Term | What it means in this project | Question for future results |
|---|---|---|
| Chiplets | Separate specialized components assembled into a larger system. | Can the components work efficiently together? |
| 3D memory | Memory technology intended to provide high data bandwidth. | What bandwidth and power figures are demonstrated? |
| Photonics-electronics integration | Combining optical and electronic approaches to move data. | How well does the integrated system perform? |
| Confidential computing | Protection-oriented features planned for CPU chiplets. | What protections are specified and tested? |
Why the project targets AI infrastructure
TNGlobal’s October 22 report describes bandwidth and power constraints as reasons for pursuing this architecture. Moving information among processors and memory is part of the system-design challenge, not merely an issue of the number of compute cores.
The report also notes that the announcement did not disclose prototype performance, a foundry arrangement or a commercialization timetable. Those omissions are important when assessing claims of a competitive breakthrough. A promising architecture still needs hardware, packaging, software and reliable production to work together.
Research funding is not a product endorsement
The participating companies naturally describe the programme in terms of its potential. Those ambitions should remain separate from verified results. We do not claim that the proposed processor is faster than a named rival, cheaper to operate or ready to ship.
NEDO’s October 20 notice, PFN’s October 20 release and Tenstorrent’s October 21 English announcement refer to the same underlying project. They are separate documents, not three independent experimental validations. TNGlobal provides reporting on the announcement, not a hands-on benchmark.
What to watch
Useful evidence will include prototype demonstrations, published power and bandwidth measurements, software support, manufacturing details and a commercial roadmap. For Asia’s AI industry, the immediate news is the selection of a multi-year research partnership. Its eventual competitive importance will depend on those measurable outputs.
Featured image: LG Electronics / Wikimedia Commons, CC BY 2.0. Licence information. Displayed with responsive cropping.
Sources
- NEDO: Implementation structure selected for advanced semiconductor development (Japanese)
- Preferred Networks: PFN and Tenstorrent selected for photonics-electronics xPU research (Japanese)
- Tenstorrent: PFN and Tenstorrent selected for NEDO project
- TNGlobal: PFN and Tenstorrent win NEDO xPU research project
The Tech Desk covers technology, artificial intelligence and cybersecurity. Stories are checked against at least two credible sources before publication.
