From Nikkei Asia March 31:
Production to be outsourced to Rapidus as Japan eyes economic security
Fujitsu plans to develop a semiconductor specifically for AI processing in servers and related systems. Fabricated using an advanced 1.4-nanometer process, the chip will be entirely domestic, with development and production carried out in Japan.
Manufacturing will be commissioned to Rapidus, a Japanese semiconductor company seeking to mass-produce cutting-edge chips.
Japan's Ministry of Economy, Trade and Industry is expected to cover part of the development cost. The effort will be a first step in responding to global moves to build AI capabilities amid economic security concerns.
The project centers on the neural processing unit, or NPU, a chip specialized for AI processing. Fujitsu plans to integrate NPUs with its CPUs -- being developed for uses such as in Japan's flagship Fugaku NEXT supercomputer -- within a single package.
NPUs enable systems to generate answers and are characterized by high energy efficiency. NPUs are typically found in consumer products such as PCs and smartphones, but the company plans to use them in servers and related systems.
Fujitsu has applied for a program operated by Japan's New Energy and Industrial Technology Development Organization (NEDO), a governmental institution. The project's initial development costs are estimated at 58 billion yen ($363 million). If the project is adopted, NEDO will cover about two-thirds of the expense.
Rapidus plans to begin construction of its second factory in fiscal 2027. Fujitsu's decision to entrust production to Rapidus marks the chip maker's second major deal from a domestic customer; Canon has already decided to order image-processing semiconductors for digital cameras.
For AI processing, graphics processing units (GPUs), which are mainly produced by Nvidia, are the core device. GPUs excel at parallel processing of complex data and are indispensable for training large language models. But when limited to inference tasks, NPUs are more efficient at calculation processing....
....MUCH MORE
