In a close collaboration with FuriosaAI and several companies and research institutes, we are building AI accelerators and their compilers. We are planning to produce a chip using TSMC 7nm process.
AI accelerators are specialized hardware purpose-built to achieve high performance by exploiting the intrinsic parallelism of the workloads. AI accelerators are expected to revolutionize the IT industry just as GPUs did a while ago.
To better serve the target workloads, accelerators break the common abstraction layers established for general-purpose hardware like CPU. So it is difficult to design compilers for those accelerators that achieve programmability and efficiency at the same time. We are focused on developing such a compiler.