TOKYO - A new chapter is emerging in the race for AI semiconductors, as a Japanese startup founded by former PlayStation chip engineers sets its sights on challenging industry giant Nvidia with a radically different approach to processor design.
The company, LENZO, is developing a next-generation AI chip known as CGLA, short for Coarse-Grained Logic Architecture. The key selling point is power efficiency. Compared with Nvidia’s GPUs, the chip is designed to run AI workloads using up to 90 percent less electricity, a claim that could have far-reaching implications for data centers and AI infrastructure worldwide.
LENZO’s core team includes engineers who previously worked on the PlayStation 2 and PlayStation 3 processors, as well as specialists who helped develop supercomputer chips during their time at Fujitsu. The company aims to bring its first commercial chip to market in the spring, with manufacturing handled by Taiwan Semiconductor Manufacturing Co. The finished chip is expected to measure roughly five millimeters square.
At a time when Nvidia dominates roughly 90 percent of the global AI chip market and boasts a market capitalization exceeding 600 trillion yen, LENZO’s challenge may seem audacious. Yet its founders argue that Nvidia’s dominance is built on an architecture that is approaching its physical limits.
At the heart of the issue is power consumption. Conventional CPUs and GPUs are based on what is known as the von Neumann architecture, in which memory and computation units are separated. This structure requires constant data movement between memory and processors, consuming vast amounts of energy in the process. In fact, studies show that moving data just one millimeter inside a chip can consume more power than performing an arithmetic operation itself.
While GPUs improved on this by processing data in large batches, they still suffer from heavy energy loss caused by frequent memory access. Today, more than half of the electricity used by AI servers is consumed not by computation, but by data transfer between memory and processing units.
Google’s Tensor Processing Unit, or TPU, addressed this issue by adopting a dataflow architecture optimized for matrix calculations used in AI. By streaming data in a fixed sequence, TPUs reduce memory access and improve efficiency. However, they are designed almost exclusively for matrix-based AI workloads, limiting their flexibility.
LENZO’s CGLA takes a different approach. Rather than fixing the data flow in advance, it allows the flow of data between processing elements to be reconfigured freely. This enables the chip to handle not only current AI models such as transformers, but also future algorithms that may rely on entirely different computational structures.
According to the company, this flexibility allows CGLA to combine high power efficiency with broad applicability, something neither GPUs nor TPUs can fully achieve. While GPUs offer versatility at the cost of power efficiency, and TPUs offer efficiency at the cost of flexibility, CGLA is designed to deliver both.
Another advantage lies in cost. Modern AI chips rely heavily on high-bandwidth memory, which has become increasingly expensive. By reducing the need for constant data movement, CGLA can operate with less memory, lowering both energy use and production costs.
Yet perhaps the biggest obstacle facing any new AI chip is not hardware, but software. Nvidia’s CUDA platform has become the de facto standard for AI development, deeply embedded in research and commercial applications alike. Many developers write their code specifically for CUDA, making it difficult for alternative hardware to gain traction.
LENZO acknowledges this challenge but sees opportunity in shifting industry trends. As cloud providers and AI developers seek alternatives to Nvidia’s ecosystem, interest in non-GPU solutions is growing. The company believes that demand for energy-efficient and flexible chips will increase as AI workloads expand and power costs rise.
The company also sees long-term value in adaptability. Today’s AI systems rely heavily on transformer models, but new approaches are already emerging. If the dominant algorithms change, hardware designed for a single method could quickly become obsolete. CGLA, by contrast, is designed to adapt through software rather than hardware redesign.
In this sense, LENZO is not simply trying to build a faster chip, but to redefine how AI processors are structured. Whether the company can overcome Nvidia’s entrenched ecosystem remains to be seen, but its technology highlights a growing recognition that the future of AI will depend not only on performance, but on efficiency, flexibility, and sustainability.
Source: テレ東BIZ
