Former PlayStation Chip Engineers Take on Nvidia

The firm, LENZO, is creating a next-generation AI chip often called CGLA, quick for Coarse-Grained Logic Architecture. The key promoting level is energy effectivity. Compared with Nvidia’s GPUs, the chip is designed to run AI workloads utilizing as much as 90 p.c much less electrical energy, a declare that might have far-reaching implications for information facilities and AI infrastructure worldwide.

LENZO’s core crew contains engineers who beforehand labored on the PlayStation 2 and PlayStation 3 processors, in addition to specialists who helped develop supercomputer chips throughout their time at Fujitsu. The firm goals to convey its first industrial chip to market within the spring, with manufacturing dealt with by Taiwan Semiconductor Manufacturing Co. The completed chip is predicted to measure roughly 5 millimeters sq..

At a time when Nvidia dominates roughly 90 p.c of the worldwide AI chip market and boasts a market capitalization exceeding 600 trillion yen, LENZO’s problem could seem audacious. Yet its founders argue that Nvidia’s dominance is constructed on an structure that’s approaching its bodily limits.

At the guts of the problem is energy consumption. Conventional CPUs and GPUs are primarily based on what is called the von Neumann structure, during which reminiscence and computation items are separated. This construction requires fixed information motion between reminiscence and processors, consuming huge quantities of power within the course of. In truth, research present that transferring information only one millimeter inside a chip can devour extra energy than performing an arithmetic operation itself.

While GPUs improved on this by processing information in giant batches, they nonetheless undergo from heavy power loss brought on by frequent reminiscence entry. Today, greater than half of the electrical energy utilized by AI servers is consumed not by computation, however by information switch between reminiscence and processing items.

Google’s Tensor Processing Unit, or TPU, addressed this problem by adopting a dataflow structure optimized for matrix calculations utilized in AI. By streaming information in a hard and fast sequence, TPUs cut back reminiscence entry and enhance effectivity. However, they’re designed nearly completely for matrix-based AI workloads, limiting their flexibility.

LENZO’s CGLA takes a special strategy. Rather than fixing the info movement prematurely, it permits the movement of information between processing parts to be reconfigured freely. This allows the chip to deal with not solely present AI fashions akin to transformers, but additionally future algorithms which will depend on completely totally different computational constructions.

According to the corporate, this flexibility permits CGLA to mix excessive energy effectivity with broad applicability, one thing neither GPUs nor TPUs can totally obtain. While GPUs supply versatility at the price of energy effectivity, and TPUs supply effectivity at the price of flexibility, CGLA is designed to ship each.

Another benefit lies in value. Modern AI chips rely closely on high-bandwidth reminiscence, which has grow to be more and more costly. By lowering the necessity for fixed information motion, CGLA can function with much less reminiscence, decreasing each power use and manufacturing prices.

Yet maybe the most important impediment going through any new AI chip shouldn’t be {hardware}, however software program. Nvidia’s CUDA platform has grow to be the de facto normal for AI growth, deeply embedded in analysis and industrial functions alike. Many builders write their code particularly for CUDA, making it tough for various {hardware} to achieve traction.

LENZO acknowledges this problem however sees alternative in shifting business traits. As cloud suppliers and AI builders search options to Nvidia’s ecosystem, curiosity in non-GPU options is rising. The firm believes that demand for energy-efficient and versatile chips will enhance as AI workloads develop and energy prices rise.

The firm additionally sees long-term worth in adaptability. Today’s AI methods rely closely on transformer fashions, however new approaches are already rising. If the dominant algorithms change, {hardware} designed for a single technique may shortly grow to be out of date. CGLA, in contrast, is designed to adapt by way of software program somewhat than {hardware} redesign.

In this sense, LENZO shouldn’t be merely attempting to construct a sooner chip, however to redefine how AI processors are structured. Whether the corporate can overcome Nvidia’s entrenched ecosystem stays to be seen, however its expertise highlights a rising recognition that the way forward for AI will rely not solely on efficiency, however on effectivity, flexibility, and sustainability.

Source: テレ東BIZ