Cerebras vs Nvidia. © Cerebras Systems.
Wafer Scale Engine (WSE), the chip developed by California startup Cerebras and introduced at Stanford University’s Hot Chips conference, is a mastodon. In fact, it’s the largest ever built, with a surface of 46.22cm2, 400,000 cores and a whopping 18GB of SRAM. This behemoth is intended for AI applications, more specifically for deep learning. One of the production challenges to overcome was the need for perfection, when the smallest imperfection in the silicon wafer can cause the failure of the entire chip. And since the larger the surface, the more likely the error… It is just about impossible to print billions of transistors error-free. When producing more than a hundred processors on a wafer, a few defective chips don’t matter. But when you produce a single chip over an entire wafer, every chip is crucial. Cerebras overcame the problem through redundancy, by adding extra cores over the entire chip to act as back-ups for neighbouring cores. Cerebras, a little-known company with 173 engineers and 112 million dollars in venture funding, discreetly developed the processor over the last few years. According to Fortune, the first commercially-produced units will be available in September, while some prototypes have already been delivered.
⇨ IEEE Spectrum, “6 things to know about the biggest chip ever built.”