Why Cerebras AI chips stand out in the Nvidia-dominated market
AI chip company Cerebras (CBRS) went public Thursday, the first of what is expected to be a handful of major AI-based initial public offerings in 2026.
Cerebras is going up against the biggest players in the AI chip industry, including Nvidia (NVDA) and AMD (AMD), as it works to steal away market share in the AI training and inference space.
But Cerebras doesn’t produce the same types of chips as its larger rivals. In fact, it doesn’t make chips like any other company.
The biggest difference between Cerebras’ offerings is their size. When you think of a computer processor, you likely think of something about the size of a postage stamp. Some are slightly larger, some a tad smaller, but that’s about the ballpark.
Cerebras went in a completely different direction, building what it says is the largest commercial chip ever produced. Go grab your iPad. It’s about that big.
Called the WSE, or wafer-sized engine, Cerebras builds individual chips on whole semiconductor wafers. GPUs from Nvidia and AMD, and CPUs from Intel, use individual chips cut from larger wafers.
Think of it like a pizza. Traditional chipmakers take a small slice at a time for their chips. If they want to build larger processors, they can tie two slices together. Cerebras, however, uses the entire pie.
Why does that matter? Because the larger the chip, the more processing power and memory you can cram into it, allowing for faster data speeds versus a multi-chip setup. The downside is that building such a massive semiconductor is incredibly complex and expensive.
University of Illinois at Urbana-Champaign professor Deming Chen, explained:
If something goes wrong on a wafer-scale chip, you can’t just throw away a small chip — you lose the whole thing,
“It’s also harder to manufacture and less flexible. Smaller chips are easier to produce, cheaper, and can be scaled up by just adding more,” he added.
When companies produce chip wafers, some individual chip segments inevitably have flaws. Those flawed chips are either tossed out or sold as lower-end processors. It’s why you buy high-end, midrange, and entry-level Intel and AMD CPUs when you shop for a computer.
Cerebras claims it has developed a fault-tolerant architecture that can route around flaws in its wafers, letting the entire wafer act as a single processor.
Cerebras also uses what’s called SRAM, or static random-access memory. Traditional chips use what’s called DRAM, or dynamic random-access memory. Without getting too into the weeds, the major differences between the two are that SRAM is far faster than DRAM, but much more complex. That complexity means that the SRAM is also larger and more expensive than DRAM.
By using such an enormous wafer, though, Cerebras chips have the space to take advantage of SRAM. What’s more, the fact that Cerebras’ processors are a single product packed together means that data has to travel shorter distances than, say, Nvidia or AMD’s systems, which pair multiple chips together that have to move data around more.
The result, Cerebras says, is a processor that provides AI responses up to 15 times faster than “leading GPU-based solutions.”
But that doesn’t necessarily mean Cerebras’ offerings are the right product for every scenario, or that these kinds of single-wafer processors will make smaller chips obsolete.
Chen explained,
Smaller chips are still more practical for most use cases because they’re cheaper, more flexible, and easier to scale across different systems,
“Cerebras shines in certain workloads, but it won’t replace everything.”
In other words, don’t expect Nvidia, AMD, or their ilk to go anywhere anytime soon.
READ the latest news shaping the AI Chips market at AI Chips News
Why Cerebras AI chips stand out in the Nvidia-dominated market, source
Add comment