Artificial intelligence (AI) chipmaker Cerebras Systems (CBRS 8.39%) went public on May 14 amid much buzz. The stock popped by 68% on its first day of trading -- a sign, according to experts, that the AI trade is alive and well.
The hype around this initial public offering makes sense, given that Cerebras is competing with major chip designers like Nvidia (NVDA 1.18%) and AMD, both of which have performed extraordinarily well in recent years.
Investors were likely even more excited to read in Cerebras' registration statement that the company says its chips are 15 times faster than those of leading competitors.
While that's undoubtedly intriguing, Nvidia's graphics processing units (GPUs) still have two distinct advantages over Cerebras' chips.
Image source: Getty Images.
Why Cerebras' chips are different
While most chip designers start the process of making their chips by cutting large silicon wafers into many small pieces that will eventually become individual chips, Cerebras uses an entire silicon wafer to make one massive chip: Its wafer-scale engines (WSEs) are 58 times larger than Nvidia's Blackwell 200 processors.Some news outlets have described Cerebras' chips as being the size of a dinner plate.
NASDAQ: CBRS
Key Data Points
Larger chips certainly have some advantages in data centers because more memory and processing power can be packed onto them. Cerebras says its WSE-3 model has 250 times more on-chip memory than Nvidia's B200 model and 2,625 times more memory bandwidth.
Having more memory on the chip helps address data-processing bottlenecks that smaller GPUs face, such as fetching data from other GPUs or from external memory adjacent to the chip.
While that could make Cerebras' chips competitive or even better in certain situations, here are two reasons Nvidia shareholders are not panicking.
1. Smaller chips are advantageous in some ways
While Cerebras naturally focuses on the disadvantages of its rivals' chips, there are scenarios in which smaller chips can offer advantages, primarily because they provide customers with more flexibility.
For one, wafer-sized chips are more expensive to make, and they contain many more cores than smaller ones. If enough parts of one of Cerebras' massive chips are defective, the whole (expensive) chip may have to be tossed. Though the company designed its chips with redundancies to reduce the odds of that happening, defects do occur in all chip manufacturing processes.Â
Furthermore, not all AI tasks require the full power of a WSE-3. Smaller tasks require fewer GPUs, so small to mid-size companies looking to deploy AI applications are likely to find more cost-effective solutions from companies like Nvidia. And Nvidia GPUs can be deployed in clusters customized to deliver any processing power capacity required.
2. CUDA
But Nvidia's biggest competitive advantage may be its CUDA (Compute Unified Device Architecture) software layer, which currently dominates the space. It's a software platform that enables developers to write code for Nvidia's GPUs -- and generally, that code can only be run on Nvidia's GPUs.
The company introduced CUDA in 2006, and over the last two decades, a robust ecosystem has developed around the platform, which has been widely adopted by consumers and businesses.
For instance, according to Nvidia, pharmaceutical companies use CUDA to help discover new drugs, automakers use it to enhance their vehicles' autonomous driving capabilities, and retailers use it to analyze customer data that impacts their product recommendations and advertising.
Transitioning away from CUDA would be a difficult process. It would not only require rewriting a lot of code, but also operating around that huge existing ecosystem. Not to mention that most developers in the space have worked with CUDA their entire careers, and use tons of other tools developed specifically for CUDA. They are used to those tools, and are efficient with them. Giving all that up would be an unattractive option for most companies.
NASDAQ: NVDA
Key Data Points
As of last year, CUDA supported over 900 libraries and AI models, according to Business Insider, with each library corresponding to an industry that is leveraging Nvidia's technology.
It is true that Cerebras has built its own software layer to compete with CUDA, but Nvidia has a big head start.
Nvidia's competitive moat is comparable to the ones enjoyed by payments companies Visa and Mastercard, which play vital roles in facilitating payment transactions. Could another company do what those two do from a technological perspective? Sure, but could one replicate their enormous global networks and scale?
Well, it hasn't happened yet.
