Cerebras Systems is an AI chip manufacturer known for creating the world's largest chip, the Wafer-Scale Engine (WSE). This chip is designed to address the inefficiencies of traditional GPUs in AI training by integrating memory directly within the processor, reducing latency and memory bottlenecks. This week, Contrary Research has put together a 12K word deep dive into Cerebras’ business, and how it represents a pulse on the broader semiconductor market!
The company's flagship product, the WSE, integrates memory directly within the GPU, addressing the memory bottleneck problem that traditional GPUs face. As a company, Cerebras doesn’t sell individual WSE chips but offers integrated computing systems that can be connected to form large clusters, providing near-linear performance scaling and significantly reducing the complexity of distributed training. For example, a single Cerebras system can deliver the computing performance of an entire room of servers with tens to hundreds of GPUs!
Cerebras has raised over $720 million in total funding from firms like Sequoia, Benchmark, and Coatue and, in August 2024, the company kicked off the process to work towards going public. With Nvidia skyrocketing in the public markets, an AI revolution everywhere you look, and Taiwan (and TSMC) increasingly in the geopolitical crosshairs, there is a huge swath of speculation about the broader semiconductor market. Cerebras sheds light on all of it.
One key consideration for Cerebras is its unique approach to AI computing. The company’s chips are designed specifically for AI workloads and offers superior performance and installation simplicity compared to traditional GPU-based systems. Cerebras' technology also allows for faster AI model training and inference, servicing the growing need for more efficient and powerful AI computing solutions.
The company’s success highlights the industry's shift towards specialized AI hardware, moving beyond traditional GPUs. In particular, Cerebras’ expansion into AI inference reflects the growing importance of efficient inference solutions in the AI ecosystem. And Cerebras’ focus on efficiency, speed, and simplified deployment aligns with the industry's key focus areas.
The shift from training to inference is one key example of how Cerebras as a company is like a pulse on the broader AI hardware market. By proving its capability to optimize models for specific hardware, Cerebras could become a crucial player in connecting AI training and inference processes, potentially improving deployment efficiency across the industry.
But any leading capability in AI is a moving target. In August 2024, Cerebras enabled inference capabilities on its latest systems, becoming the world’s fastest AI inference provider. Just a month later, in September 2024, both Groq and SambaNova made strides in faster and faster inference before Cerebras reclaimed the title. Suffice it to say, it's a rapidly evolving race.
And any breakthroughs in AI run up against the natural limits of the technology’s existing boundaries. On the technological front, Cerebras is grappling with both current limitations and future uncertainties. The company has reached the maximum possible chip size given ASML's EUV equipment constraints, potentially slowing future innovation to incremental improvements. This limitation is compounded by manufacturing challenges inherent in producing such large chips, increasing the probability of defects and adding complexity to the production process.
With the growth of AI model size and computational requirements straining data center power capacity, even leading to Microsoft pushing towards reopening Three Mile Island’s nuclear plant, these types of limitations are creating an infrastructure bottleneck that could slow the adoption of technology across the board. Whether Nvidia will see its products unseated by the likes of Cerebras and other emerging players, and how the technological infrastructure will shape the future of AI hardware, one thing is for certain:
Cerebras is like a mirror, reflecting each changing tailwind and potential pitfall as people rush to build the necessary hardware to take advantage of the promise of AI.
You can read the full memo on Cerebras here!
