Cerebras unveils CS-4, an AI accelerator with double the compute performance on the same chip

Cerebras has introduced the CS-4, a new AI accelerator that the company says delivers twice the compute performance of the previous generation without any increase in physical chip size. CEO Andrew Feldman positions the system as the fastest in the industry.

Cerebras is known for its Wafer Scale Engine, an approach in which the entire surface of a silicon wafer is used as a single large chip. This sets the company apart from vendors that combine multiple smaller chips, such as Nvidia with its H100 and H200 systems. The CS-4 builds on that same architecture but extracts more performance from the same chip area through improvements in design and manufacturing.

The source material contains limited technical details; the specifications below are based on the available announcement. Further technical documentation from Cerebras itself was not fully public at the time of writing.

What the CS-4 does differently from its predecessor

The core of the announcement is that Cerebras has managed to double compute capacity without increasing chip area. This is technically significant because larger chips are harder and more expensive to manufacture. By deriving performance gains from efficiency rather than scaling up, Cerebras keeps the production basis for the CS-4 comparable to that of the CS-3.

In its Wafer Scale Engine approach, Cerebras uses a chip roughly the size of a full silicon wafer, which is a multiple of the surface area of conventional GPUs. This gives the system a large internal memory and substantial compute power at short mutual distances, which is particularly advantageous for running large language models where data must constantly move back and forth between compute cores and memory.

Exactly how the performance gain was achieved, through new memory architecture, improved interconnects, or process improvements at the chip manufacturer, could not be verified based on the available source material. Cerebras has previously worked with TSMC for the production of its wafer-scale chips.

Position relative to the competition

CEO Andrew Feldman claims that the CS-4 is the fastest system in the industry. Such statements warrant context. The market for AI accelerators is dominated by Nvidia, but AMD, Intel, and a growing number of specialised chipmakers, including SambaNova, Groq, and Graphcore, also offer alternatives. Each vendor uses its own benchmarks and comparison criteria, making direct comparisons difficult without independent testing.

Cerebras has traditionally focused on inference speed for large language models, a segment where demand has grown sharply over the past two years driven by the rise of models such as GPT-4 and Meta's open models. The company offers its systems both as a cloud service and as hardware for data centres.

Whether the CS-4 is faster in practice than Nvidia's most recent systems depends heavily on the specific task, model, and configuration. Cerebras generally performs well in scenarios where the entire model fits on a single system and high per-user throughput is required.

Relevance for the European AI scene

For European and Dutch players in the AI infrastructure market, the CS-4 announcement is relevant for a broader reason. The market for AI accelerators is widening: alongside Nvidia, an increasing number of serious alternatives are emerging that are competitive or superior on specific applications. This expands the options available to data centre companies, cloud providers, and AI startups when composing their infrastructure.

For investors and policymakers looking at European AI sovereignty, this pattern is significant. As the market becomes more diversified, dependence on a single vendor decreases. Whether European parties will actually gain access to systems such as the CS-4, and on what terms, is a practical follow-up question that will be answered in the coming months as Cerebras elaborates on its distribution plans.