Cerebras Systems has unveiled WSE-3, the third generation of its wafer-scale AI megachip. The company says the new chip doubles the performance of its predecessor while consuming the same amount of power, a claim that puts scale at the center of its pitch for the next phase of AI training.
The chip is designed for CS-3 systems, which Cerebras says can train neural network models with up to 24 trillion parameters. That would put the system in range of models described as ten times larger than OpenAI's GPT-4 and Google's Gemini.
A larger chip built for larger AI models
WSE-3 continues Cerebras Systems' unusual approach to AI hardware: making a single chip that uses nearly an entire 300-millimeter silicon wafer. The square chip has an edge length of 21.5 centimeters, making it part of the company's line of wafer-scale processors rather than a conventional smaller accelerator.
The new chip contains 4 trillion transistors. According to the company, the move to the latest chip manufacturing technology delivers a more than 50 percent increase in transistor density compared with the earlier generation.
Cerebras also says WSE-3 is twice as powerful as its predecessor while using the same amount of power. In practical terms, that is the key claim: the company is presenting WSE-3 not only as a bigger chip, but as a way to increase AI compute without increasing power consumption at the same rate.
How WSE-3 compares with earlier Cerebras chips
The WSE line began with WSE-1 in 2019. Since then, the number of transistors has more than tripled, according to the source article.
The manufacturing process has also changed. WSE-3 will be built on TSMC's 5-nanometer technology, while the 2021 WSE-2 used the company's 7-nanometer technology.
That progression matters because Cerebras is using the newer manufacturing process to fit more computing structure into the same broad wafer-scale idea. The source frames WSE-3 as a continuation of the company's tradition of producing the world's largest single chip, rather than a break from that design philosophy.
CS-3 turns the chip into a training system
The chip itself is only one part of the announcement. Cerebras is also positioning the CS-3, the computer built around WSE-3, as a system for training very large language models.
Cerebras claims the CS-3 can train neural network models with up to 24 trillion parameters without the software tricks required by other computers. The source does not detail those tricks, but the distinction is central to the company's message: the hardware is meant to make very large model training more direct.
The company also says up to 2,048 systems can be combined. At that size, Cerebras says the configuration could train a language model such as Llama 70B in just one day.
The scale claims can be summarized in three layers:
- Chip level: WSE-3 has 4 trillion transistors and uses nearly an entire 300-millimeter silicon wafer.
- System level: CS-3 is designed to train models with up to 24 trillion parameters.
- Cluster level: Up to 2,048 systems can be combined for larger training runs.
Condor Galaxy 3 brings WSE-3 to Dallas
The first CS-3-based supercomputer will be Condor Galaxy 3 in Dallas. It will consist of 64 CS-3s and is expected to achieve 8 exaflops of performance.
Like its CS-2-based sister systems, Condor Galaxy 3 will be owned by Abu Dhabi's G42. The source does not provide additional ownership or deployment details, but the placement of Condor Galaxy 3 in the announcement gives WSE-3 an immediate system-level destination.
That is important because the WSE-3 story is not only about a single component. Cerebras is tying the chip directly to a named supercomputer, a specific system count, and a stated performance target.
The Qualcomm partnership focuses on inference
Cerebras has also entered into a partnership with Qualcomm. The stated goal is to reduce the price of AI inference by a factor of ten.
The plan described in the source has two steps. First, AI models would be trained on CS-3 systems. Then the team would make those models more efficient using methods such as pruning.
After that, the networks trained by Cerebras would run on Qualcomm's new inference chip, the AI 100 Ultra. This connects Cerebras' training hardware with Qualcomm's inference hardware, pairing large-scale model development with a separate path for running models more efficiently.
The announcement therefore covers both sides of the AI model lifecycle described in the source: training larger networks and reducing the cost of inference. WSE-3 is the training centerpiece, while the Qualcomm partnership is aimed at what happens after those models are trained.