The new favorite of AI chips, Cerebras, which recently submitted for IPO, has become popular throughout Silicon Valley. Its chip can achieve up to 20 times the inference speed of H100 in small model scenarios; And for ultra large scale models (such as 400B parameter level), the single user response speed of the Cerebras CS-3 system is about 2.4 times that of the B200 So how exactly did Cerebras do it? Will it become the Nvidia killer? We need to start from the essence of the evolution of computing power. The evolution of AI computing power is shifting from "computing power itself" to "communication and system architecture". On this evolutionary path, Cerebras Systems offers a completely different answer: not optimizing distribution, but eliminating it as much as possible. 1、 Two routes: eliminating communication vs optimizing communication The current AI computing power is essentially divided into two architectural philosophies: one is represented by NVIDIA's approach: Multi chip (GPU), high-speed interconnect (NVLink/CPO), scale out (horizontal scaling) The other path is the Cerebras pathway: achieving wafer scale on a single chip On chip network replaces cross node communication, scale up (vertical amplification) The core difference is: one is to solve "how to connect more chips", and the other is to solve "how to not need to connect". 2、 Why is this road only established now Wafer scale is not a new concept. It was attempted in the 1980s, but commercialization failed in the 1990s. as a result of: The yield cannot be sustained No fault tolerance mechanism The software cannot support it The industry has thus formed a consensus: small die+high yield+distributed. The breakthrough of Cerebras lies in the simultaneous establishment of three things: 1) Engineering of fault-tolerant mechanism 2) Mature on-chip network 3) AI workload matching (high parallelism, strong synchronization, communication dominant) The fundamental change is from "perfect hardware" to "fault-tolerant systems". 3、 Performance Comparison: Single Point Limit vs System Expansion At the communication level, the advantages and disadvantages of the two routes are very clear: 1) On-chip communication Cerebras: pure on-chip → lowest latency, lowest energy consumption CPO: There is still photoelectric conversion → Single point efficiency: Cerebras is superior 2) System Expansion Cerebras: Once cross chip → Return to communication issues CPO: Sustainable Bandwidth Expansion → System capability: CPO is superior 3) Power consumption structure Cerebras: Single machine power consumption is extremely high, but communication is extremely efficient GPU+CPO: controllable single point power consumption, more balanced system efficiency The conclusion is clear: Cerebras wins the 'single player limit', CPO wins' system scale '. 4、 Applicable scenarios: Who should use Cerebras The criteria for judgment can be simplified into three questions: 1) Is communication a bottleneck 2) Can the task be centralized 3) Is the structure regular Therefore, it is highly suitable for training large models (dense models), ultra long contexts, and some HPC (PDE, fluids, etc.) The commonality of these tasks is strong coupling, high synchronization, and high bandwidth Partially applicable for large model inference (low concurrency), graph computation (advantage decreases when the structure is complex) Not suitable for CPUs (general-purpose computing), high concurrency inference, mobile/edge chips, real-time systems The commonalities of these systems: irregularity/high concurrency/low latency 5、 Will it become mainstream Although Cerebras is extremely strong in specific scenarios, the mainstream will not take this path because: 1) Physical constraint: power density; Signal delay → Fault tolerance cannot solve these problems 2) Economy: Small die has a higher yield rate; Chiplet is more flexible 3) Industry path: TSMC and other system optimization directions are modular, multi customer reuse rather than super large individual units 4) Change in demand side: Reasoning accounts for a much higher proportion than training, and multitasking and high concurrency have become mainstream 6、 The Significance of Cerebras Rather than saying that wafer scale size is an important trend, it is more accurate to say that fault-tolerant design is a philosophy that will be widely absorbed In the future, there may be chiplet level fault tolerance and package level detours The core change is that a single hardware no longer needs to be perfect, and the system is responsible for providing security. Returning to the original question: Will Cerebras become NVIDIA's "killer"? The answer is actually quite clear. It did hit the weak spot of the GPU system at a critical point - communication. But the choice of the industry is not either or, but the simultaneous adoption of multiple technological breakthroughs: stronger interconnection, lower communication energy consumption, and higher system level efficiency. Therefore, a more accurate judgment is that Cerebras is not Nvidia's killer, but the best practice that Nvidia and all chip companies can learn from. Disclaimer: I hold the subject matter mentioned in the article, and my views are inevitably biased and not investment advice. The investment risk is enormous, and I need to be extremely cautious when entering the market (Image: A Cerebas chip)