Cerebras WSE-3: The Wafer-Scale Chip Revolutionizing AI Inference Speed

In an era where AI inference speed can determine business success or failure, Cerebras Systems is challenging the limits of AI processing with a fundamentally different approach from traditional GPU architectures. The company’s Wafer Scale Engine (WSE) uses an entire silicon wafer as a single chip—a design that defies semiconductor industry conventions and achieves unprecedented processing speeds.

Announced in March 2024, the WSE-3 features 4 trillion transistors and 900,000 AI-optimized cores, achieving processing speeds of 450 tokens per second for the Llama 3.1 70B model and 1,800 tokens per second for the 8B model. This represents over 20x the speed of comparable NVIDIA GPU solutions¹². TIME Magazine selected this technology as one of the Best Inventions of 2024³.

What is Cerebras: An AI Infrastructure Company Built by Serial Entrepreneurs

Cerebras Systems is an AI infrastructure company founded in 2015 by five co-founders, including Andrew Feldman. Feldman is known as the founder of SeaMicro, a serial entrepreneur who sold that company to AMD in 2012 for $334 million ($281 million in cash plus $53 million in stock)⁴.

Before SeaMicro, Feldman was involved in launching Force10 Networks (acquired by Dell for $800 million) and Riverstone Networks (acquired by YAGO for $280 million), with total exit values exceeding $1 billion. With an MBA from Stanford University, he is recognized as a pioneer in energy-efficient computing who forever changed the server industry trajectory by creating the microserver category⁵.

The Co-founding Team

Cerebras’ founding team consists of members who worked together at SeaMicro:

Andrew Feldman (CEO)
Gary Lauterbach
Michael James
Sean Lie
Jean-Philippe Fricker

The team has successfully raised over $720 million from top-tier venture capital firms including Benchmark, Foundation Capital, and Altimer Capital (as of November 2021)⁶.

WSE-3’s Technical Innovation: Why It’s So Fast

Fundamental Solution to the Memory Bottleneck

In traditional GPU architectures, memory is located in external HBM (High Bandwidth Memory) outside the compute cores. During LLM inference, parameters for each layer must be loaded from memory into compute cores, matrix multiplication performed, and this process repeated across all transformer blocks. During this time, most of the GPU’s compute capacity sits idle waiting for parameters to be fetched from HBM⁷.

Cerebras solved this problem fundamentally. The WSE-3 features 44GB of SRAM directly on silicon, approximately 1,000 times the capacity of an NVIDIA H100. This SRAM is distributed near compute cores, eliminating the need for external memory access during inference⁸.

Overwhelming Memory Bandwidth

WSE-3 technical specifications⁹:

Memory Bandwidth: 21 petabytes/second (7,000x the H100)
Processor Interconnect Bandwidth: 214 petabits/second (3,715x graphics processors)
Transistor Count: 4 trillion
AI Core Count: 900,000
Peak Performance: 125 petaflops
Chip Size: 46,225mm² (over 56x the maximum size of traditional chips at ~815mm²)

On-Wafer Interconnect Innovation

The WSE-3’s on-wafer interconnect eliminates communication delays and inefficiencies from connecting hundreds of small devices via wires and cables. With all communication and memory on a single silicon slice, data moves unimpeded, achieving core-to-core bandwidth of 1,000 petabits per second and SRAM-to-core bandwidth of 9 petabytes per second¹⁰.

“It’s not just a little more,” says Feldman. “It’s four orders of magnitude greater bandwidth, because we stay on silicon”¹¹.

Real-World Performance: The Numbers Tell the Story

Inference Speed Achievements

Performance at August 2024 announcement¹²:

Llama 3.1 8B: 1,800 tokens per second (2.4x faster than Groq)
Llama 3.1 70B: 450 tokens per second (the only platform enabling instant responses)

By November 2024, with further optimization¹³:

Llama 3.2 70B: Achieved 2,100 tokens per second

In May 2025, demonstrating performance exceeding NVIDIA Blackwell¹⁴:

Llama 4 Maverick (400B parameters): Over 2,500 tokens per second per user

Training Speed Improvements

WSE-3 serves as the foundation for the CS-3 computer system, compared to NVIDIA DGX H100¹⁵:

Training Speed: 8x faster
Maximum Model Size: Supports up to 24 trillion parameters
Llama 70B Training Time: 30 days on GPUs completed in 1 day on CS-3 cluster
Power Efficiency: One-third the power consumption of DGX solutions

Why Such a Dramatic Difference in LLM Inference?

The relationship between LLM characteristics and architecture creates Cerebras’ advantage. LLM inference is inherently sequential—generating each word requires passing through the entire model. One word takes one pass, 100 words require 100 passes, and since each word depends on the previous one, this process cannot be parallelized¹⁶.

Traditional GPU architectures require repeatedly loading model weights from external memory, creating a bottleneck. Cerebras’ wafer-scale approach fundamentally solves this memory bandwidth bottleneck by keeping the entire model on-chip.

From a developer’s perspective, the hardware appears as a gigantic GPU with all model weights on-chip. As a result, inference runs “unreasonably fast” at over 2,500 tokens per second¹⁷.

Key Partnerships and Deployments

Strategic Partnership with Meta

In April 2025, Meta announced a partnership with Cerebras to power the new Llama API with Cerebras technology, enabling developers to access inference speeds up to 18x faster than traditional GPU-based solutions¹⁸.

Ahmad Al-Dahle, VP of GenAI at Meta, states: “Cerebras’ WSE is one of the most innovative AI hardware architectures we’ve seen”¹⁹.

Other Major Customers

Mayo Clinic: Medical AI research applications
G42: AI application deployment in the Middle East
AlphaSense: Financial analysis platform adoption
Perplexity: Search engine inference acceleration

Denis Yarats, CTO of Perplexity, comments: “Cerebras’ speed is phenomenal. Being able to execute complex reasoning tasks instantly enables entirely new user experiences”²⁰.

Cerebras’ Vision: The Transformation Brought by Instant AI Inference

The Cerebras Scaling Law

Cerebras advocates its own “scaling law”—the concept that improved inference speed doesn’t just reduce response times but enhances AI “intelligence” itself²¹.

What high-speed inference enables:

Complex Workflows: Real-time execution of multi-model processing
Interactive Experiences: Continuous user dialogue without waiting
Large-Scale Reasoning Chains: Deep expansion of Chain of Thought

Three Deployment Options

Cerebras offers flexible deployment methods²²:

Cloud Service: Instantly available inference API
Dedicated Systems: Enterprise-exclusive instances
On-Premises: For organizations requiring complete control and security

Future Developments

Cerebras is reportedly considering an IPO in 2025, planning to further strengthen its position in the AI infrastructure market. The company’s technology has received high industry recognition, including selection for Forbes AI 50 (April 2024) and TIME’s 100 Most Influential Companies (May 2024)²³.

For those interested in deeper understanding of Cerebras’ technological innovation, several resources are available.

The Cerebras official site provides the latest technical specifications and case studies. The architecture deep dive explains technical details of hardware-software co-design. IEEE Spectrum’s technical article offers third-party technical evaluation.