"It's the fastest inference in the world," said Naor Penso, the chief information security officer of Cerebras, during our conversation at Web Summit in Vancouver today. "Last week, NVIDIA announced achieving 1,000 tokens per second on Llama 4, which is quite impressive. We just released a benchmark showing 2,500 tokens per second today."

If this sounds confusing, think of "inference" as the ability to think or act: forming sentences, images, or videos based on your input, or prompt. Imagine "tokens" as fundamental units of thought—a word, character, or symbol.

The higher the number of tokens an AI system can handle per second, the quicker it can deliver results. And speed is crucial. While it might not be a big deal for individuals, businesses need instant responses when adding AI engines to applications like suggesting one more ingredient for a complete recipe of Korean-Style BBQ Beef Tacos in a grocery shopping cart.

Interestingly, speed is about to become even more vital.

We're moving into an era where AI agents can manage complex, multi-step tasks for us, such as planning and booking a weekend getaway to Austin for a Formula 1 race. These agents aren't magical—they tackle large tasks step by step, breaking them down into 40, 50, or even 100 smaller tasks. This means significantly more work.

"AI agents require far more operations, and these operations must communicate effectively," Penso explained to me. "You can't afford slow inference."

The WSE's four billion transistors play a key role in enabling this speed. For context, the Intel Core i9 has around 33.5 billion transistors, and the Apple M2 Max chip offers just 67 billion transistors. However, it's not just the sheer number of transistors that makes a difference; it's also their co-location—everything being on one chip, along with 44 gigabytes of the fastest RAM (memory) available.

"AI computation requires a lot of memory," Penso noted. "With NVIDIA, you often need to go off-chip, but with Cerebras, you don't."

Independent agency Artificial Analysis supports the speed claims, stating they've tested the chip on Llama 4 and reached 2,522 tokens per second, compared to NVIDIA Blackwell's 1,038 tokens per second.

"We've tested dozens of vendors, and Cerebras is the only inference solution that outperforms Blackwell for Meta's leading model," said Micah Hill-Smith, CEO of Artificial Analysis.

The WSE chip represents an intriguing advancement in computer chip design.

Although we've been creating integrated circuits since the 1950s and microprocessors since the 1960s, the CPU dominated computing for many years. More recently, the GPU transitioned from supporting graphics and gaming to becoming the central processor for AI development. The WSE isn't based on x86 or ARM architectures but something entirely new that enhances GPUs, according to Julie Shin, Cerebras' chief marketing officer.

"This is not an incremental technology," she emphasized. "This is another leap forward for chips."

The above is the detailed content of World's Largest Chip Sets AI Speed Record, Beating NVIDIA. For more information, please follow other related articles on the PHP Chinese website!