Google is deepening its partnership with MediaTek to develop a specialized AI chip, codenamed 'Triggerfish', designed for the next wave of artificial intelligence.
This development is driven by the shift towards what's called 'agentic AI'. Unlike older AI that might just answer a question, agentic AI performs complex, multi-step tasks, like planning a trip and booking flights. This creates new performance challenges. Specifically, it runs into the 'CPU wall', where the main computer brain is too busy coordinating tasks, and the 'memory wall', where fetching data from slower memory creates significant delays. Triggerfish is engineered to solve these exact problems.
Its solution is twofold. First, it includes 2-3 times more on-chip SRAM, which is an extremely fast type of memory. This allows the AI agent’s “working memory” to stay directly on the chip, avoiding slow data transfers. Second, it features a dedicated 'simulation die'. Think of this as a mini-controller on the chip itself that can manage tasks and run simulations locally, without constantly needing instructions from the host CPU. Together, these features dramatically reduce latency and increase efficiency.
This move also highlights Google's broader chip strategy. For years, Google has been designing its own AI chips, known as TPUs. Now, it's clearly pursuing a dual-sourcing approach. It maintains a long-term partnership with Broadcom for its powerful, general-purpose chips used for training massive AI models. Simultaneously, it's partnering with MediaTek to create more specialized and cost-effective chips like Triggerfish, which are optimized for inference—the stage where a trained AI is put to use.
This isn't a sudden change but a logical next step. Evidence has been building for months. Google had already split its previous TPU v8 into separate training and inference versions. Industry reports repeatedly pointed to MediaTek winning more of Google’s business. And Google itself has publicly stated its focus is shifting to the 'agentic era'. Triggerfish connects all these dots, representing a hardware evolution designed to meet a new software reality.
Ultimately, Triggerfish is more than just a new piece of silicon. It's a strategic bet on a future dominated by autonomous AI agents and a clear signal that the industry is moving from raw computing power towards specialized, efficient hardware tailored for specific AI workloads.
- Glossary -
- Agentic AI: AI systems that can perform complex, multi-step tasks autonomously by using tools and reasoning.
- SRAM (Static Random-Access Memory): A very fast, but more expensive, type of memory located directly on a computer chip, used for high-speed access to frequently used data.
- Inference: The process of using a trained AI model to make predictions or decisions based on new data. It's the 'use' phase after the initial 'training' phase.