Google tapped MediaTek to design TPUv9 'Triggerfish' chip, fusing CPU and compute die into single package for agentic AI applications.
Google is working with MediaTek to develop its next-generation TPUv9 chip, codenamed Triggerfish, targeting the Agentic AI market. This partnership builds on Google's recent TPU v8 family, which included the inference-optimized v8i "Zebrafish" developed by MediaTek and the training-optimized v8t "Sunfish" developed by Broadcom.
The TPUv9 Triggerfish represents a significant shift in Google's strategy. Rather than maintaining separate chips for inference and training as it did with the v8 series, the next generation will combine both capabilities on a single processor. The chip will feature a larger SRAM cache—two to three times higher than previous generations—alongside a new CPU tile developed by MediaTek to handle seamless workload switching between training and inference modes. This CPU tile will be packaged together with the main compute die on the same carrier.
According to FundaAI's latest report, TPUv9 production will commence with Humufish in Q3 2027, followed by Triggerfish in Q4 2027. Both chips are expected to reach volume production during 2028, with total TPU shipments for 2027 projected at approximately 10 to 11 million units.
Google's TPUv9 lineup includes an optimized variant. Humufish (formerly designated v8e) will utilize Intel's EMIB packaging technology, which offers greater versatility and more scalable, lower-cost designs compared to CoWoS 2.5D approaches. Google will manufacture Humufish's main compute die in-house, while MediaTek will handle the I/O and back-end design. Triggerfish will use HBM4E memory, while Humufish will employ HBM4 DRAM.
Manufacturing will take place at TSMC, with Intel handling packaging responsibilities. There is potential for Google to shift some capacity to Intel Foundry Services to alleviate ongoing supply-chain constraints and reduce the burden on TSMC, which faces significant demand from NVIDIA, AMD, and Apple.
By integrating both training and inference capabilities onto a single die with substantially expanded SRAM and a dedicated CPU tile for workload switching, Triggerfish is designed to deliver a versatile, high-performance solution optimized for the computational demands of Agentic AI and reinforcement learning applications.