A few months ago, Google revealed its latest family of TPUs, which included the v8i “Zebrafish” (optimized for inference), developed by MediaTek, and the v8t “Sunfish” (optimized for training), developed by Broadcom.
For your next generation TPU, the giant intends to implement both inference and training capabilities on a single chip. And the internal codename would be “Triggerfish”, being part of the TPUv9 enhanced series.
Now, it appears that MediaTek will redevelop the chip for Google. Mass production of the TPUv9 series is expected to begin in the third quarter of 2027, with the “Humufish” being the first, followed by the “Triggerfish” in the fourth quarter of 2027.
Both chips will reach volume production scale in 2028.
Related News:
- Palworld creator explains why it doesn’t use AI: “players don’t want that”
- SanDisk wants to unite storage and processing on the same chip to accelerate the AI era
- Arm already dominates almost half of data center revenue and threatens the reign of x86 architecture
Single-chip inference and training
Key highlights of TPUv9 “Triggerfish” include a larger SRAM and a new CPU block (tile) added by MediaTek to manage the switching of Artificial Intelligence workloads between training and inference.
This new CPU block will be present in the same package as the die core computing. More impressively, total TPU shipments in 2027 are estimated at between 10 and 11 million units.
Google’s strategy with its TPUs has evolved substantially in recent years. Since the company entered the custom chip market, it has rapidly expanded its lineup of chips with specific accelerators for certain workloads.
With the next generation, the company seeks to create an integrated solution capable of handling inference and training workflows, while simultaneously offering a larger SRAM cache memory (2 to 3 times higher) to meet the demands of Agentic AI.
Blue participation
Google is also expected to introduce a new optimized line within the TPUv9 family, with models such as “Humufish” using Intel’s EMIB encapsulation technology. Compared to CoWoS (2.5D), EMIB is said to be much more versatile, offering scalable and low-cost designs.
The TPUv9 “Humufish” chip (formerly v8e) is expected to have its main computing die manufactured by Google itself, while the I/O and back-end design will be handled by MediaTek.
The v9 “Triggerfish” chip will use HBM4E memory, while the v9 “Humufish” will use HBM4 DRAM.
Changes?
Reports have suggested that Google’s next TPU chips could be manufactured by Intel, with volumes estimated at up to 3 million units. However, market analysts state that manufacturing will continue to be carried out by TSMC, with Intel responsible for the packaging stage.
On the other hand, considering the persistent bottlenecks in the supply chain and TSMC’s overload with orders for chips from NVIDIA, AMD and Apple, there is a possibility that Google will transfer part of the production to Intel Foundry.
With TPUv9 “Triggerfish”, Google aims to transform the hardware landscape for AI in partnership with MediaTek. By integrating training and inference capabilities on a single chip, it offers a versatile, high-performance solution designed to meet the demands of Agentic AI.
Sources: @jukan05 and @mingchikuo.
Almost religious stance
Steam Machine will not be subsidized “for the good of the consumer”
Source: www.adrenaline.com.br
Source link
