Google announced the TPUv8 series of AI chips on April 21. The company plans to launch the new hardware globally this week. Google intends to fully replace the previous TPUv7 Ironwood series with these new accelerators. The TPUv7 Ironwood series launched in 2025.
The TPUv8 series includes two distinct models. The TPUv8t codename Sunfish serves as a high-performance training accelerator. Broadcom designed this training-focused chip. The TPUv8i codename Zebrafish functions as a cost-effective inference accelerator. MediaTek developed this inference-focused chip. Both models feature tight integration with Google's Axion Arm CPU. The Axion CPU utilizes the Neoverse N3 architecture.
Google did not disclose specific pricing for the TPUv8 series. The chips will become available globally starting this week. The announcement confirms the immediate rollout of the new hardware.
The new TPUv8 chips replace the TPUv7 Ironwood series. The TPUv8t model targets high-performance training workloads. The TPUv8i model emphasizes cost-effective inference tasks. The hardware integrates closely with Google's Axion Arm CPU based on the Neoverse N3 architecture. This integration aims to support Google's AI infrastructure needs.
Information about the TPUv8 series comes from reports citing a blog post. One report noted that Marvell does not appear on the design list. This detail relies on a blog post rather than direct confirmation from Google. The source of the information remains unconfirmed by the vendor.



Discussion
0 comments
Log in to join the thread with a thoughtful take, question, or correction.