Semi-Analysis, a semiconductor research firm, disclosed on October 16th at 8:30 AM Beijing time that Google has submitted its first test results for the TPUv7 (Ironwood) to its AgentX agent inference benchmark. Under the relevant test conditions, the TPUv7's unit-cost inference performance is 1.42 times that of NVIDIA's GB300 NVL72 FP4 and 2.5 times that of AMD's MI355X. SemiAnalysis stated that even using cost metrics geared towards Google's external customers, the TPUv7 maintains its leading position. This test used FP8 precision, focusing on measuring the throughput and cost performance of AI agents in multi-turn, long-context inference tasks. The results demonstrate that Google is driving its self-developed AI chips into a wider external inference market by opening up its TPU software ecosystem.
