Google TPU v5p

TPU v5 Architecture

Active

Launched August 2023

Core Specifications

VendorGoogle

ArchitectureTPU v5

Form FactorMezzanine

VRAM—

Memory Bandwidth—

TDP400 W

Precision	TFLOPs
BF16	459

Configuration	Precision	Performance	Source
Gemini Pro 1.5, batch inference	—	8,500 tokens_per_second	View

Configuration	Precision	Performance	Source
Model FLOPs Utilization for large language models	—	65 mfu_percent	View
PaLM 540B, optimized for JAX/TensorFlow	—	14,500 tokens_per_second	View

Provider	Instance Type	Price per Hour	Region	As of
Google Cloud	—	$5.00/hr	us-central	Dec 2025

Peak Performance

459

TFLOPs (BF16)

Efficiency

1.15

TFLOPs per Watt

View other Google GPUs or compare across vendors