Skip to main content

Google Cloud TPU v5p

Overview​

Google TPU v5p (v5 Premium) is Google's 5th-generation Premium TPU, launched in 2023, focused on large-scale LLM training (e.g. Gemini). It delivers 459 TFLOPS BF16 compute, 95GB HBM, and a single Pod scale of 8,960 chips. Interconnect is via 3D Torus + OCS (Optical Circuit Switch), available exclusively through Google Cloud.

Core Specifications​

ItemSpecification
ArchitectureGoogle TPU v5p (systolic array)
Process NodeTSMC 5nm
BF16 Compute (per chip)459 TFLOPS
INT8 Compute (per chip)459 TOPS
HBM Capacity95 GB
HBM Bandwidth2,575 GB/s
ICI Interconnect Bandwidth1,200 GB/s (bidirectional)
DCN Bandwidth50 Gbps
Pod Size8,960 chips (4×4×4×140 3D Torus)
CoolingLiquid cooling
AvailabilityGoogle Cloud only

TPU v5p vs v5e vs v4​

Metricv4v5ev5p
BF16 Compute275 TFLOPS197 TFLOPS459 TFLOPS
HBM Capacity32 GB16 GB95 GB
Bandwidth1,200 GB/s400 GB/s2,575 GB/s
Pod Size4,0962568,960
PositioningTrainingInferenceTraining

Vendor Information​

ItemDetails
ManufacturerGoogle LLC
Official Websitehttps://cloud.google.com/tpu
Product Pagehttps://cloud.google.com/tpu/docs/v5p
Exclusive toGoogle Cloud Platform

Software​

Key Features​

  • Systolic Array: Efficient matrix multiplication
  • 3D Torus interconnect + OCS optical switch
  • Pathways orchestration: Scalable across multiple Pods
  • JAX first-class citizen

Use Cases​

  • LLM training (Gemini, PaLM)
  • Recommendation systems
  • Multimodal models