Skip to main content

Google TPU v4

Overview

Google TPU v4 was announced at Google I/O 2021-05, as the flagship training chip succeeding the TPU v3. It uses a 2D torus topology for connectivity, with a single Pod containing 4,096 chips and delivering 1 ExaFLOPS BF16 per Pod.

In April 2023, Google published a TPU v4 performance paper claiming 1.2-1.7× the BF16 throughput of NVIDIA A100 and 1.3-1.8× better energy efficiency (sparking a debate with NVIDIA).

Core Specifications

ItemSpec
ArchitectureTPU v4 (4th generation)
ProcessTSMC 7nm
Matrix Multiply Unit (MXU)4× 128×128 (per cycle)
HBM32 GB HBM2
HBM Bandwidth1.2 TB/s
BF16 Compute (dense)275 TFLOPS
INT8 Compute275 TOPS
SparseCore2nd generation (embedding acceleration)
TDP~170 W
Form Factor4-chip board (v4 board)
Interconnect2D Torus, ICI 800 GB/s
Pod Scale4,096 chips
Pod BF16 Compute1.1 ExaFLOPS

Pod Architecture

  • 1 board = 4× TPU v4 chips
  • 1 Pod = 4,096 TPU v4 chips (8,192 boards = 32×32 2D torus)
  • 4×4 cube = 256 chips (mid-scale)
  • Per-chip ICI bandwidth = 800 GB/s (inter-chip interconnect)
  • PaLM 540B training used 2 Pods

Performance Comparison (BF16 Training)

MetricTPU v4 Single ChipA100 80GB SXMH100 SXM
BF16 Compute275 TFLOPS312 TFLOPS989 TFLOPS
Memory32GB HBM280GB HBM2e80GB HBM3
Bandwidth1.2 TB/s2 TB/s3.35 TB/s
InterconnectICI 2D TorusNVLink 600 GB/sNVLink 900 GB/s
Large Model TrainingAdvantageTieAdvantage

Google paper data: TPU v4 Pod 4,096 chips trains GPT-3 175B 1.7× faster than an equivalently scaled A100 Pod (with -1.3× power consumption).

Software Ecosystem

  • JAX (Google-recommended framework)
  • TensorFlow (native support)
  • PyTorch/XLA (official backend)
  • TPU VM (v4 dedicated runtime)
  • Pathways (heterogeneous TPU scheduling)

Use Cases

  • Ultra-large model training (PaLM, GPT-3 class)
  • ✅ Google Cloud TPU customers
  • ✅ Recommendation systems (DLRM)
  • ❌ On-premises data centers (Google Cloud access only)
  • ❌ Low-latency inference (use v5e)

Vendor Information

ItemDetail
VendorGoogle
AccessGoogle Cloud TPU v4 Pod
Price~$3.22/hr (chip)
Target MarketGoogle Cloud large model training