Skip to main content

TPU (Tensor Processing Unit) Architecture

What is a TPU

TPU (Tensor Processing Unit) is Google's in-house dedicated AI accelerator, first deployed in 2015 (for AlphaGo) and opened to Google Cloud in 2018. Purpose-built for neural network matrix operations, its core is the MXU (Matrix Multiply Unit) — a 128×128 matrix multiplier.

The latest TPU v7 Ironwood (2025) features 192GB HBM per chip, designed for the LLM inference era.

TPU Generations

GenerationYearCompute (BF16)MemoryPod ScaleCharacteristics
TPU v1201523 TOPS INT88GB256Inference
TPU v22017180 TFLOPS16GB HBM256Training-inference fungible
TPU v32018420 TFLOPS32GB HBM1024Liquid cooling
TPU v42021275 TFLOPS32GB HBM40962D Torus
TPU v5e2023197 TOPS INT816GB HBM256Inference optimized
TPU v5p2023459 TFLOPS95GB HBM38960Training flagship
TPU v6e (Trillium)2024918 TFLOPS32GB HBM256Training-inference fungible
TPU v7 (Ironwood)20254,614 TOPS FP8192GB HBM9,216Inference-era flagship

Core Architecture: MXU

Matrix Multiply Unit (MXU)

  • 128×128 systolic array
  • Performs 16,384 multiply-adds per cycle (INT8)
  • BF16 / INT8 / FP8 support
  • Sparse acceleration: skip zero values

HBM Memory

  • High Bandwidth Memory (HBM2 / HBM2e / HBM3)
  • TPU v7 reaches 192GB (per chip)
  • 7,380 GB/s bandwidth

SparseCore

  • Second generation (TPU v4+)
  • Accelerates embedding lookups (recommendation systems)
  • Used for Google Search, YouTube

Interconnect Topology

2D Torus (TPU v4)

  • Single Pod 4,096 chips = 32×32 grid
  • ICI (Inter-Chip Interconnect) 800 GB/s
  • Data flows cyclically across the 2D grid

3D Torus (TPU v5p, v7)

  • Single Pod 8,960-9,216 chips
  • 3D cubic topology
  • 6-way interconnect (vs 2D 4-way)
  • Reduces long-distance communication latency

TPU vs GPU

DimensionTPUGPU (NVIDIA)
ComputeMXU 128×128 systolic arrayTensor Core matrix multiply
MemoryHBM (medium)HBM (larger)
Interconnect2D/3D Torus (massive Pods)NVLink (smaller Pods)
ProgrammingXLA compilationCUDA interpreted execution
EcosystemJAX / TensorFlowPyTorch / TF / JAX
DeploymentGoogle Cloud onlyDeploy anywhere
PricingHourly rentalOne-time / cloud
Best forUltra-large model training / inferenceGeneral AI

Software Stack

Programming Model

  • XLA (Accelerated Linear Algebra) compiler
  • JAX (Google-recommended framework)
  • TensorFlow (native support)
  • PyTorch/XLA (official backend)

High-Level APIs

  • Pathways (heterogeneous TPU scheduling)
  • MaxText (LLM training reference implementation)
  • TPU VM (v4+ dedicated runtime)

TPU Use Cases

  • Ultra-large model training (PaLM 540B used 2 Pods)
  • ✅ LLM inference (v7 Ironwood 192GB)
  • ✅ Recommendation systems (DLRM, embedding acceleration)
  • ✅ Google Cloud customers
  • ❌ Own data centers (Google Cloud access only)
  • ❌ Low-latency small models (GPU is more flexible)

Detailed Product Pages

Training / Training-Inference Fungible

Inference Optimized / Split Architecture

  • Google TPU v5e - 2023-Q2 inference Lite, 16GB HBM2 400 TF FP8 $1.20/hr
  • Google TPU 8t - 2026-04 first split-training, training-dedicated ASIC
  • Google TPU 8i - 2026-04 first split-inference, 288GB HBM largest per-chip inference ASIC available

Edge