Skip to main content

Google Cloud TPU 8i (Trillium 2 / Inference-Dedicated)

Overview

Google TPU 8i (codenamed Trillium 2 Inference Edition) is the latest inference-dedicated TPU, announced on 2026-04-22, forming an 8t + 8i split architecture with the simultaneously announced TPU 8t training-dedicated TPU. It features 288GB HBM (50% more than TPU v7 Ironwood), 8,601 GB/s bandwidth, and ~5,500 TFLOPS BF16 compute (dense).

TPU 8i is the core of Google's "AI Inference Era" strategy — Gemini API, Vertex AI inference, Anthropic Claude on Vertex, and Gemini 3 / 4 online serving are all powered by TPU 8i.

Core Specifications

ItemSpecification
ArchitectureTPU 8i (Trillium 2)
TypeInference-dedicated (distinct from 8t training-dedicated)
BF16 Compute (dense)~5,500 TFLOPS
FP8 Compute (dense)~11,000 TFLOPS
INT8 Compute~22,000 TOPS
HBM Capacity288 GB
HBM Bandwidth8,601 GB/s
ICI Interconnect1,200 GB/s
DCN Bandwidth200 Gbps
Pod SizeSingle chip ~256 chips
CoolingAir or liquid cooling
Announcement2026-04-22

📌 8i naming: TPU 8th-gen + i = inference. 8i is the inference ASIC with the largest memory currently, a single card at 288GB can hold a 70B model (FP16).

TPU 8i vs TPU v7 Ironwood (Inference Comparison)

MetricTPU v7 IronwoodTPU 8iImprovement
TypeTraining + InferenceInference-dedicatedType split
BF16 Compute2,307 TFLOPS~5,500 TFLOPS2.4×
FP8 Compute4,614 TFLOPS~11,000 TFLOPS2.4×
HBM Capacity192 GB288 GB1.5×
HBM Bandwidth7,380 GB/s8,601 GB/s1.17×
CoolingLiquid primaryAir/liquid flexibleFlexible
Announcement2025-112026-04-22

💡 TPU 8i compute 2.4× higher than Ironwood: 8,601 GB/s bandwidth + 288GB HBM enables TPU 8i to handle long-context inference and ultra-large-model inference with single-card capacity for 70B+ models.

TPU 8i Inference Paradigm Optimization

OptimizationDetails
Ultra-low latencyTTFT < 100ms (Time To First Token)
High throughput10,000+ tok/s (70B model FP8)
Long-context KV288GB fully retains 1M+ token context
MoE InferenceNative Expert Parallel support
Speculative DecodingInternal speculative acceleration
BatchingContinuous batching + PagedAttention
Continuous KV CacheKV Cache cross-request sharing (same-prefix optimization)

TPU 8i vs TPU 8t (Simultaneous Split)

MetricTPU 8t (Training)TPU 8i (Inference)
PositioningTrainingInference
BF16 Compute~3,500 TFLOPS~5,500 TFLOPS (higher)
HBM Capacity216 GB288 GB (larger)
HBM Bandwidth6,528 GB/s8,601 GB/s (higher)
CoolingLiquidAir/liquid
Pod Size9,216 chips256 chips
Integrated CPUArm AxionNone (standalone)

💡 Split purpose: Training emphasizes compute + interconnect; inference emphasizes memory + bandwidth + cooling flexibility. 8t = liquid + large pod; 8i = air + small pod + massive memory.

ScenarioRecommended Configuration
Gemini API Online ServingTPU 8i pod (million-level QPS)
Claude on Vertex AITPU 8i single chip / 4-chip node
Llama 4 70B InferenceTPU 8i single card (288GB fits FP16 70B)
Long-context RAGTPU 8i (1M+ token KV Cache)
Edge / Offline InferenceTPU 8i air-cooled (no liquid cooling facility required)

Software Ecosystem

  • JAX 0.5+: Inference
  • PyTorch/XLA 2.5+: Inference
  • vLLM 0.8+ (TPU backend): Low-latency inference
  • Vertex AI Inference: Google managed inference service
  • Gemini API: Primary internal user

Pricing (Estimated)

InstanceHourly PriceNotes
TPU 8i v6e-equivalent~$3-5 / chipEstimated
TPU v7 Ironwood~$6-8 / chipCurrent mainstream
TPU 8i vs TPU v7+50% price / +150% computeBetter price-performance

TPU 8i delivers 70% higher BF16 compute per dollar than TPU v7 Ironwood (based on 2.4× compute / 1.5× price).

Use Cases

  • Frontier model inference (Gemini 3/4, Claude Opus 4.5)
  • Ultra-low-latency online serving (TTFT < 100ms)
  • Long-context RAG / Agent (1M+ token inference)
  • High-throughput offline inference (10,000+ tok/s)
  • Air-cooled deployment (no liquid cooling facility required)
  • ❌ Training scenarios (use TPU 8t instead of 8i)

Vendor Information

ItemDetails
VendorGoogle Cloud
First Announced2026-04-22 (Google Cloud Next 2026)
Product Pagehttps://cloud.google.com/tpu
Cloud DeploymentGoogle Cloud only (Vertex AI / Gemini API)
CodenameTrillium 2 (Inference Edition)