AI Chips · Product page

Google TPU v5e: The Complete Guide

Google calls it “v5e” in public. Its own systems call it something else entirely. Simple to expert, every number links to its source.

Ask an AI about this chip:CAGPX
Google TPU v5e 10 sections · 2 levels 7 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

TPU v5e is Google’s cost-focused fifth-generation chip, announced 29 August 2023. Google calls it “the most cost-efficient, versatile, and scalable Cloud TPU to date,” built for both medium and large-scale training and inference 1.

Deep diveThe technical detail

Despite the “efficiency” framing, Google’s own announcement explicitly claims large-scale capability too, not just mid-size workloads 1 — v5p, covered on its own page, is the dedicated large-scale-training sibling launched a few months later.

2.Launch and history

SimpleStart here

TPU v5e arrived as Google split its fifth TPU generation into two purpose-built chips instead of one general-purpose design: v5e for cost-efficient, versatile workloads, and v5p (announced three months later) for the largest training jobs.

Deep diveThe technical detail

Google’s own launch claims: up to 2x higher training performance per dollar and up to 2.5x higher inference performance per dollar versus TPU v4, at less than half TPU v4’s cost 1. Google cited early customers AssemblyAI, reporting up to 4x greater performance per dollar on speech-recognition inference, and Gridspace, reporting a 5x training speedup and 6x inference improvement 1.

3.What’s inside it

SimpleStart here

One TensorCore per chip, with four matrix-multiply units plus a vector and scalar unit 2 — a simpler per-chip design than v5p’s two-TensorCore layout, in keeping with v5e’s cost-efficiency focus.

Deep diveThe technical detail

Google’s own documentation does not disclose which HBM generation (HBM2e or HBM3) v5e uses, only capacity and bandwidth figures 2; this page does not guess a specific memory standard. Google also does not publish a process node or foundry for v5e in any official source found.

4.Full spec table

SimpleStart here

197 teraflops (bf16) and 393 TOPs (int8) per chip, 16GB of HBM memory, connected up to 256 chips per pod 2.

Deep diveThe technical detail
SpecGoogle TPU v5e
Peak compute (bf16)197 TFLOPs per chip
Peak compute (int8)393 TOPs per chip
Memory16 GB HBM, 800 GiBps bandwidth
Inter-chip interconnect400 GBps, 4 ports
Pod size256 chips, 2D torus
Pod peak (bf16)50.63 PFLOPs
Pod peak (int8)100 PetaOps
Chips per host8

Figures per Google Cloud’s TPU v5e documentation 2. Google’s launch blog separately describes pod-level aggregate bandwidth as “more than 400 Tb/s,” consistent with the documented 51.2 TB/s all-reduce figure once converted 1.

5.Where it’s made

SimpleStart here

Google does not disclose who manufactures TPU v5e or on what process node. No official Google or Broadcom source names a foundry for this specific chip.

Deep diveThe technical detail

Google’s manufacturing partnership with Broadcom is documented at the company level for “future generations” of TPUs, but no official filing or blog post ties Broadcom, or any named foundry such as TSMC, to TPU v5e specifically. This page does not state one.

6.Which systems use it

SimpleStart here

TPU v5e is available through Google Cloud across eight regions in North America, Europe and Asia 34.

Deep diveThe technical detail

Google’s own documentation names the accelerator type internally as v5litepod (for example, v5litepod-8), not “v5e” 5. Regions offering v5e include us-central1, us-east5, us-south1, us-west1, us-west4, europe-west1, europe-west4 and asia-southeast1 34.

7.Official pricing

SimpleStart here

Google Cloud’s own pricing page lists TPU v5e on-demand at $1.20 per chip-hour in most regions, rising to $1.56 in europe-west4 and asia-southeast1 4.

Deep diveThe technical detail
RegionOn-demand ($/chip-hr)3-year committed use
us-central1, us-east5, us-west1, us-west4$1.20$0.54
us-south1$1.416$0.6372
europe-west1$1.32132$0.594594
europe-west4, asia-southeast1$1.56$0.702

Figures per Google Cloud’s own pricing page, checked September 2026 4. A flexible “Dynamic Workload Scheduler” option is also listed at $0.60/chip-hour (Flex-start) or $0.84/chip-hour (Calendar mode) across all regions. Spot pricing is not published as a fixed rate; Google states it changes dynamically, at most once every 30 days.

8.Real-world performance

SimpleStart here

Google claims up to 2x higher training performance per dollar and up to 2.5x higher inference performance per dollar versus TPU v4 1 — company-stated comparisons, not an independent benchmark.

Deep diveThe technical detail

This page reports Google’s own launch-announcement figures rather than a third-party MLPerf submission specific to v5e. Customer-reported figures cited by Google include AssemblyAI (up to 4x performance-per-dollar on inference) and Gridspace (5x training speedup, 6x inference improvement) 1 — customer claims relayed by Google, not independently audited.

9.What came before, what came next

SimpleStart here

Came before: TPU v4. Came after: Trillium (TPU v6e), announced May 2024, which Google says delivers 4.7 times the peak compute per chip 6.

Deep diveThe technical detail

Google’s own release notes state Trillium improves “over the prior generations, v5e and v5p” together 7. Its launch blog specifically compares itself to v5e: 4.7x peak compute per chip, double the HBM capacity and bandwidth, double the inter-chip interconnect bandwidth, and 67% better energy efficiency 6.

10.Hidden in plain sight

SimpleStart here

The name “TPU v5e” barely exists inside Google’s own systems — its APIs call it something else entirely.

Deep diveThe technical detail

Google’s own developer documentation for creating a Cloud TPU VM never uses the string “v5e” in the actual accelerator-type or runtime-version identifiers a developer types: they are v5litepod-8 and v2-tpuv5-litepod 5. The public-facing “e” that most people read as “efficient” is, inside Google’s own tooling, simply “lite.”

11.Sources

7 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. Google Cloud Blog: Announcing Cloud TPU v5e and A3 GPUs in GAOfficial
  2. Google Cloud: TPU v5e DocumentationOfficial
  3. Google Cloud: Cloud TPU Regions and ZonesOfficial
  4. Google Cloud: Cloud TPU PricingOfficial
  5. Google Cloud: Creating a Cloud TPU VMOfficial
  6. Google Cloud Blog: Introducing Trillium, Sixth-Generation TPUsOfficial
  7. Google Cloud: Cloud TPU Release NotesOfficial