Skip to main content

Google TPU v5e (Trillium Training-Lite, 2023)

Overview

Google TPU v5e (unofficial codename Trillium-Lite) is the entry-level / value-oriented version of Google's 5th generation TPU, released 2023-Q2. Built on TSMC 5nm, featuring 16GB HBM2 memory, 400 TFLOPS FP8 dense compute, and 180W TDP. It is positioned for inference + small-to-mid-scale training, priced ~70% lower per chip than the TPU v5p (training flagship).

Key positioning:

  • TPU v5p (2023-Q3): 96GB HBM2, 1.89 PF FP8, training-only — separate page
  • TPU v5e (2023-Q2): 16GB HBM2, 400 TF FP8, inference + small trainingthis page
  • TPU v6e (2024-Q2): 32GB HBM2, 1.5 PF FP8, Trillium — separate page
  • TPU v6p (2024-12): 96GB HBM2, 2.7 PF FP8, Pathway training — separate page

Core Specifications

ItemSpec
CodenameTrillium-Lite (Google internal: v5e)
ArchitectureGoogle TPU v5 (same generation as v5p)
ProcessTSMC 5nm
MXU128×128 (2 units, v5p has 4)
HBM16GB HBM2 (v5p: 96GB)
HBM Bandwidth400 GB/s (v5p: 1.4 TB/s)
FP8 dense400 TFLOPS (v5p: 1.89 PF)
BF16 dense200 TFLOPS
INT8400 TOPS
TDP180W (v5p: 450W)
Form FactorCloud TPU v5e pod slice
Pod Scale256 chips (v5p: 8,960)
Pod Compute102 TF FP8 dense (v5p: 16.9 EF)
Pod Bandwidth1.6 TB/s intra-domain
Production2023-Q2
Price (Google Cloud)~$1.20/hr (pod slice)

Comparison with TPU v5p

MetricTPU v5e (2023-Q2)TPU v5p (2023-Q3)Difference
PositioningInference + small trainingLarge-scale training-
Process5nm5nmSame
MXU2× 128×1284× 128×1281/2
HBM16GB HBM296GB HBM21/6
Bandwidth400 GB/s1.4 TB/s1/3.5
FP8 dense400 TF1.89 PF1/4.7
TDP180W450W1/2.5
Pod Scale2568,9601/35
Price (Google Cloud)$1.20/hr$4.20/hr1/3.5
Suitable Models7B-30B70B-540B-

TPU Product Line Comparison

GenerationCodenameMemoryFP8 densePod ScaleSuitable For
TPU v4-32GB HBM21.1 PF4,096100B+
TPU v5e-16GB HBM2400 TF2567B-30B
TPU v5p-96GB HBM21.89 PF8,96070B-540B
TPU v6eTrillium32GB HBM21.5 PF2567B-70B
TPU v6pPathway96GB HBM22.7 PF9,21670B-trillion
TPU v7Ironwood192GB HBM3E4.6 PF9,216192GB inference

TPU v5e Use Cases

  • LLM inference (7B-30B model inference)
  • Small model training (LLaMA 7B, Mistral 7B, Qwen 1.5 14B)
  • Recommendation systems (SparseCore optimized)
  • Google Cloud TPU entry point ($1.20/hr)
  • JAX / Flax training (XLA optimized)
  • Anthropic / Cohere / Mistral (Google Cloud customers)
  • Ultra-large model training (16GB memory limitation)
  • FP8 training (FP8 inference only, BF16 for training)
  • Native PyTorch (requires XLA translation)

Inference vs Training Advantages

Inference

  • TTFT < 10ms (JAX + Pathways)
  • TPOT 5-8ms (4-card interconnect)
  • Price $1.20/hr (H100 $3-5/hr, 60% cheaper)
  • 7B-30B LLM optimized

Training

  • LLaMA 7B training: v5e 256 cards = 1.5 steps/sec (H100 8 cards = 1 step/sec, comparable)
  • LLaMA 13B training: v5e 256 cards = 0.7 steps/sec (H100 8 cards = 0.5 steps/sec, v5e slightly ahead)
  • JAX + Flax + GSPMD tensor parallelism
  • Price $1.20/hr (H100 8-card $25-30/hr, 1/10 the price)

Software Stack

LayerToolDescription
AI FrameworksJAXGoogle-recommended
FlaxJAX neural network library
OptaxJAX optimizer
RLlibJAX reinforcement learning
PathwaysUnified heterogeneous accelerator programming
TensorFlowCompatible
PyTorch/XLACompatible (experimental)
CompilerXLAAccelerator compiler
DistributedGSPMDTensor parallelism
Collective CommunicationDUSProprietary
Model LibraryMaxText (Gemma 2 training)Google open-source

Vendor Information

ItemDetail
CompanyGoogle LLC
Product Pagehttps://cloud.google.com/tpu
Business UnitGoogle Cloud + Google DeepMind
FoundryTSMC 5nm (InFO_SoC packaging)
Google Cloud Pricingv5e ~$1.20/hr (pod slice)
CustomersGoogle internal (Search, YouTube, DeepMind) + Anthropic / Cohere / Mistral / Hugging Face

Comparison with NVIDIA L4 (Inference)

MetricGoogle TPU v5eNVIDIA L4Difference
INT8400 TOPS485 TOPSL4 +21%
TDP180W72Wv5e 2.5×
Energy Efficiency2.22 TOPS/W6.7 TOPS/WL4 3×
Memory16GB HBM224GB GDDR6L4 1.5×
Price$1.20/hr$0.80-1.20/hrComparable
SoftwareJAXCUDAL4 mature

TPU v5e advantage: Google Cloud integration + JAX optimization + low price. L4 advantage: 72W TDP (v5e 40% power saving) + mature software + multi-cloud deployment.

Key Features

  • 400 TF FP8: Industry 5nm entry-level TPU flagship
  • 180W TDP: 25% of H100 power
  • 16GB HBM2: Sufficient for inference, constrained for training
  • 256-chip Pod: JAX GSPMD training optimized
  • Low price: $1.20/hr
  • Drawbacks: Small memory, Google Cloud only, 5-year ecosystem