Beta
ChipsReading · ~3 min · 59 words deep

Trainium 3

Trainium 3 is AWS's third-gen AI training ASIC · announced November 2024 · claims 2× perf/watt over Trainium 2 and powers Anthropic's AWS training clusters.

Trainium 3 on hardware map
TL;DR

Trainium 3 is AWS's third-gen AI training ASIC · announced November 2024 · claims 2× perf/watt over Trainium 2 and powers Anthropic's AWS training clusters.

Level 1

Announced at re:Invent 2024, Trainium 3 is AWS's custom training chip. Key specs: 5nm process (TSMC), ~860 FP8 TFLOPS per chip, HBM3e memory. Deployed in Ultraclusters of up to 100K chips via EFA networking. Anthropic signed multi-year training commitments on Trainium 3 as part of Amazon's $8B investment.

Level 2

Trainium 3 targets training cost reduction for frontier models. AWS claims 40% better price-performance than comparable H100 clusters at equivalent throughput. The chip uses a NeuronCore architecture (custom AWS design, not CUDA-compatible). Software stack is AWS Neuron SDK + PyTorch XLA. This is AWS's play to reduce NVIDIA dependency for hyperscale training workloads · Anthropic is the flagship customer.

Level 3

Trainium 3 cluster scale: up to 100K chips in Project Rainier (joint with Anthropic), interconnected via EFA v3 at 3.2 TB/s per chip. Neuron compiler supports PyTorch and JAX via XLA; native CUDA code does NOT run. Perf/watt is ~2× T2 but absolute TFLOPS trail H200 · competitive on price per training-hour, not raw perf. General availability expected Q2 2026; Anthropic reserved bulk capacity.

The takeaway for you
If you are a
Researcher
  • ·3rd-gen AWS training ASIC · 5nm TSMC
  • ·~860 FP8 TFLOPS · HBM3e
  • ·NeuronCore architecture · not CUDA-compatible
If you are a
Builder
  • ·Access via AWS only · no retail
  • ·Requires AWS Neuron SDK + PyTorch XLA
  • ·Best for training workloads already on AWS
If you are a
Investor
  • ·AWS's counter-bet to NVIDIA dominance · key for AWS margin protection
  • ·Anthropic commitment anchors Trainium revenue
  • ·Gross margin on Trainium inference is much higher than reselling NVIDIA
If you are a
Curious · Normie
  • ·Amazon's own AI training chip · alternative to NVIDIA
  • ·Used to train Claude models on AWS
  • ·Part of Amazon's $8B Anthropic investment
Gecko's take

Trainium 3 is AWS's bet that Anthropic's training needs can anchor a real NVIDIA alternative. The Project Rainier 100K cluster proves the concept · execution is the next question.

H200 has higher peak FLOPS. Trainium 3 claims better price-performance on AWS-hosted training workloads. Different software stack (Neuron vs CUDA).