Skip to main content

AWS Trainium 3 GA: 3nm Process + 4.4× Compute + 4× Efficiency + 144-Chip UltraServer

· 4 min read
Industry Research Team

On December 2, 2025, at the re:Invent 2025 conference, AWS formally GA'd its third-generation custom AI training chip Trainium 3. This is a critical upgrade to the AWS compute landscape: 3nm process, 4.4× compute improvement, 4× efficiency improvement, Trn3 UltraServer with 144 chips. This article provides a detailed analysis.

Core Specifications

ItemTrainium 2 (2024)Trainium 3 (2025-12 GA)Improvement
ProcessTSMC 4nmTSMC 3nm+1 generation
NeuronCore8 × v38 × v4Architecture upgrade
HBM capacity96 GB144 GB1.5×
HBM bandwidth2.9 TB/s~4.5 TB/s~1.55×
FP8 compute (dense)1,299 TFLOPS5,716 TFLOPS (official 4.4×)4.4×
BF16/FP16667 TFLOPS1,300 TFLOPS
Per-chip efficiency
Memory bandwidth
NeuronLinkNeuronLink-v3NeuronLink-v4Next generation
TDP~700 W~700 Wunchanged
Release date2024-122025-12

Official 4.4× compute improvement + 4× efficiency + 4× memory bandwidth — Trainium 3 is AWS's flagship chip with simultaneous massive upgrades across three dimensions.

Trn3 UltraServer (Rack-Level)

ItemConfiguration
Chip count144 Trainium 3 chips
Total HBM~20.7 TB (144GB × 144)
NeuronLink-v4Fully interconnected, >10 TB/s bidirectional
FP8 compute (rack)52 PFLOPS (dense)
BF16 compute (rack)~187 PFLOPS
TDP (rack)~100 kW
Capable models400B+ parameter LLM training

Trn3 UltraServer = single rack can train 400B models. A single EC2 UltraCluster (>10 racks) can support 1.4T+ parameter mega-model training.

Trn3 vs Trn2 UltraServer Upgrade

MetricTrn2 UltraServerTrn3 UltraServerImprovement
Chip count641442.25×
InterconnectNeuronLink-v3NeuronLink-v4Next generation
Total HBM6.1 TB~20.7 TB3.4×
FP8 compute~83 TFLOPS52 PFLOPS~626×
Training capacity70B+ LLM400B+ LLM
Release date2024-122025-12

Trn3 UltraServer is one of the most cost-effective large-scale training solutions in 2026.

AWS Neuron SDK 3

  • Neuron SDK 3.x: PyTorch 2.4+ / JAX 0.4+ / TensorFlow 2.16+ fully optimized
  • Neuron Compiler 2.x: auto-compilation + graph optimization
  • NeuronX Distributed: large-scale distributed training library (integrated with PyTorch FSDP)
  • NeuronX Nemo: LLM fine-tuning framework (Megatron-LM equivalent)
  • vLLM 0.7+ optimized version: low-latency inference

AWS Neuron = open-source ecosystem similar to ROCm, all SDKs open-source on GitHub (aws-neuron).

EC2 Instance Types

InstanceGPUConfigurationUse Case
trn3.48xlarge1 × Trn3144GB HBMSingle-chip development
trn3.96xlarge2 × Trn3288GB HBMSmall-scale training
trn3 UltraServer144 × Trn320.7 TB HBMExtreme-scale training

Pricing and Per-Dollar Performance

InstanceEstimated Hourly Price (on-demand)
trn3.48xlarge~$32
Trainium 2 equivalent instance~$16
Price increase
Per-dollar FP8 compute improvement2.2× (at 4.4× compute / 2× price)

AWS emphasizes: Trainium 3 is significantly better than NVIDIA H100 / H200 in per-dollar FP8 compute (2-3×).

Comparison with NVIDIA Contemporaries

MetricTrainium 3NVIDIA H200NVIDIA B200
ProcessTSMC 3nmTSMC 4NTSMC 4NP
HBM capacity144 GB141 GB192 GB
HBM bandwidth4.5 TB/s4.8 TB/s8 TB/s
FP8 compute (dense)5.7 PFLOPS1.0 PFLOPS4.5 PFLOPS
FP16 compute1.3 PFLOPS1.0 PFLOPS2.25 PFLOPS
TDP700 W700 W1,000 W
InterconnectNeuronLink-v4NVLink 4NVLink 5
AvailabilityAWS Cloud onlyCommercialCommercial
SoftwareNeuron SDK 3CUDACUDA
Per-dollar performance2-3× advantage1.5×

Applicable Scenarios

  • Extreme-scale LLM training (400B-1.4T models, UltraServer)
  • AWS Bedrock model pretraining (Anthropic Claude, Meta Llama, Mistral)
  • Cost-sensitive training (priced 30-50% below NVIDIA)
  • Energy-efficiency sensitive (4× per-watt performance improvement)
  • ❌ Non-AWS deployment (Trainium only sold via EC2)
  • ❌ Legacy NVIDIA ecosystem lock-in (CUDA-only code migration costs are high)

AWS Customer Case Studies

Key customers announced by AWS at re:Invent 2025:

CustomerApplication
AnthropicClaude training (already using Trn2, now migrating to Trn3)
MetaLlama 4 training
MistralMistral Large 3 training
HuggingFaceOpen LLM training
AWS BedrockInternal managed model training

Detailed Product Pages

Summary

AWS Trainium 3 is one of the key releases in the AI chip industry in 2025:

  1. 3nm process + 4.4× compute + 4× efficiency — AWS compute landscape comprehensively upgraded
  2. Trn3 UltraServer 144 chips — single rack trains 400B+ models
  3. Per-dollar FP8 compute 2-3× NVIDIA — AWS training cost advantage
  4. Neuron SDK 3 fully open-source — lowers software migration cost
  5. Anthropic, Meta, Mistral fully adopted — AWS compute ecosystem expanded

In 2026, Trainium 3 will be the compute foundation for AWS's internal core training workloads.