AWS Trainium 3 (Trn3)
Overview
AWS Trainium 3 reached GA on 2025-12-02 at re:Invent 2025, as the third-generation AWS custom AI training chip. Built on a 3nm process, it delivers 5.7 PFLOPS FP8 per chip (dense, officially 4.4× that of Trainium 2), 4× energy efficiency over Trainium 2, and 4× memory bandwidth improvement. Trn3 UltraServer connects 144 chips via NeuronLink (2.25× Trn2 UltraServer's 64 chips).
Trainium 3 is the centerpiece of AWS's "AI Factory" strategy — Amazon's internal Bedrock, Anthropic Claude, and AWS customer core training workloads are all migrating to Trn3.
Core Specifications
| Item | Specification |
|---|---|
| Architecture | Trainium3 (NeuronCore-v4) |
| Process Node | TSMC 3nm |
| NeuronCore | 8 per chip (NeuronCore-v4) |
| HBM Capacity | 144 GB (estimated, 2× Trainium 2) |
| HBM Bandwidth | ~4.5 TB/s (estimated) |
| FP8 Compute (dense) | 5.7 PFLOPS |
| BF16/FP16 | 2,850 TFLOPS (estimated, half of FP8) |
| TDP | ~700 W |
| NeuronLink | NeuronLink-v4 |
| Launch | 2025-12-02 GA (re:Invent 2025) |
📌 Data convention: AWS Trainium uses dense compute as standard (consistent with AMD, Google); not directly comparable to NVIDIA sparse compute. 5.7 PFLOPS FP8 = dense (= 5,700 TFLOPS).
Trainium 2 vs Trainium 3 Upgrade Comparison
| Metric | Trainium 2 | Trainium 3 | Improvement |
|---|---|---|---|
| Process | TSMC 4nm | TSMC 3nm | +1 generation |
| NeuronCore | 8 v3 | 8 v4 | Architecture upgrade |
| HBM Capacity | 96 GB | 144 GB (estimated) | 1.5× |
| HBM Bandwidth | 2.9 TB/s | ~4.5 TB/s | ~1.55× |
| FP8 Compute (dense) | 1,299 TFLOPS | 5.7 PFLOPS | ~4.4× |
| BF16/FP16 | 667 TFLOPS | 2,850 TFLOPS (estimated) | ~4.3× |
| FP8 Compute per chip increase | — | 4.4× | Official |
| Energy Efficiency (perf/watt) | — | 4× | Official |
| Memory Bandwidth increase | — | 4× | Official |
| Launch | 2024-12 | 2025-12 | — |
⚠️ 5.7 PFLOPS = per chip (FP8 dense), while Trainium 2 is 1,299 TFLOPS/chip. Official 4.4× compute increase: 1,299 × 4.4 ≈ 5,716 TFLOPS ≈ 5.7 PFLOPS. Trust the official 4.4× compute improvement.
Trn3 UltraServer (Rack-Level)
| Item | Configuration |
|---|---|
| Chip Count | 144 Trainium 3 (4× Trn2 UltraServer's 64) |
| Total HBM | ~20.7 TB (144GB × 144) |
| NeuronLink-v4 | Fully connected, >10 TB/s bidirectional |
| FP8 Compute (rack) | 820 PFLOPS (dense, 144 × 5.7 PFLOPS) |
| BF16 Compute (rack) | ~410 PFLOPS |
| TDP (rack) | ~100 kW |
| Suitable Models | 400B+ parameter LLM training |
Trn3 UltraServer = single rack can train 400B models. An EC2 UltraCluster (>10 racks) can support 1.4T+ parameter mega-model training.
Trn3 vs Trn2 UltraServer
| Metric | Trn2 UltraServer | Trn3 UltraServer | Improvement |
|---|---|---|---|
| Chip Count | 64 | 144 | 2.25× |
| Interconnect | NeuronLink-v3 | NeuronLink-v4 | New generation |
| Total HBM | 6.1 TB | ~20.7 TB | 3.4× |
| FP8 Compute | ~83 TFLOPS (64×1.3) | ~365 TFLOPS (144×2.54) | ~4.4× |
| Training Capability | 70B+ LLM | 400B+ LLM | — |
| Launch | 2024-12 | 2025-12 | — |
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: Low-latency inference
AWS Neuron = open-source ecosystem similar to ROCm, all SDKs are open source on GitHub (aws-neuron).
EC2 Instance Types
| Instance | GPU | Configuration | Use Case |
|---|---|---|---|
| trn3.48xlarge | 1 Trn3 | 144GB HBM | Single-chip development |
| trn3.96xlarge | 2 Trn3 | 288GB HBM | Small-scale training |
| trn3 UltraServer | 144 Trn3 | 20.7 TB HBM | Ultra-large-scale training |
Pricing (Estimated)
| Instance | Hourly Price (on-demand) |
|---|---|
| trn3.48xlarge | ~$32 (estimated) |
| Trainium 2 equivalent | ~$16 |
| Price increase | 2× |
| FP8 compute per dollar increase | 2.2× (based on 4.4× compute / 2× price) |
AWS emphasizes: Trainium 3 is significantly better than NVIDIA H100 / H200 in FP8 compute per dollar (2-3×).
Use Cases
- ✅ Ultra-large-scale LLM training (400B-1.4T models, UltraServer)
- ✅ AWS Bedrock model pre-training (Anthropic Claude, Meta Llama, Mistral)
- ✅ Cost-sensitive training (30-50% lower price than NVIDIA)
- ✅ Energy-efficiency sensitive (4× perf/watt improvement)
- ❌ Non-AWS deployments (Trainium only available on EC2)
- ❌ Legacy NVIDIA ecosystem lock-in (high migration cost for CUDA-only code)
Vendor Information
| Item | Details |
|---|---|
| Vendor | Amazon Web Services (AWS) |
| First Release | 2025-12-02 (re:Invent 2025 GA) |
| Product Page | https://aws.amazon.com/machine-learning/trainium/ |
| Cloud Instances | EC2 trn3.48xlarge / 96xlarge / UltraServer |
| SDK | https://github.com/aws-neuron |
| Partners | Anthropic / Meta / Mistral / HuggingFace |
Related Products
- AWS Trainium 2 - Previous generation chip
- AWS Trainium 1 - First generation
- AWS Inferentia 2 - Inference counterpart
- NVIDIA H100 - Primary competitor
- NVIDIA B200 - Contemporary flagship GPU
- Google TPU v7 Ironwood - Same-generation ASIC
- Full Comparison Table