AWS Trainium 3 GA: 3nm Process + 4.4× Compute + 4× Efficiency + 144-Chip UltraServer
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
| Item | Trainium 2 (2024) | Trainium 3 (2025-12 GA) | Improvement |
|---|---|---|---|
| Process | TSMC 4nm | TSMC 3nm | +1 generation |
| NeuronCore | 8 × v3 | 8 × v4 | Architecture upgrade |
| HBM capacity | 96 GB | 144 GB | 1.5× |
| HBM bandwidth | 2.9 TB/s | ~4.5 TB/s | ~1.55× |
| FP8 compute (dense) | 1,299 TFLOPS | 5,716 TFLOPS (official 4.4×) | 4.4× |
| BF16/FP16 | 667 TFLOPS | 1,300 TFLOPS | 2× |
| Per-chip efficiency | 1× | 4× | 4× |
| Memory bandwidth | 1× | 4× | 4× |
| NeuronLink | NeuronLink-v3 | NeuronLink-v4 | Next generation |
| TDP | ~700 W | ~700 W | unchanged |
| Release date | 2024-12 | 2025-12 | — |
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)
| Item | Configuration |
|---|---|
| Chip count | 144 Trainium 3 chips |
| Total HBM | ~20.7 TB (144GB × 144) |
| NeuronLink-v4 | Fully interconnected, >10 TB/s bidirectional |
| FP8 compute (rack) | 52 PFLOPS (dense) |
| BF16 compute (rack) | ~187 PFLOPS |
| TDP (rack) | ~100 kW |
| Capable models | 400B+ 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
| Metric | Trn2 UltraServer | Trn3 UltraServer | Improvement |
|---|---|---|---|
| Chip count | 64 | 144 | 2.25× |
| Interconnect | NeuronLink-v3 | NeuronLink-v4 | Next generation |
| Total HBM | 6.1 TB | ~20.7 TB | 3.4× |
| FP8 compute | ~83 TFLOPS | 52 PFLOPS | ~626× |
| Training capacity | 70B+ LLM | 400B+ LLM | — |
| Release date | 2024-12 | 2025-12 | — |
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
| 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 | Extreme-scale training |
Pricing and Per-Dollar Performance
| Instance | Estimated Hourly Price (on-demand) |
|---|---|
| trn3.48xlarge | ~$32 |
| Trainium 2 equivalent instance | ~$16 |
| Price increase | 2× |
| Per-dollar FP8 compute improvement | 2.2× (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
| Metric | Trainium 3 | NVIDIA H200 | NVIDIA B200 |
|---|---|---|---|
| Process | TSMC 3nm | TSMC 4N | TSMC 4NP |
| HBM capacity | 144 GB | 141 GB | 192 GB |
| HBM bandwidth | 4.5 TB/s | 4.8 TB/s | 8 TB/s |
| FP8 compute (dense) | 5.7 PFLOPS | 1.0 PFLOPS | 4.5 PFLOPS |
| FP16 compute | 1.3 PFLOPS | 1.0 PFLOPS | 2.25 PFLOPS |
| TDP | 700 W | 700 W | 1,000 W |
| Interconnect | NeuronLink-v4 | NVLink 4 | NVLink 5 |
| Availability | AWS Cloud only | Commercial | Commercial |
| Software | Neuron SDK 3 | CUDA | CUDA |
| Per-dollar performance | 2-3× advantage | 1× | 1.5× |
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:
| Customer | Application |
|---|---|
| Anthropic | Claude training (already using Trn2, now migrating to Trn3) |
| Meta | Llama 4 training |
| Mistral | Mistral Large 3 training |
| HuggingFace | Open LLM training |
| AWS Bedrock | Internal managed model training |
Detailed Product Pages
- AWS Trainium 3 Full Specs
- AWS Trainium 2 (Previous Gen)
- AWS Trainium 1 (First Gen)
- AWS Inferentia 2 (Inference Counterpart)
- NVIDIA H100 (Main Competitor)
- Future Roadmap
Summary
AWS Trainium 3 is one of the key releases in the AI chip industry in 2025:
- 3nm process + 4.4× compute + 4× efficiency — AWS compute landscape comprehensively upgraded
- Trn3 UltraServer 144 chips — single rack trains 400B+ models
- Per-dollar FP8 compute 2-3× NVIDIA — AWS training cost advantage
- Neuron SDK 3 fully open-source — lowers software migration cost
- 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.