← All hardware
Amazon's newest 3nm-class training chip, up to 4.4x faster than Trn2.
Pros
- 4.4x performance and 4x perf/watt vs Trn2
- Strong cost efficiency for AWS-native training
- Massive UltraCluster scale-out
Cons
- AWS-only and Neuron-SDK dependent
- Mostly offered via Capacity Blocks / contracts, not flat on-demand
- Smaller third-party ecosystem than CUDA
✓ Where it shines / best for
- Large-scale generative-AI training on AWS
- Cost-optimized hyperscale UltraClusters
- Teams committed to the AWS Neuron stack
✕ Not the best fit for
- On-prem or multi-cloud portability
- CUDA-only workloads
- Small inference deployments
Features
- ✓ 2.52 PFLOPS FP8 compute per chip
- ✓ 144 GB HBM3e per chip, 4.9 TB/s bandwidth
- ✓ Trn3 UltraServers scale to 144 chips (362 PFLOPS FP8)
- ✓ Up to 20.7 TB HBM3e per UltraServer
- ✓ 4x better perf/watt vs Trn2 UltraServers
- ✓ EC2 UltraClusters 3.0 for hundreds of thousands of chips
- ✓ AWS Neuron SDK (PyTorch/JAX)
- ✓ 706 TB/s aggregate memory bandwidth per UltraServer
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| Trn3 UltraServer On-Demand | Varies | per hour | GA Dec 2, 2025. AWS has not published standard on-demand hourly pricing; UltraServers offered mainly via Capacity Blocks and long-term contracts. Positioned at roughly 2x better price-performance per-TFLOPS vs Trn2. |
| Savings Plans / Reserved / Capacity Blocks | Discounted | 1-3 year | Lower rates with commitments and reserved capacity |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| architecture | Trainium3, AWS custom ASIC |
| fp8_per_chip | 2.52 PFLOPS |
| perf_per_watt | 4x vs Trn2 UltraServers |
| memory_per_chip | 144 GB HBM3e, 4.9 TB/s |
| ultraserver_scale | 144 chips, 362 PFLOPS FP8, 20.7 TB HBM3e |
Sponsored
A full review is being generated for this product and will appear here shortly.