← All hardware
Amazon's 3nm training chip: 2.52 PFLOPS FP8, 144GB HBM3e
Pros
- Strong cost-per-token on AWS
- Proven at frontier-lab scale
- Deep AWS integration
- Big generational jump over Trainium2
Cons
- AWS-only
- Neuron SDK porting effort
- Not a CUDA drop-in
- Cloud-rental only
✓ Where it shines / best for
- AWS customers planning next-gen frontier-model training
- Workloads needing better performance-per-watt than Trn2
- Large-scale UltraCluster training buildouts
✕ Not the best fit for
- Anyone needing GA hardware today (preview)
- On-premises or edge deployments
- Non-Neuron software stacks unwilling to port
Features
- ✓ AI inference
- ✓ Data-center scale
- ✓ LLM
- ✓ API access
- ✓ AI Training
- ✓ Cloud Only
- ✓ Energy Efficient
- ✓ Neuron SDK
- ✓ Next Gen
- ✓ Preview
- ✓ 3nm
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| On-demand (cloud) | Not yet published | hourly | Trainium3 announced (preview); will be accessed via future EC2 Trn3 instances on a per-instance-hour basis. No public pricing yet; not sold as standalone silicon. |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| use | Cloud AI training and inference |
| power | Data center scale |
| memory | 144GB HBM3e |
| performance | 2.52 PFLOPS FP8 per chip |
| architecture | Trainium3 (3nm) |
Sponsored
A full review is being generated for this product and will appear here shortly.