AWS Trainium3
AWS Trainium3 is AWS’s fourth-generation machine-learning chip, with eight NeuronCore-v4 cores.
- Vendor
- AWS
- Architecture
- NeuronCore-v4 (8 cores)
- Memory
- 144 GB HBM3e
- Memory bandwidth (TB/s)
- 4.9 TB/s
- FP8 (TFLOPS)
- 2,517 TFLOPS
- FP4 (PFLOPS)
- 2.517 PFLOPS
- Form factor
- Cloud instances (Trainium3 UltraServers, up to 144 chips)
- Status
- Shipping
Trainium3 has 144 GB of HBM3e at 4.9 TB/s and twice Trainium2’s FP8 throughput; AWS lists 2,517 TFLOPS for MXFP8/MXFP4. Trainium3 UltraServers scale to 144 chips with up to 362 MXFP8 PFLOPS and 20.7 TB of HBM3e. AWS names Anthropic, Databricks and OpenAI among Trainium customers.
- Memory (GB)
- 144 GB
- Precision note (sparse or dense)
- AWS lists 2,517 TFLOPS for MXFP8/MXFP4 with no sparsity label; its separate sparse figure (2,517 TFLOPS) covers FP16/BF16/TF32.
- Vendor source
- awsdocs-neuron.readthedocs-hosted.com
- Checked
Questions
How much memory does AWS Trainium3 have?
Trainium3 has 144 GB of HBM3e with 4.9 TB/s of bandwidth.
How fast is Trainium3 compared with Trainium2?
Trainium3 has twice Trainium2’s FP8 throughput, with AWS listing 2,517 TFLOPS for MXFP8/MXFP4.
How big is a Trainium3 UltraServer?
Trainium3 UltraServers scale to 144 chips with up to 362 MXFP8 PFLOPS and 20.7 TB of HBM3e.
Who uses AWS Trainium?
AWS names Anthropic, Databricks and OpenAI among Trainium customers.
Source
Every figure is taken from the vendor’s own datasheet, product page, technical documentation or announcement, linked on each chip with the date it was last checked. How we pick and check sources: editorial standards.