AWS Trainium - AshokBhat/ml GitHub Wiki

AWS Trainium

• Machine Learning training chip by AWS
• vs Nvidia A100 GPU
  • 60 percent more memory
  • 2X the networking bandwidth
• Performance
  • GPT-3 training in under two weeks
    • with 130 Trainium instances.
    • vs 600 P3DN instances
    • vs 200 P4 instances

AWS EC2 Trn1

Specification

Feature	Description
Clock speed	3Ghz
FLOPS	3.4 Peta FLOPS
FP32 FLOPS	840 Tera FLOPS
Memory Bandwidth	13.1 TB/sec
NeuronLink BW between chips	768 GB/sec

Availability

• In preview (Jan 2022)

Users

Amazon users

Machine learning workflow

• Uses AWS Neuron SDK

See also

• [Groq]] ](/AshokBhat/ml/wiki/[[Habana-Labs) | Graphcore
• Google TPU
• [AWS EC2]] ](/AshokBhat/ml/wiki/[[AWS-Elastic-Inference)
• [AWS Graviton]] ](/AshokBhat/ml/wiki/[[AWS-Inferentia)