inference/research
contestedlow · 0.50updated 2026-10-15

Trainium3 vs Trainium2 on Amazon Bedrock: output tokens per megawatt ratio

5.0 x higher output tokens per megawatt at similar latency per user (vendor claim)

Fields

id
benchmarks/aws-trn3-vs-trn2-bedrock-tokens-per-mw-ratio
Metric
dimensionless_ratio
Value
5
Unit
x higher output tokens per megawatt at similar latency per user (vendor claim)
Model
unspecified (vendor claim, workload not stated)
Methodology
VENDOR CLAIM (AWS What's New, Trn3 UltraServer GA, Dec 2 2025). 'On Amazon Bedrock, Trainium3 is our fastest accelerator, delivering... over 5x higher output tokens per megawatt at similar latency per user' vs Trainium2. Energy-normalized throughput ratio, not a speedup.
Measured by
vendor
Reproducible
no
notes
Filed as dimensionless_ratio (not speedup_ratio) because it is a throughput-per-watt figure, not a faster/slower time ratio. The only energy-normalized number for the part.

Notes

Filed as dimensionless_ratio (not speedup_ratio) because it is a throughput-per-watt figure, not a faster/slower time ratio. The only energy-normalized number for the part.

Cross-references