AWS Trainium3 · Trainium

671 teraflops of dense BF16 compute a chip; AWS publishes no power figure

Shipped 2 Dec 2025; shipped by the first date promised. Epoch AI names it at 2 AI data centers and no GPU cluster. AWS across the site.

Embed

Trainium3

<iframe src="https://jesseahawkins.com/chips/embed/accelerators/aws-trainium3/" title="AWS Trainium3: 671 teraflops of dense BF16 compute a chip; AWS publishes no power figure" width="100%" height="520" style="border:0;max-width:100%" loading="lazy"></iframe>

The frame credits AWS, Epoch AI, AI data centers and Epoch AI, Data on GPU clusters and links back to its page. Preview it.

Cite

This page

Jesse Hawkins, “AWS Trainium3: 671 teraflops of dense BF16 compute a chip; AWS publishes no power figure”, jesseahawkins.com, data retrieved 23 Oct 2026. https://jesseahawkins.com/chips/accelerators/aws-trainium3/. Data: Each vendor’s own datasheets and releases, quoted; Epoch AI, “AI data centers” and “Data on GPU clusters”, CC BY 4.0.

BibTeX

@misc{hawkins2026chipsacceleratorsawstrainium3,
  author = {Hawkins, Jesse},
  title = {{AWS Trainium3: 671 teraflops of dense BF16 compute a chip; AWS publishes no power figure}},
  year = {2026},
  howpublished = {\url{https://jesseahawkins.com/chips/accelerators/aws-trainium3/}},
  note = {Data: Each vendor’s own datasheets and releases, quoted; Epoch AI, ``AI data centers'' and ``Data on GPU clusters'', CC BY 4.0, retrieved 2026-10-23}
}

The data · Epoch AI

Epoch AI, 'AI data centers'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/data/ai-data-centers' [online resource]. CC BY 4.0.

The clusters · Epoch AI

Epoch AI, 'Data on GPU clusters'. Published online at epoch.ai. Retrieved from 'https://epoch.ai/data/gpu-clusters' [online resource]. CC BY 4.0.

CSV Post on X

Each figure in AWS’s own words

one chip, dense compute; memory is high-bandwidth memory

BF16, dense: 671 TFLOPS 671 BF16/FP16/TF32 TFLOPS Trainium3 architecture (AWS Neuron documentation)
FP8, dense: 2,517 TFLOPS 2,517 MXFP8/MXFP4 TFLOPS Trainium3 architecture (AWS Neuron documentation)
FP4, dense: 2,517 TFLOPS 2,517 MXFP8/MXFP4 TFLOPS Trainium3 architecture (AWS Neuron documentation)
Memory: 144 GiB 144 GiB of device memory, with 4.9 TB/sec of bandwidth Trainium3 architecture (AWS Neuron documentation)
Memory bandwidth: 4.9 TB/s 144 GiB of device memory, with 4.9 TB/sec of bandwidth Trainium3 architecture (AWS Neuron documentation)

The FP8 and FP4 figures are for the MX block-scaled formats (MXFP8, MXFP4); AWS gives one figure for both. AWS states no chip power, and the Trn3 page states no on-demand price.

1 dated promise and the shipment

each date links to the page it is quoted from

3 Dec 2024 Promised first Trainium3 instances available in late 2025 in the second half of 2025. The first Trainium3-based instances are expected to be available in late 2025. AWS Trainium2 Instances Now Generally Available
2 Dec 2025 Shipped: Trn3 UltraServers generally available. AWS announces the general availability of Amazon Elastic Compute Cloud (Amazon EC2) Trn3 UltraServers powered by our fourth-generation AI chip Trainium3 Announcing Amazon EC2 Trn3 UltraServers for faster, lower-cost generative AI training

Named at 2 AI data centers

H100-equivalents and power are the whole site’s today, with every chip it holds

AI data centers that name this chip
Site	Owner	Country	H100e	MW	Chips
Amazon Ridgeland	Amazon	United States	171,299	273	Trainium2, Trainium3
AWS Berwick	Amazon	United States	81,354	139	Trainium2, Trainium3

In no GPU cluster of Epoch AI’s list

Epoch AI’s list of AI supercomputers, which stops in July 2025

Epoch AI’s names for this chip: Trainium3.

Other AWS chips

Trainium2 667 teraflops of dense BF16 compute a chip; AWS publishes no power figure

Sources

accelerators.csv Every dataset → Status → Glossary →

AWS’s own pages, each linked beside the figure or date it gives. Each page was read and each quote checked on 23 Oct 2026.
Where it runs: Epoch AI, AI data centers and Epoch AI, GPU clusters, CC BY 4.0.