AWS Trainium2 vs AWS Trainium3

The Trainium3 delivers 94% more FP8 throughput than the Trainium2 (2,517 vs 1,299 TFLOPS dense). The Trainium3 also carries 48 GB more memory (144 GB vs 96 GB).

Trainium2: NeuronCore-v3, 2024 Trainium3: NeuronCore-v4, 2025

FP8 dense lead

+94%

Trainium3: 2,517 vs 1,299 TFLOPS

Memory

96 vs 144 GB

48 GB more for the Trainium3

Bandwidth

2.9 TB/s vs 4.9 TB/s

Trainium3 moves data faster

Download comparison sheet Printable version →

AWS Trainium2

Full specs →

What is the AWS Trainium2?

The Trainium2 is an ASIC from AWS on the NeuronCore-v3 architecture, an accelerator built for AI workloads. Trainium2 full specifications →

Can I rent the Trainium2 in the cloud?

Flopper is not tracking any live rental listings for the Trainium2 yet. The provider directory lists the clouds that rent comparable accelerators on demand. GPU cloud providers →

How does the Trainium2 compare with other accelerators?

This page sets the Trainium2 against the Trainium3. The Trainium2 is the older generation of the two. Both are AWS parts. Flopper also compares it head to head with the H100 SXM5 80GB, across every published precision, memory and power. Trainium2 vs H100 SXM5 80GB →

Where do the Trainium2 figures come from?

Every Trainium2 figure on this page traces to AWS's own datasheet or specification table. How Flopper verifies specifications →

AWS Trainium3

Leads on FP8 Full specs →

What is the AWS Trainium3?

The Trainium3 is an ASIC from AWS on the NeuronCore-v4 architecture, an accelerator built for AI workloads. Trainium3 full specifications →

Can I rent the Trainium3 in the cloud?

Flopper is not tracking any live rental listings for the Trainium3 yet. The provider directory lists the clouds that rent comparable accelerators on demand. GPU cloud providers →

How does the Trainium3 compare with other accelerators?

This page sets the Trainium3 against the Trainium2. The Trainium3 is the newer generation of the two. Both are AWS parts. The compare tool builds the same comparison for any two parts in the catalogue. All GPU comparisons →

Where do the Trainium3 figures come from?

Every Trainium3 figure on this page traces to AWS's own datasheet or specification table. How Flopper verifies specifications →

Specifications

Trainium2

Trainium3

Architecture

NeuronCore-v3NeuronCore-v4

Launch Year

20242025

Form Factor

——

Memory

96 GB144 GB

Memory Bandwidth

2.9 TB/s4.9 TB/s

TDP

——

Process Node

—3nm

Spec

Trainium2

Trainium3

Architecture NeuronCore-v3NeuronCore-v4

Launch Year 20242025

Form Factor ——

Memory 96 GB144 GB

Memory Bandwidth 2.9 TB/s4.9 TB/s

TDP ——

Process Node —3nm

Performance (TFLOPS)

Trainium2

Trainium3

181 TFLOPS 183 TFLOPS

667 TFLOPS

2,563 TFLOPS sparse

671 TFLOPS

2,517 TFLOPS sparse

667 TFLOPS

2,563 TFLOPS sparse

671 TFLOPS

2,517 TFLOPS sparse

667 TFLOPS

2,563 TFLOPS sparse

671 TFLOPS

2,517 TFLOPS sparse

1,299 TFLOPS

2,563 TFLOPS sparse

2,517 TFLOPS

sparse not published

FP6

2,517 TFLOPS

sparse not published

Precision

Trainium2

Trainium3

181 TFLOPS 183 TFLOPS

667 TFLOPS

2,563 TFLOPS sparse

671 TFLOPS

2,517 TFLOPS sparse

667 TFLOPS

2,563 TFLOPS sparse

671 TFLOPS

2,517 TFLOPS sparse

667 TFLOPS

2,563 TFLOPS sparse

671 TFLOPS

2,517 TFLOPS sparse

1,299 TFLOPS

2,563 TFLOPS sparse

2,517 TFLOPS

sparse not published

FP6

2,517 TFLOPS

sparse not published

Figures are dense. Where a vendor publishes a figure it sits under the dense one and never decides a winner. Hover any precision for its definition, or read the GPU glossary.

FLOPS by Precision

What actually differs

The Trainium3 is the newer part: NeuronCore-v4 (2025) against the Trainium2's NeuronCore-v3 (2024).

The Trainium3 carries 48 GB more memory (144 GB against 96 GB), so larger models fit on one GPU before you shard. Bandwidth is 2.9 TB/s against 4.9 TB/s, which sets the ceiling on decode speed.

Weight the FP8 and FP16 rows for LLM work and the FP64 row for HPC.

New to the terms? The glossary explains FP32, TF32, BF16, FP16, FP8, FP4, Sparsity and TFLOPS.

Frequently asked questions

What is better, the Trainium2 or the Trainium3?

The Trainium3, on throughput. At FP8 precision the Trainium3 reaches 2,517 TFLOPS dense against 1,299 TFLOPS for the Trainium2. The performance table on this page lists every published precision for both GPUs.

Which should you choose, the Trainium2 or the Trainium3?

Choose the Trainium3 for training and high-throughput inference: it has 94% more FP8 throughput than the Trainium2. It also carries 48 GB more memory, so larger models fit on one GPU before you shard.

Which has more memory, the Trainium2 or the Trainium3?

The Trainium3 carries 144 GB of memory versus 96 GB for the Trainium2. Memory bandwidth is 2.9 TB/s for the Trainium2 and 4.9 TB/s for the Trainium3.

Get Comparison Updates

New GPUs added weekly. Be the first to see how they compare.

Open in Full Compare Tool → Trainium2 full specs Trainium3 full specs Estimate cluster cost in the calculator →

AWS Trainium2 vs AWS Trainium3

AWS Trainium2

What is the AWS Trainium2?

Can I rent the Trainium2 in the cloud?

How does the Trainium2 compare with other accelerators?

Where do the Trainium2 figures come from?

AWS Trainium3

What is the AWS Trainium3?

Can I rent the Trainium3 in the cloud?

How does the Trainium3 compare with other accelerators?

Where do the Trainium3 figures come from?

Specifications

Performance (TFLOPS)

FLOPS by Precision

What actually differs

Frequently asked questions

What is better, the Trainium2 or the Trainium3?

Which should you choose, the Trainium2 or the Trainium3?

Which has more memory, the Trainium2 or the Trainium3?

Related comparisons

Get Comparison Updates

Stay Updated on GPU Specs