AI Chips · Product page

Amazon Trainium (Trn1): The Complete Guide

The chip that took Amazon from renting other companies’ AI chips to training models on its own silicon. Simple to expert, every number links to its source.

Stuck at any point? Ask an AI about this page →
Ask an AI about this chip:CAGPX
Amazon Trainium 10 sections · 2 levels 7 linked sources Checked September 2026
How to read this page. Each section starts Simple, then goes to a Deep dive: stop wherever you have what you need. The small numbers are sources: click one to open the original document. Where the maker has not published a figure -- a price, a die size, a factory address -- this page says so rather than estimate it.

1.At a glance

SimpleStart here

Trainium — later called Trn1 once instances shipped — is Amazon’s first chip built specifically to train AI models, not just run them. AWS announced it 1 December 2020 1; the instances that use it went on general sale 10 October 2022 2.

Deep diveThe technical detail

Amazon’s own claim at general availability: Trn1 instances offer “up to 50% lower cost to train deep learning models” than comparable GPU-based EC2 instances 2. Nearly two years passed between announcement and general availability — longer than the roughly one-year cadence later Trainium generations have kept (see the Trainium2 and Trainium3 pages on this site).

2.Launch and history

SimpleStart here

Before Trainium, Amazon’s only AI chip, Inferentia, could run models but not train them — for training, AWS customers rented GPUs from NVIDIA. Trainium gave Amazon its own training silicon for the first time.

Deep diveThe technical detail

Trainium was designed by Annapurna Labs 3, the same team behind Inferentia (see that page on this site). AWS’s own architecture documentation groups the original Trainium together with Inferentia2, describing them as sharing the same underlying NeuronCore design even though one trains models and the other only runs them 4.

3.What’s inside it

SimpleStart here

Each Trainium chip is built around NeuronCores, sharing the same core design AWS later used in Inferentia2 — a large matrix-multiply engine alongside smaller engines for other math 4.

Deep diveThe technical detail

AWS’s architecture documentation describes a NeuronCore with four engines: a tensor engine for matrix math, a vector engine, a scalar engine, and a programmable “GpSimd” engine of eight processors for custom code 4. Each core has its own software-managed on-chip memory — 24 MiB on this first generation, per AWS’s own documentation 4.

4.Full spec table

SimpleStart here

3.4 PFLOPS of BF16 compute per 16-chip server, 32 GiB of HBM memory per chip at 820 GB/s — AWS’s own published figures.

Deep diveThe technical detail
SpecAmazon Trainium (Trn1)
AI compute3.4 PFLOPS BF16 per 16-chip server 5
Memory32 GiB HBM per chip 4
Memory bandwidth820 GB/s 4
On-chip memory (per core)24 MiB 4
Process nodeNot published by AWS

Server-level compute figure is AWS’s own, for a full 16-chip trn1.32xlarge instance, not a single-chip figure 5.

5.Where it’s made

SimpleStart here

Designed by Annapurna Labs. AWS has not named a foundry for the original Trainium 3.

Deep diveThe technical detail

Press reports naming TSMC as Amazon’s foundry partner (cited on the Trainium2 and Trainium3 pages on this site) refer specifically to those later chips. This page found no equivalent report naming a foundry for the original Trainium — treated here as genuinely unreported, in the same way as Inferentia and Inferentia2’s foundries (see those pages on this site).

6.Which systems use it

SimpleStart here

Rented through AWS as EC2 Trn1 instances, up to 16 chips per server, or booked in advance through AWS Capacity Blocks 6 7.

Deep diveThe technical detail

AWS makes Trn1 Capacity Blocks available in the US and in Mumbai, India, among other regions, at $9.532 per hour for a full 16-chip trn1.32xlarge — $0.596 per chip 7. This is the earliest Amazon training chip with confirmed India availability of any kind on this site; Trainium2 and Trainium3 have no equivalent published India listing found (see those pages).

7.Official pricing

SimpleStart here

Rented by the hour, not sold. AWS’s current published on-demand prices for Trn1 instances 6:

InstanceOn-demand1-yr reserved3-yr reserved
trn1.2xlarge (1 chip)$1.34/hr$0.79/hr$0.4744/hr
trn1.32xlarge (16 chips)$21.50/hr$12.60/hr$7.59/hr
Deep diveThe technical detail

Capacity Blocks (booked in advance rather than on demand) offer a lower rate: $9.532 per hour for a full trn1.32xlarge (16 chips), or $0.596 per chip-hour, available in the US, Mumbai and other regions 7. AWS says Capacity Block prices are next updated in October 2026 7.

8.Real-world performance

SimpleStart here

AWS’s own claim: “up to 50% lower cost to train deep learning models” on Trn1 than comparable GPU-based EC2 instances 2.

Deep diveThe technical detail

This is AWS’s own comparison, not an independently verified benchmark — no named third-party test against a specific GPU model, on a public benchmark, was found for the original Trainium specifically. Treat the 50% figure as AWS’s own claim, in the same way this site treats every vendor’s self-reported comparison elsewhere.

9.What came before, what came next

SimpleStart here
Came before: nothing — Trainium was Amazon’s first training-focused chip. Came after: Trainium2, covered on its own page on this site.
Deep diveThe technical detail

The gap between Trainium’s December 2020 announcement and its October 2022 general availability — nearly two years — is longer than the roughly annual cadence Amazon has kept since: Trainium2 was announced November 2023 and shipped December 2024; Trainium3 was already on sale by December 2025 (see those pages on this site). Amazon has not explained the longer gap for this first generation.

10.Hidden in plain sight

SimpleStart here

Amazon’s first training chip took almost two years from announcement to actually shipping — a gap every generation since has closed to about a year.

Deep diveThe technical detail

Trainium was announced 1 December 2020 1 but Trn1 instances did not reach general availability until 10 October 2022 2 — 22 months. By contrast, Trainium2 went from announcement (November 2023) to shipping (December 2024) in about 13 months, and Trainium3 was already generally available within roughly two years of Trainium2’s own shipping date. Amazon has not published a reason for the longer first-generation gap.

11.Sources

7 sources, checked September 2026. Where NVIDIA or another maker has not published a figure, this page says so rather than estimate it.

  1. TechCrunch (press): AWS launches Trainium, its new custom ML training chipPress
  2. Amazon Press Center: AWS announces general availability of EC2 Trn1 instancesOfficial
  3. Amazon Science: How silicon innovation became the secret sauce behind AWS’s successOfficial
  4. AWS Neuron documentation: Trainium and Inferentia2 architectureOfficial
  5. AWS News Blog: Amazon EC2 Trn1 instances are now availableOfficial
  6. AWS: Amazon EC2 Trn1 instances pageOfficial
  7. AWS: Amazon EC2 Capacity Blocks for ML pricingOfficial

Ask an AI about this page

Opens your assistant with this page as the source, and a question rather than a summary. It will ask what you are building before it answers.

ChatGPTClaudeGeminiPerplexityGrok

Nothing is sent from here. The link carries only this page’s title and address.