Amazon's AI initiative in the US is centered in Texas
Amazon is working to reduce its dependence on Nvidia by
building its own AI chips called Trainium. These chips are specially designed
for machine learning and artificial intelligence (AI).
Amazon’s cloud division, Amazon Web Services (AWS), is
developing the chips through its subsidiary Annapurna Labs in Austin, Texas.
Why Amazon Is Building Its Own Chips
For many years, Amazon relied on outside suppliers for
chips. In 2015, it bought Annapurna Labs to start designing its own processors.
It first launched:
- Graviton
chips (for cloud computing)
- Inferentia
chips (for running AI models)
The first Trainium chip came in 2020, followed by a
faster second version.
Trainium 3: Faster and Cheaper
The newest version, Trainium 3, was launched in December.
- It
is smaller than a credit card.
- It
is twice as powerful as the previous version.
- It
may reduce the cost of building and running generative AI models by up to 40%
compared to GPUs (graphics processing units), which are currently the
industry standard.
Amazon says reliability is very important. AI training
requires thousands of chips running continuously for weeks. If even one chip
fails, the process may need to restart.
How Amazon Uses Trainium
Unlike Nvidia, Amazon does not sell its chips.
Instead, AWS uses Trainium chips in its own data centres
and rents computing power to customers. The chips are designed to work smoothly
with AWS software, especially its Bedrock platform, which allows users to
choose AI models from companies like Anthropic and OpenAI.
Competing in the AI Race
The AI market currently faces a shortage of
high-performance GPUs because demand is very high. Nvidia and AMD dominate this
space.
Amazon is positioning Trainium as a cost-effective
alternative.
Even though Trainium 3 is new, Annapurna is already
working on Trainium 4, which is expected to be six times more powerful than
Trainium 3.
As companies like Google, Microsoft, OpenAI, and Meta
compete to build better AI models, chipmakers are under pressure to make chips
faster, smarter, cheaper, and more energy-efficient.
