The future of Amazon Web Services (AWS) will be built on custom silicon, AWS CEO Andy Jassy said during the AWS re:Invent 2020 virtual conference. AWS today unveiled new Graviton2-based instances alongside a partnership with Intel's Habana Labs and its own machine learning training chip called Trainium.

"We have a deep relationship with Intel and AMD, and we will for the foreseeable future," he said. "However we realized a few years ago if we wanted to continue to push the envelope of price to performance we knew we were going to have to develop some of our own chips."

Over the past few years, Amazon has aggressively developed custom silicon — often based on Arm architectures — in order to achieve superior price-to-performance ratios.

Amazon's Graviton2 chips, announced late this spring, are one such example. The chip is the company's second internally developed general-compute server chip. The Graviton2 is based on Arm's 64-bit Neoverse microarchitecture, which is built using a 7-nanometer manufacturing process, and integrates custom silicon designed by AWS for its role in the company’s cloud data centers.

According to Jassy, Graviton2-based Elastic Compute Cloud (EC2) instances offer 40% better price-to-performance over comparable x86-based instances. "Think about what you could do for your business if you have 40% better price performance," he said.

Amazon today announced plans to make a new high-throughput EC2 instance available for customers running compute and network heavy workloads. The C6GN instance is capable of 100 Gb/s of total network throughput.

"We are not close to being done investing and inventing with Graviton," Jassy said.

Turning to artificial intelligence (AI) and machine learning workloads training, Jassy stayed true to his word announcing a partnership with Intel.

Amazon will deploy the chipmaker's Habana Labs Guadi training processor in EC2. Intel acquired Habana Labs in December of last year in a deal valued at $2 billion, shortly before nerfing its Nevera AI processor project with Facebook.

"This is a partnership between AWS and Intel. It will use Gaudi accelerators that will provide 40% better price performance and the best performing GPU instances that we have today," said Jassy. "It will work with all the main machine learning frameworks — PyTorch, TensorFlow — and it'll be available in the first half of [2021]."

However, it appears Amazon's AI training partnership with Intel may be a short-term stopgap while the company prepares to bring its own machine learning chip to market. The new chip announced today, dubbed Trainium, is designed from the ground up for machine learning training.

"Trainium will be even more cost effective than the Habana chip," Jassy said.

When it debuts in the second half of 2021, it will be available as an EC2 instance and in Amazon SageMaker, and use the same Neuron software development kit (SDK) as Amazon's existing Inferentia inference chips announced last year.