Amazon's most powerful CPU chip to date. Plus flat networks, agentic AI, and formally verified VM isolation
Amazon Science
The latest news and research from Amazon’s science community. #AmazonScience
Find the latest news and research from Amazon's science community at Amazon Science.
Graviton5's improved design increases speed and energy efficiency — beyond Moore's law: Graviton5 doubles the core count from Graviton4 (96 to 192) using a four-chiplet architecture with custom die-to-die connectivity delivering up to 420 GB/s bandwidth, while moving to a three-nanometer process and adding over five times the L3 cache. The result: 25% better computational performance overall, up to 35% faster for web applications and ML inference, and up to 30% for databases.
Bridging intent and execution in agentic systems: AWS researchers introduce Simple Strands Agent (SSA), an open-source harness that achieves state-of-the-art performance across SWE-Bench-Verified, SWE-Bench-Pro, and Terminal-Bench2 by identifying invariant design principles — improved tool interfaces, diff-based feedback, and a balance between reasoning and environment interaction — that transfer across model families without any task-specific tuning.
How flat is replacing fat in AWS data center networks: AWS's new RNG (resilient network graphs) architecture — now the default for most new data centers globally — solves the longstanding challenge of making random network topologies practical, using a passive optical device called a ShuffleBox and a routing protocol called Spraypoint. The results: 69% fewer routers, up to 33% better throughput, and a projected 40% reduction in network equipment electricity consumption.
Fall 2025 Amazon Research Awards recipients announced: Awardees represent 49 universities across 11 countries, spanning six research areas including Agentic AI, Automated Reasoning, Cryptography, AI for Information Security, Cybersecurity and Anti-Abuse Technologies, and Sustainability. Recipients receive unrestricted funds, AWS Promotional Credits, and access to more than 700 Amazon public datasets.
How the next generation of AI researchers is using Amazon chips to accelerate discovery: Through Build on Trainium, a $110 million program, university researchers at UC Berkeley, MIT, Carnegie Mellon, UCLA, and others get direct access to Trainium chips and Amazon's AI experts. Early results include a 50% throughput improvement in medical imaging model training at MIT and a FlashAttention optimization breakthrough at Carnegie Mellon in just one week — with all research released as open source.
Making a Mind Season 2 coming soon: Hosted by Danielle Perszyk, Amazon cognitive scientist, the podcast features researchers tackling the hardest problems in agentic AI – from building reliable perception systems to designing training environments that mirror human learning. New episodes dropping soon, with notable guests from across the AI research community.
Recommended by LinkedIn
© 1996-2026 Amazon.com, Inc. or its affiliates | Privacy | Conditions of Use