AWS Tunes Up Graviton5 for Agentic AI, Boosts Bang for the Buck Bigtime (www.nextplatform.com)

AWS has officially launched its Graviton5 Arm server CPU in the new M9g and M9gd instances, enhancing performance capabilities specifically for agentic AI applications. Initially previewed at re:Invent 2023, the Graviton5 features a new architecture with four chiplets, each housing 48 "Poseidon" Neoverse V3 cores, totaling 192 cores. This departure from the original one-piece design allows for increased manufacturing yields and lower costs, leveraging a 3-nanometer fabrication process that enhances transistor density and power efficiency. With significant upgrades in cache sizes—2MB of L2 and 384MB of L3 per core—and advanced interconnects capable of 420 GB/s, the chip delivers up to 2.4 times the performance per socket compared to its predecessor, Graviton4. The significance of this launch lies in its ability to substantially reduce costs while improving performance, making it highly attractive for businesses leveraging AI and database operations requiring low-latency processing. With M9g instances priced approximately half of the previous generation X8g instances, AWS is responding to the current DRAM crunch by prioritizing bandwidth over memory capacity, while also hinting at future heavy-memory X9g instances. The integration of CXL 3.0 memory extension technology is expected to further optimize performance for data-intensive workloads, positioning AWS as a competitive force in the cloud computing landscape for AI applications.