AWS announces the fourth generation of the Graviton chip and the second generation of Trainium during re:Invent. There is no shortage of cores and memory.

Graviton made its debut in 2018. At that time, Amazon broke new ground by developing an in-house chip that was fully optimized for cloud services. After all, hardware forms the basis of the cloud infrastructure. The second generation followed in 2019 and the third in 2021, which was supplemented with Graviton 3E. After three comes four and so AWS announces Graviton4 during re:Invent.

Gravition4 is the most powerful chip from Amazon’s laboratory to date. The chip features no less than 96 Arm Neoverse VS cores with 2MB of L2 cache in each core. Combined with twelve DDR5-5600 channels, Graviton4 can deliver up to forty percent faster performance for database applications than its predecessor, Amazon said in a press release. For web and Java applications, the performance increases are thirty and 45 percent.

Safety was also taken into account. Graviton4 inherits all security standards of previous generations and adds two new technologies: encryption of high-speed hardware interfaces and Branch target identification, developed by Arm. Graviton4 is available via optimized Rg8 instances on AWS EC2, supporting up to three times more vCPU capacity and three times more memory.

Thirsty for AI

That’s not the only chip news AWS has to share. No announcement is complete without bridging AI somewhere, and so AWS also announces Trainium2. As the name suggests, this is a chip for training LLMs, which is why it is also called an “AI Accelerator”. Now that Microsoft also has AI chips for Azure, AWS thought it was high time to move forward with Trainium.

Technical details about the AI ​​chip remain rather sparse, both in the official announcement and during AWS CEO Adam Selipsky’s keynote speech. What AWS says is that Trainium2 can deliver four times faster training performance than the first Trainium chip and does so twice as efficiently. The storage capacity has been tripled from 32 GB to 96 GB.

In addition to its own chips, AWS has other tricks up its sleeve to attract companies to its cloud for training LLMs. AWS works closely with Nvidia and this was clearly highlighted at re:Invent. The cloud provider is the first to offer Nvidia’s new GH200 GPUs and also launched new microservices with the support of the GPU specialist.