Amazon Web Services (AWS) announced on the 15th that it has launched the new GPU-based P6e-GB200 ultra server for AI and high-performance computing (HPC) workloads.
The P6e-GB200 ultra server, powered by the NVIDIA Grace Blackwell superchip, supports training and deployment of very large and complex AI models, providing up to 72 NVIDIA Blackwell GPUs and 360 petaflops of FP8 high-density computing. This server offers more than 20 times the computing performance and more than 11 times the memory compared to the existing P5en instances, and it supports bandwidth of up to 28.8 Tbps through the fourth-generation Elastic Fabric Adapter (EFAv4).
The P6-B200 instance offers 8 GPUs and 1.4 TB of high-bandwidth GPU memory, representing an improvement of more than 2 times in GPU performance and memory size compared to the P5en instance. The P6e-GB200 is optimized for AI model training and inference at trillions of parameters in scale and can fully handle models with one trillion parameters within a single NVLink domain.
AWS emphasizes security and stability for these instances, enhancing the security and system management efficiency of the instances through the AWS Nitro system. The P6e-GB200 and P6-B200 instances offer high performance and efficiency through a liquid cooling system and demonstrate optimized performance in AWS's third-generation EC2 ultra-cluster environment.
AWS provides consistent performance and reliability in large-scale environments through the P6e-GB200, showing up to a 18% performance improvement in distributed training via EFAv4. These instances are suitable for various AI and HPC workloads, offering a safe and efficient computing environment in AWS's cloud infrastructure.