Double Bandwidth, 1.5× Memory: Boost AI Workloads with EC2 P6‑B300

Amazon EC2 P6‑B300, the latest generation instance driven by NVIDIA Blackwell Ultra GPUs, is now available in the US West (Oregon) region. Compared with the previous generation, it provides a 2× increase in network bandwidth and a 1.5× increase in GPU memory capacity.

The instance delivers 6.4 Tbps of Elastic Fabric Adapter (EFA) bandwidth and equips each node with 2.1 TB of GPU memory, allowing very large models to reside within a single NVIDIA NVLink domain and substantially reducing the need for model sharding and associated communication overhead.

Combined with the advanced virtualization and security features of the Amazon Nitro system, the P6‑B300 offers unprecedented speed, scale, and security for AI workloads, making it well‑suited for complex techniques such as mixture‑of‑experts (MoE) and multimodal processing that require handling trillion‑parameter models across thousands of GPUs.

For persistent storage, the instance can leverage multiple high‑performance options, including Amazon FSx for Lustre, Amazon S3 Express One Zone, and Amazon EBS. When paired with FSx for Lustre and NVIDIA GPUDirect Storage, the P6‑B300 can achieve up to 1.2 Tbps of Lustre file system throughput, enabling rapid model loading.

Using the instance’s dedicated 300 Gbps ENA, workloads can also benefit from high‑throughput hot storage access via Amazon S3 Express One Zone, supporting large‑scale training jobs.

Pricing follows the standard on‑demand model used by other EC2 instances, with options to purchase through EC2 Capacity Blocks for ML or Savings Plans. More details are available on the Amazon EC2 P6‑B300 product page.