Accelerate large-scale AI applications with the new Amazon EC2 P6-B300 instances
Today, we’re announcing the general availability of Amazon Elastic Compute Cloud (Amazon EC2) P6-B300 instances, our next-generation GPU platform accelerated by NVIDIA Blackwell Ultra GPUs. These instances deliver 2 times more networking bandwidth, and 1.5 times more GPU memory compared to previous generation instances, creating a balanced platform for large-scale AI applications.
With these improvements, P6-B300 instances are ideal for training and serving large-scale AI models, particularly those employing sophisticated techniques such as Mixture of Experts (MoE) and multimodal processing. For organizations working with trillion-parameter models and requiring distributed training across thousands of GPUs, these instances provide the perfect balance of compute, memory, and networking capabilities.
Improvements made compared to predecessors
The P6-B300 instances deliver 6.4Tbps Elastic Fabric Adapter (EFA) networking bandwidth, supporting efficient communication across large GPU clusters. These instances feature 2.1TB of GPU memory, allowing large models to reside within a single NVIDIA NVLink domain, which significantly reduces model sharding and communication overhead. When combined with EFA networking and the advanced virtualization and security capabilities of AWS Nitro System, these instances provide unprecedented speed, scale, and security for AI workloads.
The specs for the EC2 P6-B300 instances are as follows.
| Instance size | VCPUs | System memory | GPUs | GPU memory | GPU-GPU interconnect | EFA network bandwidth | ENA bandwidth | EBS bandwidth | Local storage |
| P6-B300.48xlarge | 192 | 4TB | 8x NVIDIA B300 GPU | 2144GB HBM3e | 1800 GB/s | 6.4 Tbps | 300 Gbps | 100 Gbps | 8x 3.84TB |
Good to know
In terms of persistent storage, AI workloads primarily use a combination of high performance persistent storage options such as Amazon FSx for Lustre, Amazon S3 Express One Zone, and Amazon Elastic Block Store (Amazon EBS), depending on price performance considerations. For illustration, the dedicated 300Gbps Elastic Network Adapter (ENA) networking on P6-B300 enables high-throughput hot storage access with S3 Express One Zone, supporting large-scale training workloads. If you’re using FSx for Lustre, you can now use EFA with NVIDIA GPUDirect Storage (GDS) to achieve up to 1.2Tbps of throughput to the Lustre file system on the P6-B300 instances to quickly load your models.
Available now
The P6-B300 instances are now available through Amazon EC2 Capacity Blocks for ML and Savings Plans in the US West (Oregon) AWS Region.
For on-demand reservation of P6-B300 instances, please reach out to your account manager. As usual with Amazon EC2, you pay only for what you use. For more information, refer to Amazon EC2 Pricing. Check out the full collection of accelerated computing instances to help you start migrating your applications.
To learn more, visit our Amazon EC2 P6-B300 instances page. Send feedback to AWS re:Post for EC2 or through your usual AWS Support contacts.
– Veliswa
Source: AWS News