New Amazon EC2 P6e-GB200 UltraServers accelerated by NVIDIA Grace Blackwell GPUs for the very best AI efficiency

Today, we’re asserting the final availability of Amazon Elastic Compute Cloud (Amazon EC2) P6e-GB200 UltraServers, accelerated by NVIDIA GB200 NVL72 to supply the very best GPU efficiency for AI coaching and inference. Amazon EC2 UltraServers join a number of EC2 cases utilizing a devoted, high-bandwidth, and low-latency accelerator interconnect throughout these cases.

The NVIDIA Grace Blackwell Superchips join two high-performance NVIDIA Blackwell tensor core GPUs and an NVIDIA Grace CPU based mostly on Arm structure utilizing the NVIDIA NVLink-C2C interconnect. Each Grace Blackwell Superchip delivers 10 petaflops of FP8 compute (with out sparsity) and as much as 372 GB HBM3e reminiscence. With the superchip structure, GPU and CPU are colocated inside one compute module, growing bandwidth between GPU and CPU considerably in comparison with present era EC2 P5en cases.

With EC2 P6e-GB200 UltraServers, you’ll be able to entry as much as 72 NVIDIA Blackwell GPUs inside one NVLink area to make use of 360 petaflops of FP8 compute (with out sparsity) and 13.4 TB of whole excessive bandwidth reminiscence (HBM3e). Powered by the AWS Nitro System, P6e-GB200 UltraServers are deployed in EC2 UltraClusters to securely and reliably scale to tens of hundreds of GPUs.

EC2 P6e-GB200 UltraServers ship as much as 28.8 Tbps of whole Elastic Fabric Adapter (EFAv4) networking. EFA can also be coupled with NVIDIA GPUDirect RDMA to allow low-latency GPU-to-GPU communication between servers with working system bypass.

EC2 P6e-GB200 UltraServers specs
EC2 P6e-GB200 UltraServers can be found in sizes starting from 36 to 72 GPUs underneath NVLink. Here are the specs for EC2 P6e-GB200 UltraServers:

P6e-GB200 UltraServers are perfect for probably the most compute and reminiscence intensive AI workloads, akin to coaching and inference of frontier fashions, together with combination of consultants fashions and reasoning fashions, on the trillion-parameter scale.

You can construct agentic and generative AI purposes, together with query answering, code era, video and picture era, speech recognition, and extra.

P6e-GB200 UltraServers in motion
You can use EC2 P6e-GB200 UltraServers within the Dallas Local Zone by EC2 Capacity Blocks for ML. The Dallas Local Zone (us-east-1-dfw-2a) is an extension of the US East (N. Virginia) Region.

To reserve your EC2 Capacity Blocks, select Capacity Reservations on the Amazon EC2 console. You can choose Purchase Capacity Blocks for ML after which select your whole capability and specify how lengthy you want the EC2 Capacity Block for u-p6e-gb200x36 or u-p6e-gb200x72 UltraServers.

Once Capacity Block is efficiently scheduled, it’s charged up entrance and its value doesn’t change after buy. The fee will probably be billed to your account inside 12 hours after you buy the EC2 Capacity Blocks. To study extra, go to Capacity Blocks for ML within the Amazon EC2 User Guide.

To run cases inside your bought Capacity Block, you should use AWS Management Console, AWS Command Line Interface (AWS CLI) or AWS SDKs. On the software program facet, you can begin with the AWS Deep Learning AMIs. These pictures are preconfigured with the frameworks and instruments that you simply in all probability already know and use: PyTorch, JAX, and much more.

You may combine EC2 P6e-GB200 UltraServers seamlessly with varied AWS managed companies. For instance:

• Amazon SageMaker Hyperpod offers managed, resilient infrastructure that mechanically handles the provisioning and administration of P6e-GB200 UltraServers, changing defective cases with preconfigured spare capability throughout the similar NVLink area to keep up efficiency.
• Amazon Elastic Kubernetes Services (Amazon EKS) permits one managed node group to span throughout a number of P6e-GB200 UltraServers as nodes, automating their provisioning and lifecycle administration inside Kubernetes clusters. You can use EKS topology-aware routing for P6e-GB200 UltraServers, enabling optimum placement of tightly coupled elements of distributed workloads inside a single UltraServer’s NVLink-connected cases.
• Amazon FSx for Lustre file programs present information entry for P6e-GB200 UltraServers on the lots of of GB/s of throughput and thousands and thousands of enter/output operations per second (IOPS) required for large-scale HPC and AI workloads. For quick entry to massive datasets, you should use as much as 405 TB of native NVMe SSD storage or nearly limitless cost-effective storage with Amazon Simple Storage Service (Amazon S3).

Now out there
Amazon EC2 P6e-GB200 UltraServers can be found at present within the Dallas Local Zone (us-east-1-dfw-2a) by EC2 Capacity Blocks for ML. For extra info, go to the Amazon EC2 pricing web page.

Give Amazon EC2 P6e-GB200 UltraServers a attempt within the Amazon EC2 console. To study extra, go to the Amazon EC2 P6e cases web page and ship suggestions to AWS re:Post for EC2 or by your standard AWS Support contacts.

— Channy