AWS launches all-new GPU-based instances for ML training and HPC

Key takeaways

AWS has introduced P4de GPU-based instances for enhanced machine learning training and high-performance computing applications.

The P4de instances feature 8 NVIDIA A100 GPUs, offering significantly improved performance and lower training costs.

Increased GPU memory allows for efficient training on large datasets, reducing model training times.

Currently, P4de instances are available in the AWS US East and US West regions with extensive system capabilities.

These instances are part of EC2 UltraClusters, aimed at providing advanced networking and storage solutions for HPC and ML.

AWS today announced that it would be allowing the preview of GPU-based instances, Amazon EC2 P4de instances. The GPU-based instances provide high performance for machine learning training (ML) and high-performance computing (HPC) applications. This includes object detection, NLP, semantic segmentation, recommender systems, seismic analysis, etc.

Inside P4de instances

Powered by 8 NVIDIA A100 GPUs with 80 GB 'high-performance' HBM2e GPU memory, the P4de instances are 2x higher than current GPUs. The new P4de instances provide 640 GB of GPU memory, which provides up to 60 per cent better ML training performance and 20 per cent lower cost to train compared to the current P4d instances.

The benefits

The company said that the improved performance would allow customers to reduce model training times and accelerate time to market faster. Also, increased GPU memory will benefit workloads that need to train on large datasets of high-resolution data.

Where is it available?

Currently, P4de instances are available in the AWS US East and US West regions. It is available in the p4de.24xl size, providing 96 vCPUs, 8 NVIDIA A100-80 GB GPUs, 1.1 TB system memory, 8 TB local NVMe-based SSD storage, 19 Gbps EBS bandwidth, and 400 Gbps networking bandwidth with EFA and GPUDirect RDMA.

P4de instances are deployed in EC2 UltraClusters, which provides petabit-scale non-blocking networking infrastructure and high-throughput, low-latency storage through FSx for scale-out HPC and ML training applications. EC2 UltraClusters is one of the most powerful supercomputers in the world. It combines high-performance computing, networking, and storage.

(Source: AWS)

What to expect?

Last year, AWS announced three new Amazon EC2 instances powered by AWS-designed chips at the AWS re:Invent. These chips were said to help customers significantly improve the performance, cost, and energy efficiency of their workloads running on Amazon EC2 instances.

With the new P4de instances, the team believes that it will continue to add to the industry's widest portfolio of accelerated compute instances, featuring platforms powered by their silicon and by accelerators from their partners, to provide the highest performing NVIDIA GPUs for customers to build, train, and deploy machine learning at scale.