Reading time ( words)
Amazon Web Services’ first GPU instance debuted 10 years ago, with the NVIDIA M2050. At that time, CUDA-based applications were focused primarily on accelerating scientific simulations, with the rise of AI and deep learning still a ways off.
Since then, AWS has added to its stable of cloud GPU instances, which has included the K80 (p2), K520 (g3), M60 (g4), V100 (p3/p3dn) and T4 (g4).
With its new P4d instance generally available today, AWS is paving the way for another bold decade of accelerated computing powered with the latest NVIDIA A100 Tensor Core GPU.
The P4d instance delivers AWS’s highest performance, most cost-effective GPU-based platform for machine learning training and high performance computing applications. The instances reduce the time to train machine learning models by up to 3x with FP16 and up to 6x with TF32 compared to the default FP32 precision.
They also provide exceptional inference performance. NVIDIA A100 GPUs just last month swept the MLPerf Inference benchmarks — providing up to 237x faster performance than CPUs.
Each P4d instance features eight NVIDIA A100 GPUs and, with AWS UltraClusters, customers can get on-demand and scalable access to over 4,000 GPUs at a time using AWS’s Elastic Fabric Adaptor (EFA) and scalable, high-performant storage with Amazon FSx. P4d offers 400Gbps networking and uses NVIDIA technologies such as NVLink, NVSwitch, NCCL and GPUDirect RDMA to further accelerate deep learning training workloads. NVIDIA GPUDirect RDMA on EFA ensures low-latency networking by passing data from GPU to GPU between servers without having to pass through the CPU and system memory.
In addition, the P4d instance is supported in many AWS services, including Amazon Elastic Container Services, Amazon Elastic Kubernetes Service, AWS ParallelCluster and Amazon SageMaker. P4d can also leverage all the optimized, containerized software available from NGC, including HPC applications, AI frameworks, pre-trained models, Helm charts and inference software like TensorRT and Triton Inference Server.
P4d instances are now available in US East and West, and coming to additional regions soon. The instances can be purchased as On-Demand, with Savings Plans, with Reserved Instances, or as Spot Instances.
The first decade of GPU cloud computing has brought over 100 exaflops of AI compute to the market. With the arrival of the Amazon EC2 P4d instance powered by NVIDIA A100 GPUs, the next decade of GPU cloud computing is off to a great start.