NVIDIA A100 Marks Daybreak of Subsequent Decade in Accelerated Cloud Computing
Amazon Net Companies’ first GPU occasion debuted 10 years in the past, with the NVIDIA M2050. At the moment, CUDA-based purposes have been centered totally on accelerating scientific simulations, with the rise of AI and deep studying nonetheless a methods off.
Since then, AWS has added to its steady of cloud GPU cases, which has included the K80 (p2), K520 (g3), M60 (g4), V100 (p3/p3dn) and T4 (g4).
With its new P4d occasion usually accessible right this moment, AWS is paving the way in which for one more daring decade of accelerated computing powered with the newest NVIDIA A100 Tensor Core GPU.
The P4d occasion delivers AWS’s highest efficiency, most cost-effective GPU-based platform for machine studying coaching and excessive efficiency computing purposes. The cases scale back the time to coach machine studying fashions by as much as 3x with FP16 and as much as 6x with TF32 in comparison with the default FP32 precision.
In addition they present distinctive inference efficiency. NVIDIA A100 GPUs simply final month swept the MLPerf Inference benchmarks — offering as much as 237x sooner efficiency than CPUs.
Every P4d occasion options eight NVIDIA A100 GPUs and, with AWS UltraClusters, prospects can get on-demand and scalable entry to over 4,000 GPUs at a time utilizing AWS’s Elastic Cloth Adaptor (EFA) and scalable, high-performant storage with Amazon FSx. P4d presents 400Gbps networking and makes use of NVIDIA applied sciences equivalent to NVLink, NVSwitch, NCCL and GPUDirect RDMA to additional speed up deep studying coaching workloads. NVIDIA GPUDirect RDMA on EFA ensures low-latency networking by passing knowledge from GPU to GPU between servers with out having to cross by the CPU and system reminiscence.
As well as, the P4d occasion is supported in lots of AWS providers, together with Amazon Elastic Container Companies, Amazon Elastic Kubernetes Service, AWS ParallelCluster and Amazon SageMaker. P4d also can leverage all of the optimized, containerized software program accessible from NGC, together with HPC purposes, AI frameworks, pre-trained fashions, Helm charts and inference software program like TensorRT and Triton Inference Server.
P4d cases are actually accessible in US East and West, and coming to extra areas quickly. The cases might be bought as On-Demand, with Financial savings Plans, with Reserved Situations, or as Spot Situations.
The primary decade of GPU cloud computing has introduced over 100 exaflops of AI compute to the market. With the arrival of the Amazon EC2 P4d occasion powered by NVIDIA A100 GPUs, the following decade of GPU cloud computing is off to a terrific begin.