NVIDIA A100 TENSOR CORE GPU

NVIDIA A100 TENSOR CORE GPU

DESCRIPTION

The NVIDIA A100 Tensor Core GPU delivers unprecedented  acceleration—at every scale—to power the world’s highest performing elastic data centers for AI, data analytics, and high performance computing (HPC) applications. As the engine of  the NVIDIA data center platform, A100 provides up to 20X higher  performance over the prior NVIDIA Volta™ generation. A100 can  efficiently scale up or be partitioned into seven isolated GPU  instances with Multi-Instance GPU (MIG), providing a unified  platform that enables elastic data centers to dynamically adjust  to shifting workload demands.  

NVIDIA A100 Tensor Core technology supports a broad range  of math precisions, providing a single accelerator for every  workload. The latest generation A100 80GB doubles GPU memory  and debuts the world’s fastest memory bandwidth at 2 terabytes  per second (TB/s), speeding time to solution for the largest  models and most massive datasets.

Groundbreaking Innovations

NVIDIA AMPERE  ARCHITECTURE

Whether using MIG to partition an  A100 GPU into smaller instances  or NVLink to connect multiple  GPUs to speed large-scale workloads, A100 can  readily handle different-sized acceleration needs,  from the smallest job to the biggest multi-node  workload. A100’s versatility means IT managers  can maximize the utility of every GPU in their data  center, around the clock.  

THIRD-GENERATION  TENSOR CORES

NVIDIA A100 delivers 312  teraFLOPS (TFLOPS) of deep  learning performance. That’s 20X  the Tensor floating-point operations per second  (FLOPS) for deep learning training and 20X the  Tensor tera operations per second (TOPS) for  deep learning inference compared to NVIDIA  Volta GPUs.

NEXT-GENERATION NVLINK

NVIDIA NVLink in A100 delivers  2X higher throughput compared  to the previous generation. When  combined with NVIDIA NVSwitch™,  up to 16 A100 GPUs can be interconnected at up  to 600 gigabytes per second (GB/sec), unleashing  the highest application performance possible on  a single server. NVLink is available in A100 SXM  GPUs via HGX A100 server boards and in PCIe  GPUs via an NVLink Bridge for up to 2 GPUs.

MULTI-INSTANCE GPU (MIG)

An A100 GPU can be partitioned  into as many as seven GPU  instances, fully isolated at  the hardware level with their  own high-bandwidth memory, cache, and  compute cores. MIG gives developers access  to breakthrough acceleration for all their  applications, and IT administrators can offer  right-sized GPU acceleration for every job,  optimizing utilization and expanding access to  every user and application.

HIGH-BANDWIDTH MEMORY  (HBM2E)

With up to 80 gigabytes of  HBM2e, A100 delivers the world’s  fastest GPU memory bandwidth  of over 2TB/s, as well as a dynamic randomaccess memory (DRAM) utilization efficiency  of 95%. A100 delivers 1.7X higher memory  bandwidth over the previous generation.

SPECIFICATION

FP64: 9.7 TFLOPS

FP64 Tensor  Core:19.5 TFLOPS

FP32:19.5 TFLOPS

Tensor Float  32 (TF32):156 TFLOPS | 312 TFLOPS

BFLOAT16  Tensor Core:312 TFLOPS | 624 TFLOPS

FP16 Tensor  Core:312 TFLOPS | 624 TFLOPS