Data Center Class Reliability
Designed for 24 x 7 data center operations and driven by power-efficient hardware and components selected for optimum performance, durability, and longevity. Every NVIDIA A100 board is designed, built and tested by NVIDIA to the most rigorous quality and performance standards, ensuring that leading OEMs and systems integrators can meet or exceed the most demanding real-world conditions.
Secure and measured boot with hardware root of trust technology within the GPU provides an additional layer of security for data centers. A100 meets the latest data center standards and is NEBS Level 3 compliant. The NVIDIA A100 includes a CEC 1712 security chip that enables secure and measured boot with hardware root of trust, ensuring that firmware has not been tampered with or corrupted.
NVIDIA Ampere Architecture
NVIDIA A100 is the world's most powerful data center GPU for AI, data analytics, and high-performance computing (HPC) applications. Building upon the major SM enhancements from the Turing GPU, the NVIDIA Ampere architecture enhances tensor matrix operations and concurrent executions of FP32 and INT32 operations.
More Efficient CUDA Cores
The NVIDIA Ampere architecture’s CUDA cores bring up to 2.5x the single-precision floating point (FP32) throughput compared to the previous generation, providing significant performance improvements for any class or algorithm or application that can benefit from embarrassingly parallel acceleration techniques.
Third Generation Tensor Cores
Purpose-built for deep learning matrix arithmetic at the heart of neural network training and inferencing functions, the NVIDIA A100 includes enhanced Tensor Cores that accelerate more datatypes (TF32 and BF16) and includes a new Fine-Grained Structured Sparsity feature that delivers up to 2x throughput for tensor matrix operations compared to the previous generation.
PCIe Gen 4
The NVIDIA A100 supports PCI Express Gen 4, which provides double the bandwidth of PCIe Gen 3, improving data-transfer speeds from CPU memory for data-intensive tasks like AI and data science.
High Speed HBM2e Memory
With 80 gigabytes (GB) of high-bandwidth memory (HBM2e), the NVIDIA A100 PCIe delivers improved raw bandwidth of 1.55TB/sec, as well as higher dynamic random access memory (DRAM) utilization efficiency at 95 percent. A100 PCIe delivers 1.7x higher memory bandwidth over the previous generation.
Error Correction Without a Performance or Capacity Hit
HBM2e memory implements error correction without any performance (bandwidth) or capacity hit, unlike competing technologies like GDDR6 or GDDR6X.
Compute Preemption
Preemption at the instruction-level provides finer grain control over compute and tasks to prevent longer-running applications from either monopolizing system resources or timing out.
Virtual GPU Software for Virtualization
Support for NVIDIA Virtual Compute Server (vCS) accelerates virtualized compute workloads such as high-performance computing, AI, data science, big-data analytics, and HPC applications.
Software Optimized for AI
Deep learning frameworks such as Caffe2, MXNet, CNTK, TensorFlow, and others deliver dramatically faster training times and higher multi-node training performance. GPU accelerated libraries such as cuDNN, cuBLAS, and TensorRT delivers higher performance for both deep learning inference and High-Performance Computing (HPC) applications.
NVIDIA CUDA Parallel Computing Platform
Natively execute standard programming languages like C/C++ and Fortran, and APIs such as OpenCL, OpenACC and Direct Compute to accelerates techniques such as ray tracing, video and image processing, and computation fluid dynamics.
Unified Memory
A single, seamless 49-bit virtual address space allows for the transparent migration of data between the full allocation of CPU and GPU memory.