What is the NVIDIA Ampere architecture?
2026-10-28
In 2020, NVIDIA introduced a new GPU architecture called NVIDIA Ampere. It became the foundation for a broad family of graphics accelerators that continued to be widely used in data-center solutions more than five years later. Popular models include the A100, A40, A30, and A10.
The NVIDIA Ampere architecture combines different types of processing cores in a single accelerator: Streaming Multiprocessors (SMs) with CUDA cores and third-generation Tensor Cores. Some models are also equipped with second-generation RT cores for graphics workloads. The high-end NVIDIA A100 supports FP64, FP32, TF32, BF16, FP16, and INT8 computing modes. Its Structured Sparsity 2:4 mechanism can increase the performance of compatible neural-network operations by using hardware acceleration for structured sparsity.
A100 data-center GPUs are available with 40 or 80 GB of HBM2 or HBM2e memory. The 80 GB versions provide memory bandwidth of more than 2 TB/s. NVIDIA’s proprietary NVLink interconnect allows multiple Ampere-based GPUs to be combined into a cluster with high-speed, direct communication between the cards. Ampere-based accelerators are widely used for machine learning, AI inference, high-performance computing, big-data analytics, scientific modeling, and many other demanding workloads. Although newer NVIDIA architectures—including Hopper, Blackwell, and Ada Lovelace—have since appeared, Ampere continues to be widely deployed in enterprise data centers.