Evolution of NVIDIA GPU Architectures for AI from Volta to Blackwell

Since the Volta era, NVIDIA's GPU architectures have increasingly focused on deep‑learning optimizations. Volta introduced the first Tensor Core, delivering a three‑fold performance boost over Pascal for AI training and inference.

Turing added integer support (INT8, INT4, INT1), a 32× performance increase over Pascal, and introduced RT Cores for ray tracing.

Ampere (2020) brought TF32 and BF16 support, sparse matrix acceleration, and NVLink for high‑bandwidth GPU‑to‑GPU communication, further improving efficiency and reducing power consumption.

Hopper (2022) featured FP8 Tensor Cores, removed RT Cores to prioritize AI compute, and added a Transformer engine for modern models.

Blackwell (2024) introduced the GB200 Superchip, second‑generation Transformer engine, FP4/FP6 precision support, and doubled NVLink bandwidth to 1800 GB/s, achieving up to 30× LLM inference performance over H100.

Rubin GPUs, named after Vera Rubin, target extreme inference workloads with 50 petaflops FP8 performance and 288 GB HBM4 memory, while the Vera Rubin NVL144 and future NVL576 systems aim for exaflop‑scale AI computing.

The article also lists recent updates to CPU, GPU, memory, storage, and system technologies, and promotes downloadable resources such as comprehensive server, storage, and AI‑chip knowledge packs.