GPU architectures
Nvidia Tesla to Blackwell, AMD GCN, RDNA and CDNA, Intel Xe, Arm Mali and Imagination PowerVR. 134 chips on EveryChip so far.
GPU architectures are the design generations of graphics and compute processors. Nvidia names each one after a scientist (Tesla, Fermi, Kepler, Pascal, Volta, Turing, Ampere, Hopper, Blackwell, Rubin), AMD uses families such as Graphics Core Next, RDNA and CDNA, and Intel has Xe. Each generation changes how the many small cores are grouped, what special units they carry and how they reach memory.
Generation by generation
Nvidia's Tesla (2006) unified the shader units and brought CUDA. Volta (2017) added tensor cores for matrix maths, Turing (2018) added ray tracing cores for realistic lighting, and Hopper (2022) and Blackwell (2024) were built for AI. AMD's Graphics Core Next (2012) served both games and compute; in 2019 and 2020 AMD split its line into RDNA for games and CDNA for data centres.
Programming models
GPU instruction sets change with each generation, so programs are written in higher-level languages such as CUDA, OpenCL, HIP or shader languages, and a driver compiles them for the chip. Nvidia's long investment in CUDA software is one reason for its lead in AI. Mobile GPUs from Arm (Mali), Imagination (PowerVR), Qualcomm (Adreno) and Apple use tile-based designs that save memory traffic.
Famous GPU architecture chips
Further reading
Photo credits on this page
The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.
- Tesla: Hyins, Public domain, source
- Fermi: Department of Energy-Office of Public Affairs, restored by Yann, Public domain, source
- Pascal: Fritzchens Fritz, CC0, source
- Graphics Core Next: AMD, Public domain, source
- RDNA: Fritzchens Fritz, CC0, source
- Hopper: 极客湾Geekerwan, CC BY 3.0, source
- Blackwell: 极客湾Geekerwan, CC BY 3.0, source
