GPU architectures

Every GPU architecture in the database, newest first, with what changed in each generation.

NVIDIA

  • Feynman2028 · Not disclosed · 1 products. 3D die stacking, custom HBM, optical NVLink.
  • Rubin2026–27 · TSMC 3 nm class · 5 products. HBM4, NVLink 6, Vera CPU with custom Olympus cores.
  • Blackwell2024–25 · TSMC 4NP / 4N · 35 products. FP4 Tensor Cores, GDDR7, dual-die data-center GPUs, DLSS 4 Multi Frame Generation and DLSS 5.
  • Hopper2022 · TSMC 4N · 3 products. Transformer Engine with FP8.
  • Ada Lovelace2022 · TSMC 4N · 22 products. DLSS 3 Frame Generation; very large L2 cache.
  • Ampere2020 · Samsung 8 nm / TSMC 7 nm · 14 products. 2nd-gen RT, 3rd-gen Tensor, GDDR6X, MIG.
  • Turing2018 · TSMC 12 nm · 14 products. First RT cores; DLSS 1.0.
  • Volta2017 · TSMC 12 nm · 1 products. First Tensor Cores.
  • Pascal2016 · TSMC 16 nm / Samsung 14 nm · 9 products. GDDR5X and HBM2; large efficiency gain over Maxwell.

AMD

  • CDNA 52026 · TSMC 2 nm + 3 nm · 3 products. HBM4 up to 432 GB; rack-scale Helios.
  • RDNA 42025 · TSMC N4P · 9 products. Much faster ray tracing, FP8/FP4 AI accelerators for FSR 4.
  • CDNA 42025 · TSMC 3 nm + 6 nm · 1 products. FP4 and FP6 support; 288 GB HBM3E.
  • RDNA 3.52024–26 · TSMC 4 nm · 2 products. Power-optimised RDNA 3 for APUs.
  • CDNA 32023–24 · TSMC 5 nm + 6 nm · 2 products. 3D-stacked chiplets; up to 256 GB HBM3E.
  • RDNA 32022 · TSMC 5 nm + 6 nm · 15 products. First chiplet consumer GPUs; AI accelerators per CU.
  • RDNA 22020 · TSMC 7 / 6 nm · 10 products. Hardware ray tracing and Infinity Cache.
  • RDNA2019 · TSMC 7 nm · 4 products. New compute-unit design replacing GCN; PCIe 4.0.

Intel

  • Xe2 (Battlemage)2024–26 · TSMC N5 · 7 products. Large per-core gains; 12–32 GB configurations.
  • Xe-HPG (Alchemist)2022 · TSMC N6 · 3 products. Intel's first discrete gaming GPUs; XMX matrix engines and AV1 encode.

Apple

  • Apple M5 GPU2025–26 · TSMC 3 nm (3rd gen) · 4 products. A Neural Accelerator in every GPU core for AI workloads.
  • Apple M4 GPU2024 · TSMC 3 nm (2nd gen) · 3 products. Faster ray-tracing engine; Thunderbolt 5 on Pro and Max.
  • Apple M3 GPU2023 · TSMC 3 nm · 4 products. Dynamic Caching, mesh shading and hardware ray tracing.
  • Apple M2 GPU2022 · TSMC 5 nm (2nd gen) · 4 products. More cores and bandwidth; up to 192 GB unified memory on M2 Ultra.
  • Apple M1 GPU2020 · TSMC 5 nm · 4 products. Tile-based deferred rendering GPU sharing one unified memory pool with the CPU.