NVIDIA AI Chip Architecture Map

Every GPU from Ampere to Rubin — specs, performance & status

5
Architectures
29
Chip Variants
2020–2028
Timeline
~80%
Market Share

Memory Capacity Comparison

HBM / GDDR per GPU (GB) — data center chips only

FP4 / FP8 Compute (TFLOPS)

Peak tensor performance per GPU (with sparsity where noted)

Power Envelope (TDP)

Watts per GPU — showing the dramatic rise across generations

Data compiled February 2026 from NVIDIA official datasheets, industry sources & analyst reports.
* Values with sparsity noted. Rack-level systems (NVL72) shown separately. Pricing is estimated street/cloud pricing.