A consumer desktop with one GeForce RTX 5090 is published by NVIDIA at over 3,352 trillion AI operations per second.1
This entry converts the published 3352 AI TOPS figure to dense INT8-equivalent throughput using the report normalization rules. Applying /2 from FP4 to INT8 and /2 to remove structured sparsity yields 838 dense INT8 TOPS, or 8.38e+14 operations per second.1
Total compute: 8.38e+14 dense INT8 operations per second.
RTX 5090-class personal systems push individual-accessible AI throughput into territory previously reserved for small server deployments while preserving comparability on the dense INT8 scale.
NVIDIA. “GeForce RTX 50 Series.” Accessed 2026. https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/. ↩ ↩2