Computing · Consumer GPUs

Consumer GPU FP32 Performance

Single-precision throughput, the metric that tracks real architectural improvement, not just transistor count.

• ~300× in 19 years
• Unified shaders + CUDA opened GPGPU
• Same silicon now powers AI training
• MAD/FMA basis, not dual-issue peaks

FP32 TFLOPS captures what gamers and developers actually feel: shader throughput. Unlike transistor count, it reflects clock speeds, memory bandwidth, and architectural efficiency together. Tracked here through NVIDIA’s flagship GeForce line, from the 8800 GTX (the first card to make GPGPU credible) to the RTX 5090, peak single-precision throughput has grown roughly 300× in under two decades. Every point uses the standard MAD/FMA basis (cores × clock × 2), not the dual-issue peaks NVIDIA quoted for some early cards. The gains came from unified shader architectures, the CUDA programming model that turned graphics cards into general-purpose accelerators, and ever-wider memory buses, the same silicon that now trains and runs most of the world’s AI.

Source: Wikipedia: NVIDIA GPU lists

Download the data: gpu_consumer.csv · gpu_consumer.json

All curves · The chart is drawn in your browser and needs JavaScript. Every number it plots is in the table above.