NVIDIA RTX 5090: Overkill Is Still the Point

The RTX 5090 costs $1,999. It has 32GB of GDDR7 memory on a 512-bit bus, a peak bandwidth of 1,792 GB/s, and a TDP of 575W. It is the fastest consumer GPU ever made by a margin that should not be possible two years after the 4090.

In Blender Cycles, the 5090 completes our standard benchmark scene in 4.2 seconds. The 4090 takes 9.8 seconds. In Stable Diffusion XL, the 5090 generates a 1024×1024 image in 0.6 seconds. The 4090 takes 1.4 seconds.

The 32GB of GDDR7 enables something genuinely new for consumer hardware: running full fine-tuning passes on 7B parameter language models locally, without quantisation. This is not a use case for most people. It is a use case for a rapidly growing number of AI developers who currently pay cloud API costs for work they could run at home.

The 8.8 is not a 10.0 because the power requirements are extreme, the price is difficult to justify for gaming alone, and DLSS 4 upscaling still introduces perceptible artifacts in fast motion. But for the person this card is built for, the 5090 is not just the best option. It is the only option.