RTX 3090 vs RTX 4080

The RTX 4080 produces a higher dense FP16 figure of 97.5 TFLOPS compared with the RTX 3090 dense FP16 figure of 71 TFLOPS and a higher FP32 figure of 48.7 TFLOPS compared with 35.6 TFLOPS. This difference supports faster dense FP16 operations during both training and inference workloads on the RTX 4080 while the RTX 3090 higher memory bandwidth of 936 GB/s versus 717 GB/s permits larger batch sizes in memory constrained scenarios. Sparse FP16 performance reaches 195 TFLOPS on the RTX 4080 against 142 TFLOPS on the RTX 3090 and the RTX 3090 further supplies NVLink interconnect absent from the RTX 4080.

The RTX 3090 suits workloads that require the 24 GB VRAM capacity for larger models or datasets that exceed 16 GB. Its 936 GB/s memory bandwidth also supports higher throughput in bandwidth sensitive tasks and its NVLink interconnect enables multi GPU scaling not available on the RTX 4080.

The RTX 4080 suits workloads that benefit from its 97.5 TFLOPS dense FP16 and 48.7 TFLOPS FP32 performance at a TDP of 320W. Its additional FP8 dense rating of 194.9 TFLOPS and INT8 dense rating of 389.9 TOPS further favor inference pipelines that utilize those data types.