RTX 2080 vs RTX 3060

The 40.3 dense TFLOPS FP16 rating of the RTX 2080 enables higher throughput in mixed precision training loops than the RTX 3060 can demonstrate from its published 12.7 FP32 TFLOPS alone. Inference tasks that rely on dense INT8 operations benefit from the 161.1 dense TOPS figure available only on the RTX 2080. Memory bandwidth of 448 GB/s on the RTX 2080 supports larger batch sizes in memory bound kernels compared with the 360 GB/s limit of the RTX 3060. The 10 TFLOPS FP32 of the RTX 2080 versus 12.7 TFLOPS FP32 of the RTX 3060 indicates the newer card holds an edge in single precision workloads while the 45W lower TDP of the RTX 3060 reduces system power draw during sustained FP32 execution.

Choose the RTX 2080 when your model and batch fit inside 8 GB and the job is bandwidth bound, since 448 GB/s beats the 360 GB/s of the RTX 3060. Its tensor cores are rated at 40.3 TFLOPS dense FP16, so small model FP16 inference runs well. Do not pick it for anything that needs more than 8 GB, and treat the NVLink listing as a footnote rather than a reason to rent it.

Choose the RTX 3060 for most rentals. The 12 GB configuration loads models that an 8 GB RTX 2080 cannot, and its 112 Ampere tensor cores run FP16 and mixed precision work normally even though the table lists no figure. It also draws 170W against 215W. The only thing you give up is bandwidth, 360 GB/s against 448 GB/s, which matters mainly for small models that already fit in 8 GB.