RTX 5080 vs V100

The FP32 difference of 56.3 TFLOPS versus 15.7 TFLOPS favors the RTX 5080 by a factor of 3.59 for single precision workloads. Dense FP16 performance shows the V100 at 125 TFLOPS ahead of the RTX 5080 at 112.6 TFLOPS so training loops that stay within dense FP16 may favor the older card. The RTX 5080 supplies additional sparsity modes at 225.1 TFLOPS FP16 that the V100 does not list. Memory bandwidth of 960 GB/s on the RTX 5080 versus 900 GB/s on the V100 permits modestly larger batch sizes before memory stalls occur. TDP values of 360 W and 300 W indicate the V100 draws less power under identical dense FP16 loads.

Choose the RTX 5080 for inference of models that fit in 16 GB, including FP8 serving at 225.1 TFLOPS dense, a format the V100 does not publish. Its 960 GB/s of GDDR7 bandwidth feeds token generation at least as well as the V100's 900 GB/s, and 56.3 TFLOPS of FP32 is 3.6x the V100 figure for anything that is not tensor core bound. It also fits fine tuning of small models with LoRA style methods, as long as the whole job stays within the 16 GB card.

Choose the V100 when your model or working set needs more than 16 GB, since the V100 comes in a 32 GB HBM2 configuration and the RTX 5080 stops at 16 GB. Choose it when you need NVLink to pool memory across several cards, or when the job is double precision, where its 7.8 TFLOPS of FP64 has no published counterpart on the RTX 5080. Its 300W TDP against 360W is a minor point; the memory ceiling and FP64 are the real reasons to pick it.