RTX 5090 vs V100

Dense FP16 throughput reaches 209.5 TFLOPS on the RTX 5090 against 125 TFLOPS on the V100 yielding a ratio of 1.676 times higher dense FP16 capacity. The same ratio holds when comparing dense FP16 figures directly without mixing sparsity values. Memory bandwidth at 1792 GB/s on the RTX 5090 versus 900 GB/s on the V100 permits larger batch sizes during training or inference before memory capacity limits are reached. FP32 throughput of 104.8 TFLOPS on the RTX 5090 versus 15.7 TFLOPS on the V100 further widens the gap for precision sensitive operations that rely on dense FP32 arithmetic.

The RTX 5090 suits workloads that demand the highest dense FP16 throughput of 209.5 TFLOPS and memory bandwidth of 1792 GB/s. Its 32 GB GDDR7 configuration and 104.8 TFLOPS dense FP32 rating support larger models and faster iteration cycles than the V100 can achieve.

Choose the V100 only for FP64 work or for small batch inference of older models. Its 7.8 TFLOPS FP64 is about 4.9 times the 1.6 TFLOPS of the RTX 5090, so legacy double precision HPC codes still run better on it, and the SXM2 version has NVLink for multi-GPU jobs that the PCIe only RTX 5090 lacks. For modern LLM inference it is a poor fit: it lists no FP8 support, predates BF16, and its 900 GB/s bandwidth is half of the 1792 GB/s on the RTX 5090.