P100 vs RTX 3080

The FP16 dense figure of 59.5 TFLOPS on the RTX 3080 exceeds the 18.7 TFLOPS dense FP16 on the P100 by a ratio of 3.18 to 1. This delta means the RTX 3080 can complete training steps or inference passes in less time when workloads rely on dense FP16 operations. The FP32 value of 29.8 TFLOPS on the RTX 3080 compared with 9.3 TFLOPS on the P100 produces a similar ratio of 3.2 to 1 and favors the newer card for precision sensitive training. Memory bandwidth of 760 GB/s on the RTX 3080 versus 732 GB/s on the P100 allows modestly larger batch sizes before memory capacity limits are reached although the P100 holds 16 GB against the 10-12 GB on the RTX 3080. When sparsity is available the RTX 3080 reaches 119 TFLOPS in FP16 with sparsity while the P100 provides no published sparse figure so only dense to dense comparisons apply.

The P100 suits workloads that require 16 GB of HBM2 VRAM and NVLink interconnect at a TDP of 250W. Scientific computing tasks that fit within the 9.3 TFLOPS FP32 or 18.7 TFLOPS dense FP16 of the P100 benefit from its higher memory capacity and SXM2 form factor options.

The RTX 3080 fits workloads that need 59.5 TFLOPS dense FP16 or 29.8 TFLOPS FP32 at PCIe 4.0 bandwidth. Inference and training jobs that can use the 10-12 GB GDDR6X and 320W TDP gain from the higher dense FP16 throughput relative to the P100.