P100 vs V100

The dense FP16 rating reaches 18.7 TFLOPS on the P100 and 125 TFLOPS on the V100. This gap indicates that workloads using dense FP16 operations complete faster on the V100. Memory bandwidth at 900 GB/s on the V100 compared with 732 GB/s on the P100 supports larger batch sizes during training sessions that rely on dense FP16 arithmetic. FP32 throughput at 15.7 TFLOPS versus 9.3 TFLOPS further favors the V100 for mixed precision pipelines that alternate between dense FP32 and dense FP16 stages.

Choose the P100 only for FP32 or FP64 work that fits in 16 GB and does not use tensor cores: it lists 9.3 TFLOPS FP32 and 4.7 TFLOPS FP64, both roughly 60 percent of the V100 figures, and 732 GB/s of HBM2 bandwidth. It also has NVLink and the same SXM2 and PCIe form factors, so multi card FP64 simulations are possible. Do not choose it for LLM inference, fine tuning or image generation; with no published tensor core count, its 18.7 TFLOPS of FP16 is about 15 percent of the V100 figure.

The V100 suits training sessions that exploit dense FP16 throughput of 125 TFLOPS and memory bandwidth of 900 GB/s. Workloads that scale with 16 to 32 GB HBM2 configurations benefit from the higher FP32 rating of 15.7 TFLOPS at 300W TDP.