P100 vs RTX A6000

Dense FP16 performance reaches 18.7 TFLOPS on the P100 compared with 154.8 TFLOPS dense on the A6000. The resulting eightfold increase allows the A6000 to complete training iterations and inference passes on larger models within the same time window. FP32 figures of 9.3 TFLOPS on the P100 versus 38.7 TFLOPS on the A6000 produce similar scaling for single precision workloads. Memory bandwidth of 732 GB/s on the P100 versus 768 GB/s on the A6000 supports modestly larger batch sizes on the A6000 before memory capacity of 16 GB versus 48 GB becomes the limiting factor. Sparse FP16 performance of 309.6 TFLOPS on the A6000 has no counterpart on the P100 so only dense figures are compared directly. INT8 dense performance of 309.7 TOPS on the A6000 further widens the gap for quantized inference paths.

The P100 suits workloads that remain within 16 GB HBM2 and operate under a 250W TDP limit. Its SXM2 form factor and PCIe 3.0 interconnect match legacy server designs that cannot accommodate the A6000 PCIe layout. Scientific computing tasks that rely on the published 9.3 TFLOPS FP32 rating avoid the higher 300W draw of the A6000.

The A6000 fits workloads that require 48 GB GDDR6 and dense FP16 throughput of 154.8 TFLOPS. Its PCIe 4.0 interconnect and NVLink support higher data movement rates than the P100 PCIe 3.0 path.