A10 vs V100

FP16 dense performance matches exactly at 125 TFLOPS for both the A10 and the V100 so training workloads that stay within dense FP16 see no arithmetic advantage either way. The A10 FP32 figure of 31.2 TFLOPS is exactly double the V100 figure of 15.7 TFLOPS therefore mixed precision or FP32 heavy fine tuning finishes in fewer cycles on the A10. When sparsity is enabled the A10 reaches 250 TFLOPS in FP16 with sparsity while the V100 supplies no sparsity rating for FP16. Memory bandwidth of 600 GB/s on the A10 versus 900 GB/s on the V100 limits maximum batch size on the A10 for bandwidth bound kernels. The A10 TDP of 150W versus the V100 TDP of 300W further reduces operating cost per dense FP16 TFLOPS on the A10.

The A10 suits deployments that require the lowest power draw at 150W while retaining 24 GB memory and 31.2 TFLOPS FP32. PCIe only installations also favor the A10 because it omits any requirement for SXM2 sockets. Workloads that benefit from INT8 dense performance at 250 TOPS find the A10 supplies published ratings absent from the V100.

The V100 fits systems already equipped for NVLink interconnect and SXM2 form factor. Its 900 GB/s memory bandwidth supports larger batch sizes than the 600 GB/s available on the A10 in memory intensive scientific kernels. Dual slot installations that accept the 300W TDP can utilize the V100 range of 16 to 32 GB HBM2 memory.