GTX 1080 vs RTX 4070

The FP32 rating of 8.9 TFLOPS on the GTX 1080 versus 29.1 TFLOPS on the RTX 4070 indicates the newer card processes roughly three times the single precision operations per second for training workloads. The RTX 4070 further lists dense FP16 at 58.3 TFLOPS and dense FP16 with sparsity at 116.6 TFLOPS whereas the GTX 1080 lists no FP16 figure. Memory bandwidth of 320 GB/s on the GTX 1080 versus 504 GB/s on the RTX 4070 limits batch sizes in memory bound inference steps by a factor of 504 divided by 320. Dense FP16 figures must be compared only against other dense FP16 figures and sparse figures only against other sparse figures so the RTX 4070 dense FP16 value of 58.3 TFLOPS stands alone against the GTX 1080 FP32 value of 8.9 TFLOPS for mixed precision tasks.

The GTX 1080 suits legacy scientific computing codes that run exclusively on FP32 at 8.9 TFLOPS and require no newer interconnect features. Its 180 W TDP also fits power constrained servers that already host PCIe 3.0 infrastructure and need only 8 to 11 GB of GDDR5X memory for small batch inference.

The RTX 4070 suits workloads that exploit its dense FP16 rating of 58.3 TFLOPS or its dense INT8 rating of 233.2 TOPS together with 504 GB/s bandwidth and PCIe 4.0 interconnect. Its 12 to 16 GB GDDR6X capacity further supports larger models than the 8 to 11 GB limit of the GTX 1080 at a TDP of 200 W.