A bunch of CompuBench tests of NVIDIA’s next-gen Ampere flagship, the RTX 3080 has been shared by APISAK on Twitter, giving us an approximate measure of the compute capabilities of the new GPUs compared to the existing Turing parts. Keep in mind that these are the CUDA and OpenCL benchmarks that represent the GPGPU (general-purpose computing) performance of these GPUs. That means compute operations rather than the graphics rendering, tasks usually that are usually performed by the CPU but offloaded to the GPU for acceleration and faster batch execution.
Having a look at these benchmarks, you can see that the performance varies greatly from test to test, with some showing a gain of 2x while the rest improving by around 50-70% while going from the RTX 2080 to 3080.
The RTX 2080 Ti, on the other hand, is around 25-50% slower than the 3080 across different tests. This huge variance is likely due to the revamped structure of the SM in Ampere which basically doubles FP32 performance at the expense of integer workloads. As a result, benchmarks that are INT32-intensive see trivial to moderate gains.