In order to execute the Cuda PBS benchmark you can run:
RUSTFLAGS="-Ctarget-cpu=native" cargo run -p concrete-core-bench --release --features=backend_cuda -- --bench LweCiphertextVectorDiscardingBootstrapFixture2
The benchmarks in the post have been obtained on a V100 GPU. You can change the parameters of concrete-core-fixture/src/fixture/lwe_ciphertext_vector_discarding_bootstrap_2 to increase the number of inputs, change the polynomial size, etc.
Let me know if you have issues with the command or if you have further questions.