The SYCL 2020 Shared USM feature, sycl::malloc_shared, maps 1:1 with CUDA managed memory allocations using the Driver API cuMemAllocManaged. Programmers should be aware that currently the DPC++ runtime does not have sophisticated heuristics for auto-management of Shared USM. Therefore the default performance of Shared USM may not be as good as the CUDA Runtime API equivalent using cudaMallocManaged. However it is still possible for users to manually use sycl::prefetch and sycl::queue::mem_advise / sycl::handler::mem_advise APIs; judicious usage of which can lead to optimal performance of shared (managed) USM in DPC++. sycl::prefetch and sycl::queue::mem_advise (or sycl::handler::mem_advise) map 1:1 to cuMemPrefetchAsync and cuMemAdvise respectively. Users should consult the latest Nvidia documentation for expected behaviour and best practices when prefetching and advising for their hardware. The purpose of the following section is to describe the mapping between cudaMemAdvise* advice and their equivalents in the CUDA backend of DPC++.