Embodiments that dynamically conserve power in non-uniform cache access (NUCA) caches are contemplated. Various embodiments comprise a computing device, having one or more processors coupled with one or more NUCA cache elements. The NUCA cache elements may comprise one or more banks of cache memory, wherein ways of the cache are vertically distributed across multiple banks. To conserve power, the computing devices generally turn off groups of banks, in a sequential manner according to different power states, based on the access latencies of the banks. The computing devices may first turn off groups having the greatest access latencies. The computing devices may conserve additional power by turning of more groups of banks according to different power states, continuing to turn off groups with larger access latencies before turning off groups with the smaller access latencies.