If your stack is NUMA aware, and can manage that 4/8 partitions, enabling it would seem to reduce memory contention between chiplets; even if there isn't a latency penalty for going a bit farther to memory (which there likely is, but probably small)