This paper is old but the ratios should still be pretty accurate. Main memory is 240 clock cycles away and L1d is only 3. If the experiments in this paper still hold true loading the 8K block into L1d is far more expensive than the CPU processing done once the block is in cache.
When one adds in NUMA to the contention on this shared resource, its not that hard for a 40 core machine to starve for memory bandwidth, and for cores to sit idle waiting for main memory. Eliminating wasted space seems far more important even when everything could fit in memory already.
