An arXiv preprint reports lower calculated CPU bandwidth and fewer cache misses, but its throughput comparison uses unmatched models.