Note that those 13.1 TFLOPs are FP64, which isn't supported natively on the MacBook GPU. On the other hand, local/per-node memory bandwidth is significantly hig

Note that those 13.1 TFLOPs are FP64, which isn't supported natively on the MacBook GPU. On the other hand, local/per-node memory bandwidth is significantly higher on the MacBook. (Apparently, SGI Altix only had 8.5 to 12.8 GB/s.) Total memory bandwidth on larger Altix systems was of course much higher due to the ridiculous node count. Access to remote memory on other nodes could be quite slow because it had to go through multiple router hops.