A lot of server workloads also aren't about maximizing flops/Watt or iops/Watt, but rather setting up DMA between the disk controller and the network card and having a low idle power, to get the most network throughput per Watt. A lot of the high-end computation is also about handing off data to the GPU and having a low idle power.
It took Intel way too long to get the power consumption on the Atom processors down (not ops/Watt, but total Watts), back when Atom had a chance in the phone market. (A difficult task, which is why I bought ARM Holdings stock shortly after Android came out because I realized (1) nearly everyone was going to soon have a smartphone and (2) whether iPhone or Android won, it was going to be very difficult for Intel to get competitive in the smartphone market fast enough to prevent ossification of the market around ARM. Unfortunately, SoftBank took ARM private and closed out my trade.)
Now, the same dynamic of the low-end x86 processor eating its way up the stack to threaten the high-margin mainframe and RISC servers is playing out with low-end ARM eating its way up the stack and displacing a lot of uses for the high-margin x86 server chips.
It took Intel way too long to get the power consumption on the Atom processors down (not ops/Watt, but total Watts), back when Atom had a chance in the phone market. (A difficult task, which is why I bought ARM Holdings stock shortly after Android came out because I realized (1) nearly everyone was going to soon have a smartphone and (2) whether iPhone or Android won, it was going to be very difficult for Intel to get competitive in the smartphone market fast enough to prevent ossification of the market around ARM. Unfortunately, SoftBank took ARM private and closed out my trade.)
Now, the same dynamic of the low-end x86 processor eating its way up the stack to threaten the high-margin mainframe and RISC servers is playing out with low-end ARM eating its way up the stack and displacing a lot of uses for the high-margin x86 server chips.