Out-of-the-box Computing - The Mill general purpose CPU architecture

[DNS]

Hi,

I have been casual reader of this forum for a number of years now, so I thought it might be time to join in and contribute something I have not seen discussed yet. The Mill general purpose CPU.

http://www.youtube.com/watch?v=LgLNyMAi-0I

 

[sebbbi]

I liked many of their approaches (the front end, the belt, etc). However I am not yet fully sold on the idea that their latency hiding mechanism can hide memory latency equally well compared to modern OoO processors in pointer (double and triple indirection) heavy code. Their solution to handle false aliasing (without needing heavy store forwarding mechanism or register renaming) is great, and will be a big improvement over older in-order cores. It will solve the issue of hiding L1 and L2 latency, but it cannot compete with 192 entry ROBs of Haswell in hiding memory latency. Naturally neither will solve the issue of long pointer chains (linked lists), but "modern" object oriented code contains lots of shorter pointer chains that are quite well handled by massive OoO machinery.