| snowdog said:
Great post, but thought I'd add two important things. Firstly that Expresso has a ridiculously short pipeline (4 stages!) and Jaguar has a 17 stages, so over 4 times more stages to do before a process is completed and secondly that Expresso also has access to 32MB of eDRAM in addition to the 3MB of CPU cache. The second point above is HUGE. |
Short pipeline stages are a limiting factor for several things. For starters, they don't allow such heavy amounts of floating point instruction sets, as we've seen that Espresso shares the same 32 x 2 Paired Singles instructions that Broadway and Gekko had. There are advantages and disadvantages to having short pipeline stages. But considering how this processor can run modern software decently, I think it's something that pushes well above its weight. If I were to code for a CPU in 2013 that used 32x2 SIMD, I would definitely have some doubts about it. Its architecture is very well designed for what it was made to do, and has aged very well.
I guess the odd-ball for Espresso is having so much cache for one core, I don't really understand their purpose for doing that. Based on the die shots, there isn't anything different about Core 01 compared to the other two, aside from having access to 4x more cache than the other two. One plus for the cache though, is having eDRAM as cache.
