5 ms·
I don't think that ARM is as clear a win in high performance areas as you seem to think. ARM does let you get away with not having most of the decode logic tha
by Symmetry 16y ago
I don't think that ARM is as clear a win in high performance areas as you seem to think. ARM does let you get away with not having most of the decode logic that an x86 has to carry around, but thats pretty small penuts compared to out-or-order logic with a decent reorder window size, or a wide execution cluster with the associated bypass network.
And while instruction predication does let you avoid a lot of branching nearly for free in the ARM architecture it makes out of order logic much less effective because now there are a lot more dependencies between instructions. That was the reason the Alpha team decided not to use predication, even though they were familiar with the idea. Its also why the Itanium, an in order architecture, used predication.
- Andys 16y agoYep, taking a look at a modern x86 core in 32nm (and 22nm is just around the corner), the x86 decoding is now just a tiny part of the core. Intel has been been picking low-hanging fruit and not trying particularly hard to make things very low power. The current area of work is GPU integration with CPU on the same die (see Sandy Bridge). Consumer electronics is driven by volume & price, so all we are really seeing is small, low power cores becoming acceptably fast for mainstream computing.
- hackermom 16y agoThis is not just about the underlying fundaments and how the core is implemented, but also about the inefficiency of the x86 ISA.
- tptacek 16y agoDidn't he just make the opposite argument, and with evidence?
- hackermom 16y agoNo, he detailed specific differences in the branch predictors of the two mentioned architectures. Branch prediction is part of the microarchitecture of a CPU, not its ISA. These two should not be confused.
- tptacek 16y agoPredication is one of the signature features of the ARM ISA. Could you be specific about the costly X86 ISA features you're talking about?
- Symmetry 16y agoAs the other commenter pointed out, I'm talking about instruction predication (also called conditional execution) instead of branch prediction. In any ISA you have branches that might or might not execute depending on some conditions. With ARM, the first four bits of every single instruction - not just branches but loads and stores, addition and subtraction, etc - are used to specify under what conditions the instruction will be executed - so you can have small 'if' statements that don't actually need any branches when they're compiled. All that is really pretty awesome if you're designing your processor to go in order, but if you want to try out of order execution it becomes more complicated. Now you have to check every instruction to see if its predicated, and if it is you now have a new dependency on the previous arithmetic instruction even if there isn't any data dependency between the two. Of course you could argue that if you find a predicated instruction in ARM code then its replacing a branch that would be in x86 code, and that overall your job is no harder. And you could go back and forth arguing over it. The important thing to remember, though, is the things that give the ARM ISA inherent advantages when you're making low power, in order processors aren't necessarily advantages when you're talking about high performance, high power processors.
- philwelch 16y agoIf the inefficiency of the x86 ISA was a killer, we'd all be using Alpha, PowerPC, SPARC, PA-RISC, or Itanium processors today. The basic argument you're giving is about 20 years old by now and the market has seen it tested.