Chat > Tech Chat

No wonder Apple is moving to ARM

(1/2) > >>

phi2008:

--- Quote ---Arm promises Intel Core i5 performance at lower power

Arm has published a blog post about how it is accelerating mobile and laptop performance. The key graphic shared by the chip designer shows of the firm's first ever public forward looking roadmap with performance numbers. In brief, later this year, Arm expects its <5W Cortex-A76 to offer performance on a par with the 15W Intel Core i5-7300U. (To judge comparative performance it used the single-core SPEC CINT2006 benchmark.) Furthermore, Arm predicts that its upcoming processor generations will deliver >15 per cent computer performance uplifts through 2020 to easily break beyond Intel's trajectory.

...

https://hexus.net/tech/news/cpu/121217-arm-promises-intel-core-i5-performance-lower-power/

--- End quote ---

Looks like ARM is going to corner the laptop market at least.  ;)

Weaver:
The thing is, the logic to allow multiple instructions to execute in parallel in one thread safely is possibly costly. I don't know how much power that is. That is going to spoil the low power characteristics of these machines, and they would be wasting energy in order to get superb performance. On Intel Haswell and later new processors, if logic allows, you can execute four instructions in parallel, more where instructions get deleted altogether, register-to-register moves are an example. This crucifies ARM for straightforward computing tasks. Sometimes programs have sequential dependancies and no room for instruction parallelism and in that case your luck is out. But if your luck is in, you can easily get a three time or four times or more speed up, so your 2GHz processor turns into an 8GHz one effectively if you have the right kind of sequence of ordinary common instructions in an inner loop.

If ARM designers wish to get the very best battery performance then they simply will not deliver this and so they will always be three or four times slower in some fairly common situations. Or not. Mains powered machines, who cares, so there would be an argument for a massive gap between Mac CPUs and iPhone CPUs and who knows where you would put the iPad, probably in the iPhone category. However perhaps it would be an idea to give users a choice of iPad models depending on whether or not you, like me, never run your iPad on battery.

22over7:
@weaver, I'm interested. I can't find the string "thread" in the link @phi2008 posted.  (Plus I've lingering interests in processor design, and have terminally annoyed everyone I knew who still does it.  Did you once (do it)?)

To my mind, the field started to go insane a couple of decades ago when performance began to mean how many machine/architectural instructions you could retire per (lucky) microsecond, regardless of area, power, heat, complexity/bugs, address-spaces and multithreading.  I reckon we're pretty close to the wall in terms of what human beings can design reliably with a (basically) 1960's mind-set.

(I'm quite open minded really...)

Weaver:
The thing about threads is that some processors, some intel processors have two program counters (iirc Intel call it the IP register) and two sets of registers so two complete sets of state and can execute two threads in parallel. They call it hyperthreading. Then there is the issue of n complete processors on one chip. So this could be n or 2n threads.

So I was restricting this to talk about one single thread, and the number of instructions executed from a small region around the same PC value.

I have never done any work involving hardware design, although I would love to start thinking about a design for a processor architecture and have given it some small amount of thought. I have done a lot of detailed reading in the subject in recent years. I used to write a lot of assembly language for a living, that's all, so I am machine code literate.

I am not sure that I am convinced that there are not more very radical things that could be carried out, as I have a couple of ideas that have never been done, but I don't know whether they would actually speed anything up, although some might tolerate higher clock speeds and that would be enough. The limits of electronics will soon be reached and then it will be down to extremely radical architectural changes or even more parallelism.

phi2008:

--- Quote ---I have never done any work involving hardware design ...
--- End quote ---

Maybe a good place to start is FPGAs - people often emulate various CPUs on them.

DESIGNING A CPU IN VHDL FOR FPGAS

 :)

Navigation

[0] Message Index

[#] Next page