1 |
Am 09.12.2012 04:51, schrieb Michael Mol: |
2 |
> On Sat, Dec 8, 2012 at 10:25 PM, Grant <emailgrant@×××××.com> wrote: |
3 |
>> It seems like ARM processors will destroy x86 before too long. Does anyone |
4 |
>> think this won't happen? |
5 |
> |
6 |
> It's looking promising. Not that I have a horse in the race, but I |
7 |
> very much like ARM's low power consumption. The way I see it, they're |
8 |
> only a short list of features away from obliterating x86: |
9 |
> |
10 |
> * I'd like to see fast division. |
11 |
> I keep hearing about how this or that is slow because of ARM's lack of |
12 |
> strong division. |
13 |
> |
14 |
> * I'd like to see a modern baseline of strong instructions. |
15 |
> x86 kept continually improving in a very fragmented way, but there |
16 |
> were, from time to time, baseline collections of feature sets you |
17 |
> could expect all processors to have. i386 represented one. i686 |
18 |
> represented one. Currently, it's x86_64, which implies not only a |
19 |
> 64-bit flattened address space and a departure from real mode, but |
20 |
> also a collection of SIMD instruction sets and other features |
21 |
> developed between the release of the Pentium Pro and AMD's Hammer |
22 |
> architecture. |
23 |
> |
24 |
> ARM just feels...fragmented. And I don't have the impression I could |
25 |
> write my code assuming the availability of SIMD (presuming I use |
26 |
> things like OpenMP to expand my code to leverage it, rather than |
27 |
> writing processor-specific code. Though OpenCL could very well |
28 |
> alleviate that issue.) |
29 |
> |
30 |
|
31 |
+1 with regard to fragmentation. What I especially despise is the lack |
32 |
of a common boot infrastructure. If I'm not mistaken, it is still |
33 |
impossible to make a kernel that boots on all (or at least a large |
34 |
subset of all) ARM platforms [1]. |
35 |
|
36 |
And then, there is the simple fact that current ARMs lack the raw power |
37 |
of an x86 and I guess if you scale them up to the point where they can |
38 |
compete with x86s with regard to computing power per core, there is no |
39 |
point in switching to ARM to begin with. Sure, you can parallelize and |
40 |
make a large array of "wimpy" nodes, but you cannot fool Amdahl's law. |
41 |
And even where you can parallelize nearly 100%, you risk high latency |
42 |
[2, 3]. |
43 |
|
44 |
[1] https://lwn.net/Articles/496400/ |
45 |
[2] http://pages.cs.wisc.edu/~jignesh/publ/nonwimpy.pdf |
46 |
[3] http://research.google.com/pubs/archive/36448.pdf |
47 |
|
48 |
Regards, |
49 |
Florian Philipp |