1 |
Fernando Rodriguez <frodriguez.developer <at> outlook.com> writes: |
2 |
|
3 |
|
4 |
|
5 |
> > albeit in it's infancy. Naturally it's going to take a while to |
6 |
> > become mainstream useful; but that more like a year or 2, at most. |
7 |
> |
8 |
> The value I see on that technology for desktop computing is that we get the |
9 |
> GPUs for what they're made (graphics processing) but their resources go |
10 |
unused |
11 |
> by most applications, not in buying powerful GPUs for the purpose of |
12 |
offloading |
13 |
> general purpose code, if that's the goal you're better off investing in more |
14 |
> general purpose cores that are more suited for the task. |
15 |
|
16 |
|
17 |
I think most folks when purchasing a workstation include a graphics |
18 |
card on the list of items to include. So my suggestions where geared |
19 |
towards informing folks about some of the new features of gcc that |
20 |
may intice them to consider the graphics card resources in an |
21 |
expanded vision of general resources for their workstation. |
22 |
|
23 |
|
24 |
> To trully take advantage of the GPU the actual algorithms need to be |
25 |
rewritten |
26 |
> to use features like SIMD and other advanced parallelization features, most |
27 |
> desktop workloads don't lend themselves for that kind of parallelization. |
28 |
|
29 |
Not true if what openacc hopes to achived indeed does become a reality. |
30 |
Currently, you are most correct. Things change; I'm an optimist because |
31 |
I see what is occuring in embedded devices, arm64, and cluster codes. |
32 |
ymmv. |
33 |
|
34 |
> That |
35 |
> is why despite similar predictions about how OpenMP-like parallel models |
36 |
would |
37 |
> obsolete the current threads model since they where first proposed, it |
38 |
hasn't |
39 |
> happened yet. |
40 |
|
41 |
Yes it's still new technology, controversial, just like systemd, clusters, |
42 |
and Software Defined Networks. |
43 |
|
44 |
|
45 |
> Even for the purpose of offloading general purpose code, it seems with all |
46 |
the |
47 |
> limitations on OpanACC kernels few desktop applications can take advantage of |
48 |
> it (and noticeably benefit from it) without major rewrites. Off the top of my |
49 |
> head audio, video/graphics encoders, and a few other things that max out the |
50 |
> cpu and can be broken into independent execution units. |
51 |
|
52 |
|
53 |
You are taking a very conservative view of things. Codes being worked |
54 |
out now for clusters, will find their way to expand the use of the |
55 |
video card resources, for general purpose things. Most of this will |
56 |
occur as compiler enhancements, not rewriting by hand or modifying |
57 |
algorithmic designs of existing codes. Granted they are going to |
58 |
mostly apply to multi-threaded application codes. |
59 |
|
60 |
|
61 |
When folks buy new hardware, it is often a good time to look at what |
62 |
is on the horizon for computers they use. All I have pointed out is |
63 |
a very active area that benefits folks to review for themselves. I not |
64 |
pushing expenditures of any kind on any hardware. |
65 |
|
66 |
Caveat Emptor. |
67 |
|
68 |
|
69 |
|
70 |
James |