1 |
Duncan, mused, then expounded: |
2 |
> |
3 |
> As for AMD, here's what I skipped over in the previous post. They've |
4 |
> already teamed up with various third parties to develop and sell physics |
5 |
> and floating point processors slotted into additional CPU sockets, linked |
6 |
> directly to the multi-core CPUs via Cohesive HyperTransport. The next |
7 |
> step is doing the same thing with full video processors, a decent portion |
8 |
> of which are physics processors anyway. There are however a couple issues |
9 |
> with the idea, including the fact that getting video folks to commit to an |
10 |
> AMD specific platform when they already have PCI-E would have been rather |
11 |
> difficult. This is supposedly one of the big reasons they bought ATI -- |
12 |
> to give AMD the where-with-all to follow up on that idea. |
13 |
|
14 |
Sorry, the conclusion is very likely wrong. True AMD wants several things |
15 |
that ATI brings to the table. But, there is little chance of Gfx being |
16 |
integrated fully into the cpu or even into a socket via a HyperTransport |
17 |
link. It's just too costly and the performance, for all but UMA (Unified |
18 |
Memory Archittecture) memory interface would suffer greatly. It's also |
19 |
one of the reasons AMD is adding a PCIe interface onto their cpus - the |
20 |
1207 pin Socket F (is that the correct socket?) models, later next year. |
21 |
|
22 |
Gfx, especially 3D, is about memory bandwidth. Move the memory out |
23 |
of direct contact with the gpu chip - say via a socket, and it's |
24 |
necessary to drop the frequency that the memory interface runs at. |
25 |
It's basic electronics - add more capacitaince and inductance, and |
26 |
the speed of the interface goes down. |
27 |
|
28 |
True, it's possible to put low-end, UMA based, graphics into a socket, |
29 |
but that's not really a long-term goal. Rather, just an interesting |
30 |
Engineering exercise. Plus, in laptops, where it would be best, the |
31 |
socket cost - both dollars and size, negates any benefit of moving |
32 |
away from it existing on the southbridge where it is today. |
33 |
|
34 |
|
35 |
> Now, in ordered |
36 |
> to fully populate the ecosystem, they'll /have/ to open things up. Keep |
37 |
> in mind that Intel is designing similar stuff with it's integrated |
38 |
> solutions, and if AMD didn't get in the game, it wouldn't be long until |
39 |
> they were no more relevant than Via in the x86 CPU market. |
40 |
|
41 |
Both AMD and Intel are moving, but not in the direction of more complexity. |
42 |
They are moving to simpler parallel execution units - like the parallel |
43 |
Gfx pipes found on GPU chips today, only for generic computing. All this |
44 |
time, cpus have had a very diffcult time moving from serial processing |
45 |
and the efforts for massivly parallel units have not faired well - Sun's |
46 |
last UltrsSparc. At the same time Gfx systems have become more and more |
47 |
parallel, and with the push from the late 3DLabs, have become more |
48 |
like generic, programable, compute blocks, though still diffcult to |
49 |
access from main menory. |
50 |
|
51 |
|
52 |
> |
53 |
> So... I think AMD/ATI /could/ open their video specs. However, it's still |
54 |
> an open question of whether they /will/, tho I think the chances are |
55 |
> pretty good, as Intel will have them between a rock and a hard place if |
56 |
> they don't. |
57 |
> |
58 |
|
59 |
Uh, no. Remember, Intel doesn't make real 3D Gfx chips. Unlike, Nvidia and |
60 |
Ati, Intel does most of it's 3D processing in software. Thus opening up |
61 |
the chips specs and driver has little impact on any IP outside of Intel as |
62 |
it doesn't expose any IP that might belong to Micrsoft or SGI. |
63 |
|
64 |
Both ATI and Nvidia have IP in their hardware that came from outside |
65 |
of those companies. Opening up their drivers exposes this IP and it's |
66 |
not theirs to give away. And given that some IP was sold outright |
67 |
to Microsoft, even if SGI were to open up the Gfx IP that is licensed, some |
68 |
would still be held up in Redmond, as would any implementation in hardware |
69 |
that insures decent performance with DirectX 9 and 10. So outside of the |
70 |
binary blobs that hide the IP, like Nvidia implements today, there will |
71 |
be little change in drivers from Nvidia. With luck AMD will get ATI |
72 |
to at least get their Linux drivers into a mroe suitable form on par |
73 |
with Nvidia. |
74 |
|
75 |
> (My sources for much of this were articles at ArsTechnica and the |
76 |
> Register, in turn quoting trade mags and industry analysts, plus AMD's own |
77 |
> PR.) |
78 |
> |
79 |
|
80 |
While ArsTechnica tends to be fairly reliable, I've found stories from the |
81 |
Register to be very, very, wrong - the assumptions were way off base or the |
82 |
reporter got the entire story wrong because they ignored documented history. |
83 |
I'd be very wary of believing a story in the Register without cross checking |
84 |
it with independent sources. Sometimes the Register will get one sentence |
85 |
in the story correct. Sometimes, the entire story is made up from that one |
86 |
sentence. |
87 |
|
88 |
Bob |
89 |
- |
90 |
-- |
91 |
gentoo-amd64@g.o mailing list |