1 |
Duncan, mused, then expounded: |
2 |
> Bob Sanders <rsanders@×××.com> posted 20060814151338.GA102938@×××.com, |
3 |
> excerpted below, on Mon, 14 Aug 2006 08:13:38 -0700: |
4 |
> |
5 |
> > Gfx, especially 3D, is about memory bandwidth. Move the memory out |
6 |
> > of direct contact with the gpu chip - say via a socket, and it's |
7 |
> > necessary to drop the frequency that the memory interface runs at. |
8 |
> > It's basic electronics - add more capacitaince and inductance, and |
9 |
> > the speed of the interface goes down. |
10 |
> |
11 |
> Two factors in counter-point. We're talking FB-DIMM tech time this comes |
12 |
> out, and this will be using upgrades to the same memory controller on the |
13 |
> CPU (only in this case GPU) AMD is already using to maintain its lead in |
14 |
> multi-socket performance against Intel (even while Intel's Core-Duo has |
15 |
> caught up at the low single-socket end). |
16 |
> |
17 |
|
18 |
Problems with this approach - it takes 4 FB-DIMMs to a special northbridge |
19 |
for Intel to acieve 21.5 GB/s bandwidth to the cpu in a Woodcrest based server. |
20 |
And that's a fair chunk of change for the northbridge and dimms. Plus the |
21 |
increased latency. Btw - the current DDR-2 controller in the AMD Socket AM-2s |
22 |
can talk to FB-DIMMs. It's not allowed to yet, probably not tweaked up right |
23 |
and AMD like to take things slowly so that risk is minimized - one or two |
24 |
things at a time. But the current memory controller should work |
25 |
with FB-DIMMs. |
26 |
|
27 |
While 21.5 GB/s may seem fast, that's the peak speed, not the sustained |
28 |
bandwidth. Effective bandwidth is going to be less. And it's a lot |
29 |
cheaper to build a dedicated PCIe card with memory than to even purchase |
30 |
one stick of FB-DIMM memory. Even with just 2 DIMMS, DDR-2 or FB-DIMMS, |
31 |
the memory bandwidth peaks around 10.5 GB/s with a sustained bandwidth |
32 |
of around 8 GB/s. Still less than the 12.8 GB/s an Nvidia 600GS card |
33 |
attains while retailing for $109. |
34 |
|
35 |
> > Remember, Intel doesn't make real 3D Gfx chips. Unlike, Nvidia and |
36 |
> > Ati, Intel does most of it's 3D processing in software. Thus opening up |
37 |
> > the chips specs and driver has little impact on any IP outside of Intel as |
38 |
> > it doesn't expose any IP that might belong to Micrsoft or SGI. |
39 |
> |
40 |
> Does <> will do. With aeroglass driving dx10 requirements, the low end of |
41 |
> the gfx market is about to get very legacy, very fast. Their new chips |
42 |
> do more in hardware because they must, due to the performance |
43 |
> requirements. That's the reason Intel's latest free driver release made |
44 |
> the headlines in the FLOSS world -- it's the first time we've had full |
45 |
> sources to something that capable. |
46 |
> |
47 |
|
48 |
|
49 |
Based on all I've read, what I'm putting forth is strictly my take on the |
50 |
situation. No source to go to, just speculation. What Intel is doing |
51 |
is stopping Gfx development any laying off most of it's Gfx Engineering |
52 |
staff. They are outsourcing most of the IP development by licensing |
53 |
cores from Imagination Technologies - the PowerVR SGX, that has been widely |
54 |
reported, but not explicitly confirmed (unless I missed it). As the SDK |
55 |
for the PowerVR is already publically available, all Intel did was go |
56 |
and say - all this old crap we made and are abandoning, well - here are |
57 |
the specs for it. And since the new cores already have an open api and |
58 |
sdk, we'll just "say" we're opening the development to the Open Source |
59 |
Community, because, well, it already is. Great Marketing move. |
60 |
|
61 |
Of course I could be wrong on this, but time will tell. |
62 |
|
63 |
Bob |
64 |
- |
65 |
-- |
66 |
gentoo-amd64@g.o mailing list |