Nvidia's Hybrid SLI

[PSU-failure]

Thats basically the same as PowerXpress which has been talked about with AMD's Puma platform. The issue is that with Vista you can only have one graphics driver loaded, so switching between on and the other on-the-fly is only possible if the same driver can operate all the devices you are switching between.

In fact it's quite easy as long as you consider the IGP as a DX9/2D chip and the GPU as a DX10 chip.

Look at your graphics drivers list, you'll see 3 related to graphics:

- 2D driver
- D3D driver
- OGL driver

That's with XP, I think Vista has 2 different D3D drivers (D3D9 & D3D10), which is just fine in this configuration.

Set the 2D and D3D9 driver as related to the IGP (D3D9 for Aero compatibility) and D3D10/OGL as related to the GPU which no more has any 2D use, transfer the framebuffer to the IGP when rendering is done and all is ok.

Assuming 120fps throughput, you'll need 1.83GB/s copy rate between GPU and IGP at most, which won't be a bottleneck on a 16x PCI-E 2.0 slot, and that's assuming there's no data compression.

 

[Razor1]

You know, I really want the head of Alfredo Garcia. Or whoever came up with the idea of doing GPU Physics nearly exclusively through Havok FX. And I also want the head of whoever came up with the idea of telling people that it was more efficient to do it on a separate board than on the same one, unless limited by the amount of video memory.

The whole GPU Physics thing has been so horribly mismanaged that I truly am at a loss of words regarding the whole thing. The fact that it was made nearly exclusive to a presumably expensive add-on for an expensive middleware solution is downright ridiculous. I know Havok did some of the R&D, but FFS, come on... This is laughable at best.

Although my current signature ("you can't compete with free.") is not related to GPU Physics, I feel that it also applies just about perfectly to its core problem.

Heh, I think Crysis will support your statement very well , really with multicore CPU's, physics and ai can be done on seperate threads and that alone alievates much of the issues we have seen in previous generation games. Of course small things like particle physics are nice on a gpu, but I don't see much else that is an absolute necessacity.

 

[INKster]

I think Vista has 2 different D3D drivers (D3D9 & D3D10), which is just fine in this configuration.

How do you explain Vista's support for non-DWM managed desktop modes, like AERO Basic or Classic ?
The 2D driver is still there for compatibility reasons. It's simply downplayed by MS.

 

[PSU-failure]

How do you explain Vista's support for non-DWM managed desktop modes, like AERO Basic or Classic ?
The 2D driver is still there for compatibility reasons. It's simply downplayed by MS.

M$ still uses the 2D driver as an HAL, just as it has ever been.

 

[Twinkie]

http://www.vr-zone.com/?i=5126

NVIDIA is targeting to launch their first GeForce 8 DX10 UMA chipset called MCP78 for AMD processors in November. MCP78 support the new Hybrid SLI technology, HyperTransport 3.0 and PCI Express 2.0. NVIDA hopes to create a new category of motherboard graphics that offers entry level discrete GPU performance but cheaper. NVIDIA touts a Hybrid performance boost up to 40% by pairing up with a mainstream discrete graphics card, ~20% with a performance card and ~5% with enthusiast card.


There will be 3 variants; MCP78 U, MCP78 S and MCP78 D and all supports 1x16 PCIe 2.0. MCP78 U is priced at US$70-80 where its 3DMark performance on par with entry level graphics card and PureVideoHD support. MCP78 S is priced at US$55-65 with decent graphics performance and PureVideo support. Lastly, MCP78 D priced at US$55-65 has no int. graphics capability and thus no Hybrid SLI and PureVideo support. NVIDIA plans to deliver A01 samples to customers in August and A02 samples in October where mass production and launch slated for November.

Pretty interesting as this means that you could get 8600GT performance with a 8500GT. (save a few bucks )

Looks like november is the big date for nVIDIA this year with mcp73/72 and G92 all released on that month.

 

[Silent_Buddha]

Interesting tech, but unless it supports multiple displays while in SLI mode it's useless for me. Likewise until SLI/Crossfire don't require profiles...it's more of a bother than a benefit to me.

Is there any technical reason that 3DFX style SLI (scan line interleve) isn't possible with current 3D acceleration?

Regards,
SB

 

[ShaidarHaran]

Interesting tech, but unless it supports multiple displays while in SLI mode it's useless for me. Likewise until SLI/Crossfire don't require profiles...it's more of a bother than a benefit to me.

tru.dat
I used to run a multi-mon CF setup and I had more bugs because of it than any other config I'd previously run. There's got to be a way to overcome this problem...

Is there any technical reason that 3DFX style SLI (scan line interleve) isn't possible with current 3D acceleration?

Regards,
SB

Sure, but why?

 

[Davros]

because it worked well i guess..

 

[Arun]

Looks like november is the big date for nVIDIA this year with mcp73/72 and G92 all released on that month.

Indeed, but I heavily question the existence of what VR-Zone calls 'C72' and 'C73'. My guess is that what they refer to as the former is really MCP72+C55, while the latter is MCP72+C??, where C?? is the next-gen discrete memory controller chip from NVIDIA Bangalore.

As for Hybrid SLI on Intel chipsets, I'll admit not to really understand what they can do without either using three chips or skipping a cycle. Unless, of course, C?? actually includes an IGP on the northbridge... Hmm!

 

[trinibwoy]

because it worked well i guess..

It worked well with a vastly simpler API and rendering techniques. I'm baffled why people think it will magically work with all the advances in those areas since the 3dfx era.

 

[Dave Baumann]

Scan Line Interleave was an analogue process. By the time 3dfx moved to a digital SLI system, with VSA-100, we were working on quads of pixels at that point and SLI actually became BLI - changing from scan lines, to bands of pixels rendered on each board. In effect this is little different from from the scissor modes on Crossfire/SLI or even the tiled mode of Crossfire.

 

[LunchBox]

That's a very cool concept!
coz nothing gets wasted in terms of contributing performance.

 

[AlexV]

It worked well with a vastly simpler API and rendering techniques. I'm baffled why people think it will magically work with all the advances in those areas since the 3dfx era.

Because it's magic, why else?I mean, how can you miss such an obvious thing

 

[AnarchX]

http://my.ocworkbench.com/bbs/showthread.php?p=421362#post421362

So you have to put your heating on this winter.

 

[Arun]

Woah, so MCP72 includes a G98-like IGP too, and the only real difference between MCP72 and MCP78 is that the latter is less feature-rich? That does explain quite a few things if true, and makes a lot of financial sense.

I can't help but ponder what the actual *cost* savings between MCP72P and MCP78 would be then, though. I mean, the only real difference I can think of is 4 SATA ports vs 6 SATA ports, but that isn't exactly very significant... So call me skeptical on this one for now.

 

[AnarchX]

This should be the secret of cost-saving:

http://www.fudzilla.com/index.php?option=com_content&task=view&id=3330
http://my.ocworkbench.com/bbs/showthread.php?p=421355#post421355

... the needed PCIe-lanes for 3x x16 through an external-chip.

And not to forget, this chipset is supposed to made in 65nm.

 

[Rufus]

Kind of interesting how the BR chips have gone from AGP/PCIe bridge (BR02) to PCIe switch on the GX2 (BR03) to now PCIe switch on a motherboard (BR04, on both this and I assume Intel's Skull Trail).

Edit: Apparently Skull Trail is using "NVIDIA nForce 100 MCP" and is PCIe 1.0 while this is "NF200" and PCIe 2.0. I guess Skull Trail uses BR03 while 780i uses BR04.

 

[Arun]

This should be the secret of cost-saving

That does explain the difference between MCP72 XE and MCP72 P, both in terms of features and cost... But it doesn't explain the cost difference between MCP72P and MCP78... All of this still doesn't make much sense to me, really.

 

[AnarchX]

MCP78 has maybe fewer ports and is already manifatured in 55nm?
http://vr-zone.com/index.php?i=5126

And nF780a should use two chip-setup: MCP72+C72.

 

[Arun]

Hmm, well AFAICT:
- 750a is MCP72.
- 780a is MCP72+BR04.
- 750i is MCP72+C55.
- 780i is MCP72+C55+BR04.
- 790i is MCP72+C73.

Based on this, it should also be kind of obvious that 780i is a stop-gap solution. The 790i should, in theory at least, be both cheaper and faster.

Anyway, what I'm asking is where the cost and feature differences between the 750a and the MCP78 SKUs lies. I was presuming that MCP72 didn't have an integrated IGP, but it does seem to have one now, so I'm rather confused... AFAIK, both MCP72 and MCP78 are 65nm, but then again maybe they decided to switch process last-minute...

Even then, this doesn't give us a big feature/value difference between MCP72 and MCP78 single-chips. Unless mayb the MCP72 IGP is faster... That wouldn't make a lot of sense though in my eyes.