AMD has DX10 IGP and 'Hybrid CrossFire' working

[B3D News]

The lads at [H] seem to have temporarily had a new toy to play with. It was an AMD "Hybrid CrossFire" system sporting a working DX10 IGP chipset cooperating with an RV620-based discrete graphics card.

Read the full news item

 

[Geo]

And in the spirit of the Christmas season, I shall even sign up for Kyle's wishlist:

We have given AMD and NVIDIA a wish list of sorts. What I would truly like to see is an integrated IGP that will run all my desktop and video applications while I have a single or double high end video gaming card configuration powered down. I don’t run a multi-card configuration now for two simple reasons. First, I don’t want 200 watts worth of idling 3D cards sitting under my desk all day. Second, I don’t want to deal with the multi-monitor situation that continually forces me to reboot when toggling between multi-monitor desktop use and gaming on a single monitor. I want enthusiast level Hybrid that fully turns off (or incredibly close to it) my high end 3D gaming card when I am not using it. I also don’t want to juggle reboots every time I want to get a gaming session on.

 

[Dave Baumann]

Already announced or shipping elements actually mostly address those concerns. We already announced Extended Desktop support for Crossfire, which will be coming in the new year - this addresses the issue of dual monitors with Crossfire gaming on a single card. Secondly, PowerPlay for desktop also goes a considerable amount towards reducing the power consuption of high end cards when they are mostly idle; this of course still operates in Crossfire configurations so both cards can get into their lower power states at the desktop.

 

[tEd]

...and with SLI there's no need to make a restart to switch mulit-monitor/SLI although multi-monitor while SLI would be nice indeed.

 

[Zaphod]

There is also a piece of discrete RAM on the motherboard that can be offered in either 16MB or 32MB capacities that acts as part of the framebuffer and is within close proximity to the IGP to remedy latency issues.

Is that (latency) really the main reason? I would have thought the primary use for such a minute piece of local ram would be for the graphics core to alway control one coherent display buffer regardless of what other pieces of the system does (CPU sleep states, shutting down buses, memory access, or whatnot). In other words: Power saving.

 

[DavidC]

Is that (latency) really the main reason? I would have thought the primary use for such a minute piece of local ram would be for the graphics core to alway control one coherent display buffer regardless of what other pieces of the system does (CPU sleep states, shutting down buses, memory access, or whatnot). In other words: Power saving.

That's also what Anandtech said with their Xpress 200 chipset review: http://www.anandtech.com/showdoc.aspx?i=2269&p=19

 

[Rufus]

Is that (latency) really the main reason? I would have thought the primary use for such a minute piece of local ram would be for the graphics core to alway control one coherent display buffer regardless of what other pieces of the system does (CPU sleep states, shutting down buses, memory access, or whatnot). In other words: Power saving.

It's a combination of a number of factors, all based around how much an integrated memory controller makes integrated graphics suck. In the old days an IGPU was sitting next to the northbridge's memory controller, so it was easy for it to get any memory it wanted and reasonably good bandwidth. Now the MC is behind a CPU, meaning the IGPU has to ask the CPU for everything.

This arrangement gets you really strange request patterns. Like CPU reading from the frame buffer requires: CPU read from IGPU, IGPU send read request to CPU, CPU reads from MEM, CPU responds to IGPU, IGPU responds to CPU. Bounce more than a few of those requests around and you're asking for extreme latency, deadlocks, satturated connections, and all kinds of badness.

16-32MB of local memory would fit the entire FB and a reasonably sized cache. Having the FB local should substantially improve performance, lower bandwidth usage / power, and simplify the design.

This is all nice and good except for the obvious problem: IGPUs target the near-0 performance / near-0 cost segment. Will anyone care about these benifits if the mobo's $10 more to build?

 

[Geo]

This is all nice and good except for the obvious problem: IGPUs target the near-0 performance / near-0 cost segment. Will anyone care about these benifits if the mobo's $10 more to build?

Not if its aimed at the value segment. ATI found that out with less than 20% margins back in the day for chipsets when they figured "a better mousetrap" must surely command a better price. Value market said, "umm, no".

The more interesting point is does it save you some costs elsewhere to get certified for everything you want to be certified for? If it just moves the dollars around then that's a different calculation.

 

[Tim Murray]

This is all nice and good except for the obvious problem: IGPUs target the near-0 performance / near-0 cost segment. Will anyone care about these benifits if the mobo's $10 more to build?

If I could run hybrid CF/SLI and not run an 8800GT/GTS/3870/3850 in Aero, I would absolutely buy one of these.

 

[BRiT]

Does anyone know the details of "Hybrid Crossfire"? Are they still using AFR or did they get SFR or SuperTiling to work?

Given the large disparity between an IGP and a Discrete card, am I right in thinking that SuperTiling or SFR would give a better benefit than AFR?

I hadn't fully thought this through, so here's me whipping something up on the fly. Lets use the performance ratio between them is 4 to 1.

In AFR, whenever it's the IGP's turn to render there would be horrible input lag. If the Discrete card could render a frame in X milliseconds, the discrete card would take 4x milliseconds. During that 4x milliseconds, the discrete card could have rendered 4 frames (an additional 3 frames). The discrete card will be waiting for the IGP to finish rendering.

In SuperTiling, the screen could be broken out into 5(4+1) tiles. When the discrete card finishes with its 4 tiles, the IGP will be finished with it's 1 tile. That could/would bring a net benefit. That would also prevent the Discrete card from having to wait on the IGP to finish.

So where is the flaw in that on-the-fly off-the-cuff thinking? I don't see how Hybrid Crossfire can help unless the Discrete card is spec'd as lowly as the IGP. Or is that not the market they're aiming this at?

 

[Arun]

So where is the flaw in that on-the-fly off-the-cuff thinking? I don't see how Hybrid Crossfire can help unless the Discrete card is spec'd as lowly as the IGP. Or is that not the market they're aiming this at?

In AMD's case, from my understanding, that is indeed the market they're aiming at. Think about it this way: if an OEM uses their IGP, the buyer can choose to add in a RV620 and get twice the performance instead of just a few percentage points more.

Given how many configs at OEMs are based on an IGP to which you can add a discrete GPU, this is a fairly big advantage by itself. Of course, getting a 20% performance advantage on a mid-range card would be appealing too for OEMs and the market in general, and NVIDIA has claimed that's what they're doing - but they haven't given any details about it yet, nor has there been any public demonstration of this ability.

 

[Rufus]

If I could run hybrid CF/SLI and not run an 8800GT/GTS/3870/3850 in Aero, I would absolutely buy one of these.

How close are current IGPUs to Aero compatible? I could see an argument that making an IGPU Aero compatible would be worth some extra cost. However won't most IGPU-based systems be shipping with Vista Basic which lacks Aero anyway?

So where is the flaw in that on-the-fly off-the-cuff thinking? I don't see how Hybrid Crossfire can help unless the Discrete card is spec'd as lowly as the IGP. Or is that not the market they're aiming this at?

Doesn't SFR do automatic load-balancing between the two cards? Assuming the load balancing works on identical cards with very biased scenes (empty sky vs a city), then load balancing different performance levels shouldn't be much harder.

 

[Dave Baumann]

How close are current IGPUs to Aero compatible? I could see an argument that making an IGPU Aero compatible would be worth some extra cost. However won't most IGPU-based systems be shipping with Vista Basic which lacks Aero anyway?

Virtually all recent (i.e DX9) integrated are Aero compatible, and Vista Premium Logo compliant - IGP's likes 690G run Aero very well in fact.