The FULL Llano die

[IntelUser2000]

http://www.semiaccurate.com/forums/showthread.php?t=1076

Post #2, presentation titled "Now the Fusion Era of Computing Begins" shows the FULL Llano die without the GPU being cut-off as in the bigger die picture spread everywhere on the web.

Xbitlabs reports that the on-die IGP will feature 6 SIMD clusters containing 80 SPs each, or 480SPs in total.

 

[jvroig]

on-die IGP will feature 6 SIMD clusters containing 80 SPs each, or 480SPs in total.

Can you or anybody put that in context? Is that comparable to a 43xx, 45xx, or 47xx?

 

[Idontcare]

jvroig it depends entirely on the clockspeed of those SPs as well as the actual bandwidth (and size) those SPs ultimately have access with and to.

But at first glance it does appear to be quite capable of being formidable as a lower-end mainstream GPU, the hardware is there, but as we see with Cypress vs RV770 the presence of superior hardware statistics does not necessarily ensure the performance will follow.

 

[IntelUser2000]

Well, the Radeon 4770 and 4830 has 640SPs, and 4670 and 4650 has 320SPs. And the Radeon 5750 has 720SPs. The memory bandwidth is similar to the 4650/4670 series, and the architecture is rumored to be Radeon HD5x00 derivative. The rogue report of being approx. 4700 performance might be true.

Anyhow, its an IGP that doesn't look like one.

 

[jvroig]

Thanks, IDC. So as of now, not enough info to tell?

but as we see with Cypress vs RV770 the presence of superior hardware statistics does not necessarily ensure the performance will follow.

BFG10K seems to think AMD's simply holding out with the drivers for it. Or that it's just a driver problem, even if not intentional.

 

[jvroig]

Well, the Radeon 4770 and 4830 has 640SPs, and 4670 and 4650 has 320SPs. And the Radeon 5750 has 720SPs. The memory bandwidth is similar to the 4650/4670 series, and the architecture is rumored to be Radeon HD5x00 derivative. The rogue report of being approx. 4700 performance might be true.

Thanks! Still, it's just hardware and incomplete info, but at least it looks good to start with.

 

[Idontcare]

IntelUser2000 how did you come by memory bandwidth estimates? Is that speculation or has AMD actually alluded to memory bandwidth for the SPs somewhere?

I only ask because this is something I have been curious about ever since the idea of putting the GPU into the same socket as the CPU has been bandied about...all is well and good until you need >100GB/s bandwidth to that socket and the mobo is setup to only access DDR3 memory.

So if memory bandwidth is indeed going to at 4650/4670 levels then what manner of configuration of DDR3 speeds and channels are they going to have to implement to achieve that?

 

[jvroig]

I didn't notice at first that he mentioned bandwidth estimates (I just glossed over his post looking at SP numbers, sorry!) That is interesting if true, that would be one heck of a powerful IGP. a 4770 is certainly no slouch, in essence we'd get an 8800/9800GT power in an IGP... sounds too good to be true. IGPs have always been underpowered so as not to cannibalize the sales of discrete GPU cards, at least as far as I understood it.

 

[Idontcare]

BFG10K seems to think AMD's simply holding out with the drivers for it. Or that it's just a driver problem, even if not intentional.

That's actually a theory I put out there the day Cypress was released, not saying BFG10K didn't rationalize it on his own (great minds think alike [/laugh]), just saying I am well acquainted with the theory and the rationalizations behind it.

My comments above regarding the RV770->Cypress scaling in specs not being mirrored in the scaling of performance are made with the conservative assumption that AMD really has no cunning strategy up their sleeve regarding the drivers.

We either believe in the data we have and use it for extrapolation (as I have done) or we believe something far more complex is afoot and then attempt extrapolation at risk of incurring far larger/fatter error bars. For now, when it comes to Llano I am opting for the conservative estimations until AMD does something to prove otherwise is warranted.

 

[jvroig]

Understood. I am rather pessimistic about it myself, so as not to be too disappointed if it does not materialize.

 

[IntelUser2000]

I was half expecting someone to ask that.

Well, Llano is said to have Dual Channel DDR3-1600 memory controller, which turns out to be 25.6GB/s memory bandwidth. It's not the Llano itself has lots of bandwidth, its that Radeon 4650/4670 has low amount of bandwidth for a modern GPU.

http://www.anandtech.com/video/showdoc.aspx?i=3405
http://hi-techreviews.com/index.php/review-links-mainmenu-80/9511-sapphire-hd4670-review

4650: 1GHz 128-bit for 16GB/s bandwidth
4670: 1.8-2.2GHz 128-bit for 28.8 to 35.2GB/s bandwidth

It's really not that far off.

 

[21stHermit]

"Now the Fusion Era of Computing Begins" shows the FULL Llano die without the GPU being cut-off as in the bigger die picture spread everywhere on the web.

All well and good, but I also read 2011, not exactly the day after tomorrow.

Intel will be shipping 32nm IGP in Jan 2010 (Pine Trail, Clarkdale & Arrandale). But in package not on die, as is their practice. Not clear if we will see 32nm on die IGP, likely at 22nm, 2012?

 

[IntelUser2000]

Nah, that'll happen with Sandy Bridge, the tock that's coming by end 2010/early 2011.

Indeed the IGP is becoming powerful. But this I think, is only the shape of things to come. Memory bandwidth on the on-die GPUs will be fully solved by Haswell timeframe with mass eDRAM integration with hundreds of gigabytes to a terabyte of bandwidth. By then, IGPs might be able to perform like mainstream performance graphics cards by that timeframe. Now, it might be a hint of why Nvidia is becoming frantic.

When Clarkdale arrives, you'll see Intel isn't too incompetent in graphics either.

Update:

Intel will be shipping 32nm IGP in Jan 2010 (Pine Trail, Clarkdale & Arrandale). But in package not on die, as is their practice.

By the way, PineTrail uses on-die IGP, a heavily modified GMA950 named GMA3150.

 

[cdbular]

I am also concerned about the memory bandwith. I think that if they can use triple channel and use a dedicated channel for the GPU and the two others shared for CPU/GPU, there are many chances of getting 4650 performance. Not high end performance , but remarkable for an integrated video.

Unlike inteluser2000, I am not too confident in intel compentence in graphics. They still lack from high end experience in that area. ATI has been making high performance 3D chips for a while.

 

[jvroig]

Not high end performance , but remarkable for an integrated video.

Definitely, because no one who makes discrete video cards will want to make an IGP that performs like a high-end part. That's just crazy because you will effectively obsolete your own business, unless you were really planning on shutting down your discrete GPU business anyway.

 

[cdbular]

Definitely, because no one who makes discrete video cards will want to make an IGP that performs like a high-end part. That's just crazy because you will effectively obsolete your own business, unless you were really planning on shutting down your discrete GPU business anyway.

That is so true. I don't think AMD will kill their high end parts. However, you get what you pay for and I don't see integrated graphics performing as high end parts for low price in a near future, that would require special plataform with dedicated graphics memory. So I don't see intel completing with high end GPU cards, not because they don't want (as they don't have discrete GPU parts to kill), it's just because they can't do it at competitive prices.

 

[21stHermit]

Nah, that'll happen with Sandy Bridge, the tock that's coming by end 2010/early 2011.

I'm not as up to date on the road map.

Indeed the IGP is becoming powerful. But this I think, is only the shape of things to come. Memory bandwidth on the on-die GPUs will be fully solved by Haswell timeframe with mass eDRAM integration with hundreds of gigabytes to a terabyte of bandwidth. By then, IGPs might be able to perform like mainstream performance graphics cards by that timeframe.

You're way ahead of me on this.

Now, it might be a hint of why Nvidia is becoming frantic.

I've noticed, even talk of x86.

When Clarkdale arrives, you'll see Intel isn't too incompetent in graphics either.

Planning on that. I'm going to be a Clarkdale early adopter. Hope to do both a drastic performance boost Atom 330 -> Core i3-530 at a significant idle power drop: Atom @ 37W, i3-530 will get a picoPSU and/or a laptop HDD + Win7.

Update:

By the way, PineTrail uses on-die IGP, a heavily modified GMA950 named GMA3150.

Are you sure, linky please?

 

[21stHermit]

Definitely, because no one who makes discrete video cards will want to make an IGP that performs like a high-end part. That's just crazy because you will effectively obsolete your own business, unless you were really planning on shutting down your discrete GPU business anyway.

Not true, we long ago learned that if possible, far better to obsolete your own product than have your competition do it before you.

 

[jvroig]

Not true, we long ago learned that if possible, far better to obsolete your own product than have your competition do it before you.

Yes, but if a company knew that was coming, they'd be planning to shut down that part of the business already.

The fact that AMD has this new marketing "Velocity" thing where GPU architectures improve every year implies they aren't dismantling their high-end graphics division in favor of just releasing mega-IGPs.

If mega-IGPs become the trend, then AMD will (should) follow suit, but that's not right now, and that's obvious since they still have big plans for their discrete GPU line of products.

 

[IntelUser2000]

Are you sure, linky please?

http://www.techfuels.com/latest-hardware-news/22665-drawings-new-atom-called-gma-3150-a.html

 

[jvroig]

The reason mega IGPs start making sense is because as IGPs become more acceptable, the less financial incentive is to develop a discrete one.

Yes, and that's exactly why AMD won't go there now, even if they could, because their discrete GPU division is currently "kicking ass". When it starts to become less profitable, and their projections for such mega-IGPs become brighter, they'll probably turn that corner fast.

As for Intel? If Intel could actually do it, I'd say they'd have done it already, but so far all their IGPs are crap. Larrabee hasn't exactly impressed anybody yet, even at IDF. If they are wanting to make a mega-IGP in order to obsolete the discrete GPU business (which they don't have, so only AMD and nVidia will lose), I'd say they aren't doing so well.

Not saying it's impossible for them to do so, but right now, using existing products as basis, AMD is in a better position to start the mega-IGP trend, and they sure as hell won't do that yet.

 

[21stHermit]

http://www.techfuels.com/latest-hardware-news/22665-drawings-new-atom-called-gma-3150-a.html

Sorry, nothing in there saying on die GPU. My only conclusion from that paragraph was that the authors primary language was NOT English.

I'll look and see if I can find a link supporting my original assertion: In package IGP.

 

[21stHermit]

By the way, PineTrail uses on-die IGP, a heavily modified GMA950 named GMA3150.

Well we have this quote from someone by the name of Anand Lal Shimpi from here http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3565

The CPU in this case is the next version of the Atom core. Still based on Intel’s 45nm process, this new Atom (codenamed Pineview...yes, confusing) brings both the memory controller and GPU core on-die.

Look as if you're correct, first time for everything.

So what happens when you have a dual core Pine Trail? Do you just disable a MC & GPU?

 

[IntelUser2000]

Look as if you're correct, first time for everything.

Har har. Weren't you the one claiming Clarkdale's IGP and Pinetrail's IGP is one and the same.

 

[21stHermit]

Not saying it's impossible for them to do so, but right now, using existing products as basis, AMD is in a better position to start the mega-IGP trend, and they sure as hell won't do that yet.

Clearly AMD/ATI is a dominant graphics player, Intel is not. Larrabie? is just vaporware. But this thread started with Llano/Fusion and that's 2011, also vaporware. It will be interesting to see how your "mega-IGP" plays out.

So Intel will have shipping in-package/on-die GPUs in three product lines, Pine Trail, Clarkdale & Arrandale, in Jan. 2010, that's not vaporware. I'd be surprised if Anand didn't have at least one of those on his test bench, under NDA, as we speak.

For me the IGP in Pine Trail, Clarkdale & Arrandale is of more interest, I'm concerned with low power, not high performance. Consider that Intel has over one-million transistors devoted solely to power management on many of its latest chips, pretty incredible.