GF108 rumours, this is a budget part?

[DanceswithUnix]

Fudzilla have pictures of the upcoming GF108:

http://www.fudzilla.com/graphics/gra...ized-and-sized

Die size of 127mm^2. Compare that to the GTX 460 with a die of 367mm^2, that is a ratio of nearly 3 so I guess this part will have 128 shaders (maybe some fused off for yield).

So my 8800GT is about to become a budget class board

[kalniel]

Your 8800GT has a 256bit memory bus.. I'm pretty certain this part won't!

[scaryjim]

This shouldn't surprise you: the GT240 has 96 shaders, after all. I doubt 128 will be the minimal shader card: this is likely to be the replacement for the 9600GT / GT240...

But yes. Let's not forget that ATI did this in one generation between the (320 shader) 3870 and 4670

[DanceswithUnix]

Your 8800GT has a 256bit memory bus.. I'm pretty certain this part won't!

But my 8800GT has an old DDR3 256 bit bus, doesn't sound like half width DDR5 is any penalty from looking at the ATI benchmarks.

[DanceswithUnix]

This shouldn't surprise you: the GT240 has 96 shaders, after all. I doubt 128 will be the minimal shader card: this is likely to be the replacement for the 9600GT / GT240...

But yes. Let's not forget that ATI did this in one generation between the (320 shader) 3870 and 4670

Doesn't surprise me at all, will just be amusing to see it happen.

In a way I wonder if all discrete graphics are being forced upmarket. Sounds like the AMD Fusion might be 5670 class integrated graphics, which will make the low end discrete look dumb if that is true. If AMD are indeed aiming that high, then no doubt Intel will be forced to step up their game and an 8800GT class board isn't much of an upgrade!

[scaryjim]

Sounds like the AMD Fusion might be 5670 class integrated graphics, which will make the low end discrete look dumb if that is true.

It'll be frakking amazing if that's true! Can't see it though - it'd make no sense to have their last integrated graphics chipset run 40 shaders, then suddenly jump up to 400 for the on die one! Plus the 5670 dissapates something like 65W on its own - add that to a processor and you're talking a phenomenal TDP... and add to that the fact that it would completely cannibalise sales for their low-to-mid range cards...

it's a lovely rumour, but I really can't see it...

[kalniel]

The 40nm 5670 is 65W TDP...
Llano is 32nm GPU, up to 25W per core. So it's not all that far off.

But I think it could be a slower part equivalent.

[scaryjim]

*thought* on-die GPU based on 5650 Mobility, perhaps? Already in the wild, 40nm, 400 shader, 450MHz - 650MHz clock speed, and 15W - 19W TDP. *that* I could see being a possibility.

Although I still reckon that would cannibalise their discrete card market too much - wonder if there'll be a range of GPU options available...

[DanceswithUnix]

I must admit I will believe it when I see it.

I mostly don't see how you can feed lots of GPU and CPU off the existing AM3 DDR3 bus, it would have to go way faster or wider.

It also doesn't tie in with the 890 chipset. Lets face it, 890 IGP is rubbish and I assumed it is to make the Fusion kit look good when they come out. AMD have been doing 40 shader chipsets now for ages, I really expected this generation to double that so why else did they hold back?

[kalniel]

I mostly don't see how you can feed lots of GPU and CPU off the existing AM3 DDR3 bus, it would have to go way faster or wider.

I don't think that's an issue - intergrated chips have been using shared RAM for years, and while it's not especially fast, it's fast enough for what they're used for. Having the GPU integrated closely into the CPU memory controller+crossbar is going to provide much better access to RAM than previous shared RAM solutions.

Lets face it, 890 IGP is rubbish and I assumed it is to make the Fusion kit look good when they come out.

Interesting call.

[scaryjim]

AMD have been doing 40 shader chipsets now for ages, I really expected this generation to double that ...

780G released 4th March 2008, so more than 2 years. I was expecting the 8x0 series chipsets to ship with a 4350-based IGP, so 80 shaders (and the option of up to 256MB sideport, but that was almost certainly wishful thinking ). I can only assume that they were unable to shrink the process sufficiently to fit that many shaders within the TDP tolerance of standard chipset cooling (although given the small passive coolers on most 4350 cards that doesn't ring *entirely* true) *and* get sufficient yields to make such a chipset profitable...

[DanceswithUnix]

Was trying to dig up the rumour sites that indicated Llano performance. Only came up with these two, I'm sure there has been more that this though.

Firstly, for the Ontario variant aimed at Atom type systems:
"Performance will be essentially what you see in the low-end mainstream notebook chips today"
then "Llano will bring improvements to both CPU power, and GPU power"

http://www.semiaccurate.com/2010/07/...etails-galore/


Secondly, Llano rumoured to use Redwood graphics core:
http://www.fudzilla.com/graphics/gra...dwood-graphics

Now, by "low end mainstream notebook" for Ontario do they mean 3200 IGP, our old 40 shader friend?

[scaryjim]

Now, by "low end mainstream notebook" for Ontario do they mean 3200 IGP, our old 40 shader friend?

“Performance will be essentially what you see in the low-end mainstream notebook chips today”
<- snip ->
A “low-end mainstream” chip today is the Athlon II Dual-Core Mobile

So it looks like SemiAccurate are assuming the performance quote relates to CPU performance, rather than GPU performance, about which they say:

Graphics performance of Ontario will be a “Game changer” in the industry (according to AMD)

Which frankly could mean anything, since these devices are targetting Atom / GMA3150: hardly a sterling graphics performer...

[Blaineoliver]

They showed a demo on llano that ran an AVP game solid 30fps.

[DanceswithUnix]

So it looks like SemiAccurate are assuming the performance quote relates to CPU performance, rather than GPU performance, about which they say:



Which frankly could mean anything, since these devices are targetting Atom / GMA3150: hardly a sterling graphics performer...

But AMD already push 40 shader chips for netbook use (the Ferrari One), so I think we can assume Ontario is better than that and Ontario is the low end version.

More Llano stuff, Llano has 1B transistors (from the title):
http://www.anandtech.com/show/2871

and the CPU part is Phenom II core with zero L3 cache, so basically an Athlon II X4
http://www.anandtech.com/show/2933

and and Athlon II X4 has ~300M transistors
http://www.anandtech.com/show/2836

so that leaves 700M transistors for graphics?

Redwood 5570/5670 uses 627M transistors
http://www.anandtech.com/show/2935

Now that sum includes stuff like duplicate memory controllers and comes up ~100M short, easily soaked up with some cache somewhere.

[kalniel]

And the crossbar?

Don't forget llano CPU has double the L2 cache of the X4 - it's more like 2x X2's (2x234m trannys) in that respect.