Fusion die-shot - 2009 Analyst Day

Memory access by the HT interface would require external memory controller ASIC, which will drive the implementation costs up.
Those HT links have some special even probably sacred meaning.

 

I'm thinking if its not for the GPU, then possibly for a Magny-Cours type setup to go into some OpenCL compute based mega-computer.

Thats partly why I was hoping someone could say which socket it'll be on.

Hmm, but now I think about it, with no L3 & so no cache snooping thingie (that is supposed to be what will make Magny-Cours work nicely) that wouldn't work very well

 

After one more cearfull visual disecting of the die shot, it looks like only one type of HT physical interface is present in there, being Rx or Tx pad-row type, but not both! If you look in any other AMD K8 or K10 hi-res die shot, you'll spot the distinctive Rx and Tx pad rows:

 

Maybe it's a copy/paste job?

 

http://realworldtech.com/page.cfm?ArticleID=RWT120409180449&p=2

AMD definitely has a window of opportunity there with fusion, question is whether they can get it out the door in time. A 480 SP part will nuke anything Intel has to offer in it's IGP's.

 

Fusion will make use of hypertransport 3.1 and will benefit from ddr3 PC12800, so there's a bit of progress regarding bandwith as well. we can guesstimate it's similar to a ddr3 GT240, 4670 or 5650 in terms of performance.

very good : it qualifies as a gaming PC in my eyes, and the CPU part is good for photoshop, music software, 3D content creation etc.

I will sure recommend it. pair it with a 350W PSU, overclock it and you have awesome bang for the buck.

 

I find this questionable, Intel could end with some SCC cores instead of larrabee.
In the low end market SCC cores may well be better/easier as general purpose accelerator than larrabee or GPU cores, who know ow end market could go away with software renderer, I would like to have Nick opinion in this regard for example. But Intel has clearly let AMD a head start by giving up on larrabee as they can't be ready by the time AMD will.

 

SCC cores are meant for integer workloads and not fp workloads. The two have vast differences in their nature, and the SCC cores certainly don't have the massive vector unit that is essential to lrb.

 

So far we don't know much about them (width? in order? out of order? even if I would put my bet on a narrow in order core, I'm not sure they passed on SIMD units but I don't expect 8 wide units either) but it's not really my point my point is not about graphics prowess, I agree with you SCC is not to compete with high end NVDA or ATI products as graphic part (it's not what I meant) but as on chip accelerators they may prove more useful /easier to deal with than any fusion products. A sandy brigdge / SCC "fusion" chip may shine as a desktop part which user would care if his desktop/profesionnal computer is back in the software rendering realm if all the really relevant application are faster?

 

that paragraph about Intel's IGP benefitting from a canned Larrabee, it's hard to make sense of it. first I read of it as implyingg Larrabee tech would be used in IGP, but it's not about that at all.
It's fishy : Larrabee would probably be in the $500 range (or $300 at the least) and use gobs of power, so I don't quite see the relationships with IGP that you'll find on desktop computers costing as much as Larrabee itself.

It's not like granma would buy a cheap Dell or Acer then upgrade with a 300W Larrabee card down the road.

Intel can beef up their current IGP tech, too. double units, clock faster etc. ; I'm curious to see how sandy bridge's IGP makes progress (it will have a good available bandwith).
I don't believe Larrabee is desirable for an IGP : you absolutely want perf/area/watt, which is not achieved by ditching ROPs, texture filtering etc.

SCC cores should not be useful. The cores themselves are nothing special.
It's a research product, a headstart for Intel in manycores development, an advantage they buy on competition. They have so much resources that they can build right now such chips, to give insight on architectures that you'll see in the second half of the coming decade. They learn a lot on the hardware side, bandwith (the one problem), plus it can be a testbed for software.
Larrabee 1 itself turns out similar. They'll make compilers, dev tools, libraries etc. as well as seeding it out to developers.

it's a long term approach, only meaningful with lots of cores so you won't see Larrabee/SCC a-likes in the low end for a while.

 

Depending on whether the GPU clocks up linearly with the CPU (IE - overclock one you overclock the other), will probably go a ways to determining whether it might or might not overclock well.

AMD over the past couple decades has been a bit hit and miss on overclockability of parts...

Regards,
SB

 

I doubt we'll ever see processors as wide as GPUs clock nearly as high as CPUs.

 

The way I understood that, it means that due to LRB's failure, Intel will have to step up it's IGP efforts and get some really good things in there, since fusion will eat any Intel IGP for breakfast.

If you look at it, Intel now has 3 lines of graphics parts,

• Larrabee, delayed to 2011
• their own GMA series, crappiest shipping GPU on the planet
• IP licensed from IMG, good, but not shipping in desktops and laptops yet.

IMG IP looks like their best bet to get competitive IGP's out.

 

IMG tech has shipped on netbooks and have worthless drivers, at least on the linux side.
apparently Intel contracts out their GMA500 drivers ; they don't take it seriously or something, again Intel doesn't want the Atom platform to be too competitive.

given that factoid I would hope for an Intel IGP rather than a PowerVR one in the Sandy Bridge.

 

Intel's powervr drivers are known to be "a bloody mess", but improvements are supposed to be coming in Mobiln 2.2.

 

Only a fraction of the buyers know and can understand the difference, and infact amd igp are actually eating intel's, but looks like nobody cares

 

"Nobody cares" if you look at the sales numbers but IGPs don't sell platforms, overall system performance, price, and sadly brand name recognition do. Obviously Intel is winning here on name alone. I still run into people that don't know who or what AMD is or perceive them to be inferior to Intel simply because they're not Intel.

 

maybe nvidia is better at the marketing game then, but they do sell an IGP as a platform : ION.
though I don't quite understand why they aren't competing with AMD IGP anymore, there's not even geforce 8200/8300 AM3 motherboards, let alone a newer offering.

 

My statement "IGPs don't sell platforms" means customers don't shop for systems based upon what IGP they use, generally speaking. Sure there are a few informed buyers out there looking for a system on a budget that may choose an NV or AMD IGP over Intel, but the vast majority of the market does no such thing.

 

I'd count that as four, since intel's own gma series still features both gen3 (i915,i i945, pinetrail is supposedly still this generation) and gen4 (i965, g45, ironlake) parts, which have absolutely nothing in common (as far as 3d is concerned, they have similar display controllers).
I don't know what intel exactly has licensed from IMG, but if it's only sgx535 that part certainly isn't going to be competitive in notebooks, and I very highly doubt sgx545 (highest performing IMG part you could "buy" right now) would be.