[Computex 2010]AMD to show Fusion tomorrow

[haylui]

Yea but anything Intel puts out (no matter what it is) would have been lame for you.

really?
since Intel demo the capability of its SB's GPU I'm expecting it is comparable rather than not comparable.

[freeloader]

I think that is true. The need for really powerful GPUs will remain in the future, the 100$ market and up is here to stay, i cannot imagine how someone could implement for example a GPU the size of 5870 with a CPU onto the same die and keep it under 100W, it's just not possible.

Not possible as of today, but who knows what node we will be on (sub 10nm?) in ten years? I'm pretty sure you'd get quadcores and a full 5870 on that node under 100 watts.

[Manicdan]

Not possible as of today, but who knows what node we will be on (sub 10nm?) in ten years? I'm pretty sure you'd get quadcores and a full 5870 on that node under 100 watts.

hes talking about a mainstream or semi enthusisit card, not just that we will never have something more powerful than a 5870.

the fusion market may be able to offer the same performance as a 200$ cpu + 200$ gpu, but for people who want a 500$ gpu, those will still have to exist, and its very doubtful you will be able to buy a 300mm2 cpu combined with a 400mm2 gpu in a fusion like package.

[Florinmocanu]

Not possible as of today, but who knows what node we will be on (sub 10nm?) in ten years? I'm pretty sure you'd get quadcores and a full 5870 on that node under 100 watts.

And by that time, a GPU of 200W on a 10nm process will probably be way more powerfull than the fusion GPU on 10nm.

Fusion is low power, i doubt we will see a fusion product over 100W, that's why we have normal CPUs with lots of cores for. Fusion i think will stay around 60-70W for top-end part.

A GPU of 200-250W is immensely more powerful in games and will always be.

PS: Manicdan explained it pretty decent, 130W CPUs and 100W+ GPUs will always have a market, APUs are designed for low power.

[Manicdan]

PS: Manicdan explained it pretty decent, 130W CPUs and 100W+ GPUs will always have a market, APUs are designed for low power.

just like 6 years ago you couldnt get a card that was much more than 100W, now you can have over a killowatt of gpu power. looking at the consoles ps1 was 150$, ps2 was 300, and ps3 was 600 new. as time changes people seem to want just more and more and more.

id be scared that 2P will become consumer level in the next decade, lol

[Mechromancer]

Those who are concerned about how a Fusion APU will run games aren't seeing the big picture. APUs are for OpenCL acceleration and other GPGPU applications. In fact, AMD should have done a much better job of emphasizing this fact than showing AVP at low frame rates. 480 shaders on an APU is decent for games, but baller for a GPGPU app. Hell, when the OpenCL GPU3 Folding@Home client shows up, a measly Llano will crank out decent PPD on a laptop.

[informal]

Those who are concerned about how a Fusion APU will run games aren't seeing the big picture. APUs are for OpenCL acceleration and other GPGPU applications. In fact, AMD should have done a much better job of emphasizing this fact than showing AVP at low frame rates. 480 shaders on an APU is decent for games, but baller for a GPGPU app. Hell, when the OpenCL GPU3 Folding@Home client shows up, a measly Llano will crank out decent PPD on a laptop.

I agree there is a nice possibility for GPGPU,but AMD had to show they had a working DX11 Fusion product.And btw,that was Ontario/Bobcat that run the AVP ,with its low power/low clock x86 cores and lowly 80 Stream Processors

[Mechromancer]

^^Only 80 stream processors?!?!? Well that isn't as bad as I'd expect performance to be in that game.

[Sn0wm@n]

hes talking about a mainstream or semi enthusisit card, not just that we will never have something more powerful than a 5870.

the fusion market may be able to offer the same performance as a 200$ cpu + 200$ gpu, but for people who want a 500$ gpu, those will still have to exist, and its very doubtful you will be able to buy a 300mm2 cpu combined with a 400mm2 gpu in a fusion like package.

Llano wasnt rumored to have a 400 shader gpu ??? so that puts it at 5670 performance depending on clock ...... shrink to the next lowest full node ... and bam 5870 performance is reacheable ... either by increasing shader count or raising clock speed while staying at the same thermal of the 32nm Llano

[kl0012]

Llano wasnt rumored to have a 400 shader gpu ??? so that puts it at 5670 performance depending on clock ...... shrink to the next lowest full node ... and bam 5870 performance is reacheable ... either by increasing shader count or raising clock speed while staying at the same thermal of the 32nm Llano

I wonder what performance will have 5870 with 128 bit bus and DDR3 mem.

[wuttz]

id be scared that 2P will become consumer level in the next decade, lol

1P is the new 2P,
2P is the new 4P.

G34 is turning the market upside-down.
8 real cores for $257.00, 16 for $514.

if you can find it.

[Manicdan]

Llano wasnt rumored to have a 400 shader gpu ??? so that puts it at 5670 performance depending on clock ...... shrink to the next lowest full node ... and bam 5870 performance is reacheable ... either by increasing shader count or raising clock speed while staying at the same thermal of the 32nm Llano

you might be a little off on the expectations. if we can get 480SPs on Llano at 32nm, then at the next step, 22nm, (or is it 24? idk...) we can expect 50% more, with maybe 10% more clock rate at the same power. so its 800SPs if were lucky, and clocking probably 1ghz, sounds like an OCed 5770, or maybe up to a 5830. the next one will probably be well within reach of a 5870 though. so that means <1 year till this, and then 2 years for each node after, so just under 5 years away and u can have a 5870 with probably an 8 core BD chip under 100W.

kinda scary what laptops and nettops will be able to do in 5 years.

[Manicdan]

1P is the new 2P,
2P is the new 4P.

G34 is turning the market upside-down.
8 real cores for $257.00, 16 for $514.

if you can find it.

lol, its limit 1 per customer for a 2P chip

[Sn0wm@n]

you might be a little off on the expectations. if we can get 480SPs on Llano at 32nm, then at the next step, 22nm, (or is it 24? idk...) we can expect 50% more, with maybe 10% more clock rate at the same power. so its 800SPs if were lucky, and clocking probably 1ghz, sounds like an OCed 5770, or maybe up to a 5830. the next one will probably be well within reach of a 5870 though. so that means <1 year till this, and then 2 years for each node after, so just under 5 years away and u can have a 5870 with probably an 8 core BD chip under 100W.

kinda scary what laptops and nettops will be able to do in 5 years.

the 22nm will reduce the die size by close to 40% - 50% cut power down to 30% .... so using that die space to add more shader and use the norther island shader design wouldnt be a bad idea on the 22nm node gaining another 20% more performance ... and bam 5870 performance right there .... on a decent tdp

[Manicdan]

50% more shaders, using 30% less power, with 20% ipc gains over Llano will = 5870? very doubtful.

if Llano is 1x, then 1 * 1.5 for space, and * 1.2 for IPC = 1.8x. so a 22nm version of Llano is 1.8x better than 32nm version. so by that standard, the current Llano has to be better than a 5770. and even if it is, with only a 30% power drop and being 50% bigger, your going to have a much higher TDP at the same clocks, effectively reducing how high you can clock it, say 800mhz vs 700mhz with a .1v drop.

[Aerwidh]

lol, its limit 1 per customer for a 2P chip

Now there's the laugh of the day

[wuttz]

lol, its limit 1 per customer for a 2P chip

news flash! it isn't 2P only, that chip goes across the board from 1P to 4P.
one reason it's 1/customer only is probably the tight demand.

the other reason is this. did i mention G34 is blurring server market segements?

[Chumbucket843]

the 22nm will reduce the die size by close to 40% - 50% cut power down to 30% .... so using that die space to add more shader and use the norther island shader design wouldnt be a bad idea on the 22nm node gaining another 20% more performance ... and bam 5870 performance right there .... on a decent tdp

different types of logic scale differently and there is a huge difference in density. most of the gpu is SRAM and datapaths which will scale very well. think 2x more logic from 32nm to 22nm for the gpu. hopefully Glofo will quit using SOI past 32nm and density will increase even more.

[snoro]

news flash! it isn't 2P only, that chip goes across the board from 1P to 4P.
one reason it's 1/customer only is probably the tight demand.

the other reason is this. did i mention G34 is blurring server market segements?

wrong, it can go all the way up to 8p in one rig with cpu extension board.

But the cost for only a 4 socket motherboard and 2x 2socket cpu extension board would cost more than 8 opteron 6128 ( the 300$ 8 core 2.0ghz cpu). But hey, you would have a 64 core rig !!!!
And the n you can upgrade to lowest bulldozer server cpu which will be 12 core and have 96 core or go batcrazy and get the 16 core one and have 128 core.

[FischOderAal]

hopefully Glofo will quit using SOI past 32nm and density will increase even more.

So SOI has worse transistor density?

[JF-AMD]

wrong, it can go all the way up to 8p in one rig with cpu extension board.

But the cost for only a 4 socket motherboard and 2x 2socket cpu extension board would cost more than 8 opteron 6128 ( the 300$ 8 core 2.0ghz cpu). But hey, you would have a 64 core rig !!!!
And the n you can upgrade to lowest bulldozer server cpu which will be 12 core and have 96 core or go batcrazy and get the 16 core one and have 128 core.

What is this "CPU extension board" that you speak of?

[snoro]

my bad, i thought one of the 2 cpu supermicro motherboard pcie express 16x slot was a proprietary link to act as a extension bridge.

SM 2 socket g34 motherboard with one very strange pcie 16x slot placement

[Chumbucket843]

So SOI has worse transistor density?

yup. their are tradeoffs for SOI. you'll get higher clocks but there are numerous disadvantages. for instance the insulator is also an insulator of thermal energy which means it will get hot unlike a normal bulk process where the heat can diffuse into the substrate. it's also more complex and costs more. i think amd used it b/c they collaborated with ibm on their 45nm node and probably some other tech.

[superrugal]

L2-cache is shared ?

http://pc.watch.impress.co.jp/docs/c...04_371991.html

[Hans de Vries]

L2-cache is shared ?
http://pc.watch.impress.co.jp/docs/c...04_371991.html

Rather unlikely in my opinion....

From the same page, AMD's Ontario wafer:


AMD's dual Bobcat core + DX11 GPU Ontario is remarkably small: ~75mm^2
(>900 dies per wafer) which is smaller as the dual core atom based D510
Pineview which has 87mm^2

Intel dual core Atom D510: http://ark.intel.com/Product.aspx?id=43098



Regards, Hans