Llano IGP vs SNB IGP vs IVB IGP [Archive] - Page 2

Alexko

13-Jun-2011, 21:50

3450 can run some games ok. It will run UT3 alright at reduced settings. But then you try KOTOR and it's not entirely smooth. It really reminded me of a Radeon 9600 with support for more shader features.

The inability to accelerate flash is really a problem though when you are on a system with a wimpy CPU. It can accelerate H.264 and VC-1 but its UVD lacks something to allow flash accel.

To be fair, KOTOR's engine was pretty buggy and would go from running fine to lagging badly for no apparent reason. You had to close the game and relaunch it. Of course, on current high-end graphics cards, even when it switches to "super-slow mode" you don't notice it because it's still above 60 FPS, but on a 3450 I can imagine how it might be noticeable.

ToTTenTranz

13-Jun-2011, 21:57

3450 can run some games ok. It will run UT3 alright at reduced settings. But then you try KOTOR and it's not entirely smooth. It really reminded me of a Radeon 9600 with support for more shader features.

The inability to accelerate flash is really a problem though when you are on a system with a wimpy CPU. It can accelerate H.264 and VC-1 but its UVD lacks something to allow flash accel.

Wut? The HD3200 IGP (RV610 as well) in my Ferrari One accelerates Flash video and all kinds of browser acceleration just fine.
It's probably the one thing that's holding that dual-core 1.2GHz Athlon Neo together for a very decent web browsing experience.

I've also tried KOTOR in it and it played fine with 1024*768 medium-ish settings @ ~30fps.
I even played Mass Effect 1 in a 960*540 custom resolution with all settings low, at some ~24fps.

Perhaps you used an old driver?

mczak

13-Jun-2011, 22:06

3D Mark numbers are very good for integrated solution for laptop, and that is with low clocked 4 core APU where we all know 3DMark 06 loves fast cores.

I doubt it would make a lot of difference for the gpu class we're talking here.

And as AMD doesn't have anything to compete with nVidia's Optimus for Intel yet, I'd say a Llano with the discrete GPU turned off should be more power efficient than an i5+HD66xx.

There are some hybrid notebooks on the market with AMD graphics without the display mux which have the display outputs just on the IGP. Looks to me AMD introduced its own Optimus solution rather quietly without anyone noticing yet :-).

ToTTenTranz

13-Jun-2011, 22:12

There are some hybrid notebooks on the market with AMD graphics without the display mux which have the display outputs just on the IGP. Looks to me AMD introduced its own Optimus solution rather quietly without anyone noticing yet :-).

For Intel CPUs and IGPs?

Really? Could you point out a model or two? (Honest question, I really thought AMD's switchable graphics (http://www.amd.com/uk/products/technologies/switchable-graphics/Pages/dynamic-switchable-graphics.aspx) only supported AMD IGPs so far.)

swaaye

13-Jun-2011, 23:02

Wut? The HD3200 IGP (RV610 as well) in my Ferrari One accelerates Flash video and all kinds of browser acceleration just fine.
It's probably the one thing that's holding that dual-core 1.2GHz Athlon Neo together for a very decent web browsing experience.
780G does accelerate flash but most if not all of the discrete 3xxx GPUs do not for some reason.

I've also tried KOTOR in it and it played fine with 1024*768 medium-ish settings @ ~30fps.

Yes it's similar to a Radeon 9600 Pro (or X600 or X1300). Not exciting but is expected for a part with a tiny transistor budget and yet full DX10.1 featureset. I've messed with 780G too but on the desktop. I got the IGP up to 930 MHz. ;) It's too bandwidth constrained to really benefit though.

The modern stuff like 6450 is more interesting because it has far more ALU power and often GDDR5. They've also improved the texturing performance since the olden R600 times.

32nm Llano could make for some interesting notebooks....

mczak

13-Jun-2011, 23:12

For Intel CPUs and IGPs?

Really? Could you point out a model or two? (Honest question, I really thought AMD's switchable graphics (http://www.amd.com/uk/products/technologies/switchable-graphics/Pages/dynamic-switchable-graphics.aspx) only supported AMD IGPs so far.)
Dell Vostro 3550 for example.
http://www.notebookcheck.net/AMD-PowerXPress-4-0-aka-BACON-vs-Nvidia-Optimus.55204.0.html

edit: there are others, at least other Vostro's, I thought I saw some non-dell notebook as well. Since AMD doesn't really have a name for it yet though it's sort of hard to search notebooks specifically which support it :-).
Judging by a bug filed against the linux kernel (can't switch gpu...), there's also at least HP notebooks out which support this though I don't know the exact model (recent dv7).

Dave Baumann

13-Jun-2011, 23:50

Really? Could you point out a model or two? (Honest question, I really thought AMD's switchable graphics (http://www.amd.com/uk/products/technologies/switchable-graphics/Pages/dynamic-switchable-graphics.aspx) only supported AMD IGPs so far.)
Switchable graphics is actually for non-AMD solutions.

Psycho

14-Jun-2011, 00:46

There's also the MBP: http://www.apple.com/macbookpro/features.html
The integrated Intel HD Graphics 3000 processor — now included across the MacBook Pro lineup — handles the things you do every day. It encodes video quickly, making HD video calls with FaceTime possible. And it decodes efficiently, so you get long playback time for DVDs and iTunes movies. For graphics-intensive applications, the 15- and 17-inch MacBook Pro models automatically switch to new high-performance AMD Radeon graphics processors.

liem107

14-Jun-2011, 05:23

AnarchX

14-Jun-2011, 07:38

Nearly 1 billion transistors for a 32nm SOI Redwood (627M @ 40nm bulk)? :shock:

This chip weighs in at 1.45 billion transistors, nearly 50% more than Sandy Bridge. Around half of the chip is dedicated to the GPU however, so those are tightly packed transistors resulting in a die size that's only 5% larger than Sandy Bridge.
http://www.anandtech.com/show/4444/amd-llano-notebook-review-a-series-fusion-apu-a8-3500m/2

btw.
http://img811.imageshack.us/img811/4153/trinityze.jpg (http://img811.imageshack.us/i/trinityze.jpg/)
http://www.anandtech.com/show/4444/amd-llano-notebook-review-a-series-fusion-apu-a8-3500m/13
~900 GFLOPs for Trinity.

Blazkowicz

14-Jun-2011, 08:37

I hate being pedantic, but the constant misuse of this word is annoying.

niche - a distinct segment of a market.

Of course x86 is a niche of the market... niche != small.

I disagree. x86 on phone/PDA-like devices such as UMPC is a niche, but you can't say that for mainstream computers, servers and the like.

if a niche is a subset of a market, then if 100% of all computing was all x86, it would be a niche as the full set is a subset of itself :razz:

more seriously, x86 is present in many markets including tablets (anecdotal), industrial, embedded, and dominates the desktop, laptops and servers.
a niche instruction set or technology would have significant presence in one market (or dominate it) and be mostly irrelevant in others.

it was even put in consoles (twice) and in a nokia smartphone back when the word smartphone didn't exist.

Florin

14-Jun-2011, 09:35

Silent_Buddha

14-Jun-2011, 11:00

I haven't seen it mentionned but anandtech previews are up.
as expected K10.X gen cpu performances don't exactly shine vs corei3....yep AMD really n eeds to push strong on APU programming to try bridge the gap with SDB cpu perfs.
but GPU perf is quite enough for low end gaming.
http://www.anandtech.com/show/4448/amd-llano-desktop-performance-preview
http://www.anandtech.com/show/4444/amd-llano-notebook-review-a-series-fusion-apu-a8-3500m

And right about where I expected it to land. Around entry level discrete graphics performance (6450) versus the somewhat exaggerated expectations of some that were hoping for midrange performance (x5xx/x6xx) from it.

Still a pretty good first step and with an, IMO, better balance between GPU/CPU than Sandy Bridge. Sandy Bridge still has far better pure CPU power. But Llano strikes a better balance if a user does occasional gaming.

Wish they had been able to indicate power consumption however.

Regards,
SB

rpg.314

14-Jun-2011, 11:19

Nearly 1 billion transistors for a 32nm SOI Redwood (627M @ 40nm bulk)? :shock:

SOI is generally a bit denser than bulk at the same node.

Custom logic has more density than synthesized logic.

Caches also increase density. There's not much of it in GPUs.

~900 GFLOPs for Trinity.
I hope we'll get to see this sometime in 2012.

Zaphod

14-Jun-2011, 11:28

Wish they had been able to indicate power consumption however.

One can do some back of napkin calculations based on the notebook preview. I.e. looping 3DMark06, the system used sin the neighborhood of 21W.

ToTTenTranz

14-Jun-2011, 12:10

http://www.anandtech.com/show/4444/amd-llano-notebook-review-a-series-fusion-apu-a8-3500m/10

It's such a shame that AMD screwed up with the drivers for Hybrid Crossfire at launch!

Just look at those 3DMark Vantage & 11 results, and they're not even using the best APU nor the best dGPU available, far from it. I bet the "HD6775G2" combination could turn out faster than some of those GTX460M results, even with the CPU advantage.

There's a huge potential in there, but the games aren´t using it at all... It seems the driver team just rushed the Crossfire profiles for 3DMark and that's it..

And what's up with the "DX10\11 only" crossfire support?! It's such a shot in the foot... so there won't be any Crossfire for Skyrim, Bioshock 2, Unreal 3, and previous games?
Dave, please tell us DX9 will be supported for "G2" combinations!

And I wonder why send laptopts with DDR3-1333MHz, if the platform supports DDR3-1600? Llano should be pretty much bandwidth-hungry, right?

Alexko

14-Jun-2011, 12:57

And right about where I expected it to land. Around entry level discrete graphics performance (6450) versus the somewhat exaggerated expectations of some that were hoping for midrange performance (x5xx/x6xx) from it.

Still a pretty good first step and with an, IMO, better balance between GPU/CPU than Sandy Bridge. Sandy Bridge still has far better pure CPU power. But Llano strikes a better balance if a user does occasional gaming.

Wish they had been able to indicate power consumption however.

Regards,
SB

Anand tested his APU with DDR3-1333, while Llano officially supports DDR3-1600 and 1866.

That's almost 40% more bandwidth left on the table.

Lightman

14-Jun-2011, 14:01

Anand tested his APU with DDR3-1333, while Llano officially supports DDR3-1600 and 1866.

That's almost 40% more bandwidth left on the table.

Not mentioning this in article was unfair.
Also comparing hd6450 + 2500k to APU bias results even more thanks to much faster CPU.
If they would put Athlon Ii X4 + 6450 we could draw better conclusions.

At least Anand is going to do prpoer review in coming weeks.

ToTTenTranz

14-Jun-2011, 14:18

Anand tested his APU with DDR3-1333, while Llano officially supports DDR3-1600 and 1866.

That's almost 40% more bandwidth left on the table.

In a system where memory bandwidth definitely makes some difference, and with DDR3-1600 being pretty much the "minimum standard" nowadays (equally priced to DDR3-1333, at least), it's just odd that he's decided to couple the system with slow memory.

Albuquerque

14-Jun-2011, 15:07

In a system where memory bandwidth definitely makes some difference, and with DDR3-1600 being pretty much the "minimum standard" nowadays (equally priced to DDR3-1333, at least), it's just odd that he's decided to couple the system with slow memory.

Are we sure that he chose that configuration, or is it more likely that this configuration was provided to him somehow? I don't know, just asking...

ninelven

14-Jun-2011, 15:36

I disagree. It does not matter if you disagree. That is how it is defined. Using the word niche to imply size or significance is using it incorrectly, period.

Lightman

14-Jun-2011, 16:29

Are we sure that he chose that configuration, or is it more likely that this configuration was provided to him somehow? I don't know, just asking...

In this case we are talking about Desktop Llano and Anand has several boards from manufactures, so this implies he is deciding which components to build it around.
In case of SB memory support is limited to 1333 (officially), but Llano Desktop officially supports DDR3 1833.
It's like testing SB platform with DDR3 800, does it make sense from performance point of view? No! Can it be done when comparing performance impact of faster/slower memory on different tasks? Yes!

Alexko

14-Jun-2011, 16:51

It does not matter if you disagree. That is how it is defined. Using the word niche to imply size or significance is using it incorrectly, period.

This is a little bit off-topic but here is Oxford's definition:

a specialized but profitable segment of the market:
[as modifier] :
a niche market for quality food

http://oxforddictionaries.com/definition/niche

To me, specialized => small.

swaaye

14-Jun-2011, 17:45

It looks like Llano performs similar to a M5650 in games and that's nice. Maybe that 35W variant will end up in some interesting little notebooks.

But wow the CPU tech is looking really poor up against recent Intel hardware. There are a few tests at Anandtech that show Intel more than doubling CPU performance. It's interesting that they dumped the Phenom shared L3 design and stuck with the Athlon II design. Tom's shows a Phenom II really beating up Llano (they think it might have to do with TDP limitation).

Regarding DDR3 speed, 1066 and 1333 are common in notebooks and that's where this chip is most interesting I think.

Gipsel

14-Jun-2011, 18:11

Caches also increase density. There's not much of it in GPUs.But as the register files are done with SRAM, there is more than roughly nothing of it in Llano (256 kB register files + 32kB local memory per SIMD). That alone amounts to ~1.4 MB SRAM arrays. Add the texture L1-caches (40kB), the (probably) 256 kB texture- L2, 64 kB GDS, the ROP caches (color and Z) and a few buffers here and there and I would be surprised if the Llano GPU has a lot less than 2 MB SRAM (roughly 100 Million transistors?). It's less than on the CPU side though (4.5 MB L1 + L2 cache).

I would guess a part of the high density of the GPU part also results from the somewhat relaxed latency and frequency requirements of the GPU. 600 MHz of the GPU part in the desktop versions in 32nm SOI does not appear very challenging when one considers a HD6870 runs at 900 MHz in 40nm bulk from TSMC (and has even a higher transistor density as the whole Llano).

edit: Llano die shot (each of the groups with 4 VLIW-units are accompanied by 32 small blocks of SRAM [each 2kB in size]):

http://pics.computerbase.de/3/4/9/8/5/11.jpg

What looks a bit confusing to me is that the upper part of the SIMDs are slightly different from the lower half.

Blazkowicz

14-Jun-2011, 18:24

anands announces a smaller llano chip with dual core CPU and 240SP GPU.
that one will be pretty good on light laptops, or ones with a large battery life and no overheating.
though no earth shattering in any way - it feels like a core2duo with a geforce 8600M :p

a good laptop imo would have a 15.6" or 16" IPS, high res screen, a 2C llano, 4 or 8GB memory, a 60GB or something SSD, a 1TB hard drive and no optical drive.

Psycho

14-Jun-2011, 18:36

The review at Toms (http://www.tomshardware.com/reviews/a8-3500m-llano-apu,2959.html) is (for once) pretty good with comparisons with Phenom2 X4 at 1.5 and 2.5 ghz, and a 5570 for the graphics. Only missing an X2 to see how threaded the various CPU benchmarks really are.
But feels like Turbo Core is rarely kicking in, which is strange considering that the GPU should be mostly idle in those tests, so even with 4 cores it should be quite a bit more than 1.5 ghz.

swaaye

14-Jun-2011, 19:10

ToTTenTranz

14-Jun-2011, 19:15

Tom decided to only use the DX9 versions of DX10/11 games for the antialiasing benchmarks, so the Crossfire wouldn't work.

wtf?

swaaye

14-Jun-2011, 19:32

I am not really interested in problematic Crossfire/SLI, especially when weak GPUs are involved. There are plenty of such results at Anandtech though. This CPU is too weak for a really fast GPU setup anyway.

ToTTenTranz

14-Jun-2011, 19:37

I personally couldn't care less about crossfiring gimpy GPUs (or crossfire in general for that matter). There are plenty of such results at Anandtech though.

"Gimpy"?!
We're talking about gaming performance on a 700€ 14" 2Kg laptop that goes awfuly close to 1600€ 17" 4.5Kg laptops!

Thank god you're not even marginally representative of budget-conscious gamers..

mczak

14-Jun-2011, 19:38

Yup the Tom's review shows the Phenom II laying some smackdown. It doesn't always keep up with the 1.5 GHz Phenom II. The lack of shared L3 is probably hurting it in highly threaded apps ala Athlon II. Turbo looks very weak. Gimpy CPU with a decent GPU. ATI saves the day.
Doesn't look like the lack of L3 makes that much difference to me (except abby finereader), just turbo seems to fail to clock up even a tiny bit with multithreaded loads and never reaching anything close to the max even with single-threaded loads.
I think skipping L3 cache makes sense for now. Just look at the AII and PII parts on the desktop, the L3 cache usually doesn't do all that much except for games when they are cpu-bound with high-end graphic cards - no issue for notebooks, and not much of an issue for lower-end desktops neither.
Even with L3 cache (I guess you'd really need 4MB at least, allowing reduction of L2 to 512KB for it to make any positive impact) the cpu just would be nowhere close to Sandy Bridge anyway, and the die is already huge compared to dual-core Sandy Bridge parts.
L3 cache makes a LOT of sense if you make that available to the IGP too, but I think AMD tried to stick to "old" graphics (and cpu) core without serious redesign (after all, it's their first 32nm chip). Trinity should probably show what this can do, at least I'd hope so, given intel already does that (and Ivy Bridge could potentially come close to Llano in IGP performance).

swaaye

14-Jun-2011, 19:38

Thank god you're not even marginally representative of budget-conscious gamers..

Lol yeah I guess it's a good thing I'm not. This CPU is a dog compared to a lot of the budget Intel notebooks with discrete graphics. Strategy gamers beware probably.

swaaye

14-Jun-2011, 19:45

Doesn't look like the lack of L3 makes that much difference to me (except abby finereader), just turbo seems to fail to clock up even a tiny bit with multithreaded loads and never reaching anything close to the max even with single-threaded loads.

Yea you are right that L3 would have cost a lot for a little gain. Turbo seems to either be very limited or bugged.

mczak

14-Jun-2011, 19:49

Regarding DDR3 speed, 1066 and 1333 are common in notebooks and that's where this chip is most interesting I think.
Still poor excuse for using ddr3-1333 even on the desktop. Maybe though we'll see ddr3-1600 a bit more in notebooks now? Let's face it up to now there just wasn't much incentive to use faster memory in notebooks. Well intel didn't support it at all, not sure about AMD but it would have been totally pointless anyway. The (presumably pretty small) power draw disadvantage should be well worth the faster graphics you get from ddr3-1600 I think. Though I'm wondering why it only supports low-voltage ddr3-1333 and not the low-voltage ddr3-1600 version - but few notebooks use low-voltage dimms today.

Blazkowicz

14-Jun-2011, 19:53

L3 cache makes a LOT of sense if you make that available to the IGP too, but I think AMD tried to stick to "old" graphics (and cpu) core without serious redesign (after all, it's their first 32nm chip). Trinity should probably show what this can do, at least I'd hope so, given intel already does that (and Ivy Bridge could potentially come close to Llano in IGP performance).

and that classic shared memory architecture worked better.
sandy bridge has this little problem where the performance tanks if all CPU threads are fully loaded while amd keeps the GPU usable.
once again Intel is the high rev and high tech small gasoline engine, and AMD is the clunky turbo diesel that falls short on power, but is steady when hauling big loads and is quite advanced on its own too. :twisted:

mczak

14-Jun-2011, 20:09

and that classic shared memory architecture worked better.
sandy bridge has this little problem where the performance tanks if all CPU threads are fully loaded while amd keeps the GPU usable.
once again Intel is the high rev and high tech small gasoline engine, and AMD is the clunky turbo diesel that falls short on power, but is steady when hauling big loads and is quite advanced on its own too. :twisted:
I've seen these benches but there was no proof it was due to the L3 sharing. Even if it was though there are ways to solve this (could reserve portions of the L3 cache to either cpus or gpu for instance), in any case I expect sharing L3 between cpu and gpu is going to be inevitable for next gen, the advantages are just too big.

btw anand has updated his desktop benches with ddr3-1866:
http://www.anandtech.com/show/4448/amd-llano-desktop-performance-preview/3
That gets things up to roughly gt430 performance...
So for 40% more memory bandwidth, performance improved by 20-30% (though mostly close to 20%). Definitely better than HD6450 gddr5 now (which is unavailable anyway but that's a different topic).

swaaye

14-Jun-2011, 20:36

Yeah faster RAM gives it a nice boost. Still, being below a 5570 is somewhat sobering in the grand scheme. I want to see the 35W Llano put to good use in a smaller form factor where a discrete setup is less than ideal.

ToTTenTranz

14-Jun-2011, 20:47

Lol yeah I guess it's a good thing I'm not. This CPU is a dog compared to a lot of the budget Intel notebooks with discrete graphics. Strategy gamers beware probably.
If we're talking gaming performance, I think the 4-core Llano is actually quite balanced to pair with a graphics system with a processing power equivalent to a downclocked desktop HD5750.
Just look at the 3dmark11 results in the anandtech review. The measly 1.5GHz quad-core + HD6690G2 (fGPU + HD6630M) is bitting the toes of Intel's mightiest laptop Sandybridge + dedicated 675MHz GF106. The latter are top-end big-sized gaming laptops with a ~2.5 hour battery life, not some mid-sized budget notebooks with 6+ hours.

btw anand has updated his desktop benches with ddr3-1866:
http://www.anandtech.com/show/4448/amd-llano-desktop-performance-preview/3
That gets things up to roughly gt430 performance...
So for 40% more memory bandwidth, performance improved by 20-30% (though mostly close to 20%). Definitely better than HD6450 gddr5 now (which is unavailable anyway but that's a different topic).

Wow, that's by far the largest jump in overall-performance per memory bandwidth I've ever seen!

I guess the memory controller tweaks make no miracles and the Fusion APUs are, in fact, starving in bandwidth.
The jump to DDR3 1333MHz for Brazos UMPCs should give them a large performance bump too. I've seen benchmarks where the CPU gets much higher scores when the fGPU is turned off.

swaaye

14-Jun-2011, 20:53

If we're talking gaming performance, I think the 4-core Llano is actually quite balanced to pair with a graphics system with a processing power equivalent to a downclocked desktop HD5750.
Just look at the 3dmark11 results in the anandtech review. The measly 1.5GHz quad-core + HD6690G2 (fGPU + HD6630M) is bitting the toes of Intel's mightiest laptop Sandybridge + dedicated 675MHz GF106. The latter are top-end big-sized gaming laptops with a ~2.5 hour battery life, not some mid-sized budget notebooks with 6+ hours.
Of course you have to remember that multi GPU setups come with issues in some games and require profiles to perform well. Sometimes it simply doesn't benefit a particular game at all. That's why I tend to ignore Crossfire and SLI.

mczak

14-Jun-2011, 21:21

Just look at the 3dmark11 results in the anandtech review. The measly 1.5GHz quad-core + HD6690G2 (fGPU + HD6630M) is bitting the toes of Intel's mightiest laptop Sandybridge + dedicated 675MHz GF106. The latter are top-end big-sized gaming laptops with a ~2.5 hour battery life, not some mid-sized budget notebooks with 6+ hours.

I tend to think that's about the best scaling you'll ever see though (and it's still 20% off that GTX460M btw).
I think something like the 3dmark Vantage result is probably more representative what you'll get with asymmetric CF. Which means it'll beat GT540M but it's probably not even enough to beat the top-end mobile Turks alone (HD6770M - still the same graphic chip...). Of course you could use CF on that too but I doubt it's worth it. Now if you could run OpenCL physics on the IGP and graphics on the GPU that would be more interesting probably...
And really asymmetric CF is useless right now. It didn't really work in anything but 3dmark for now. I'm sure that'll get fixed but I'm not even sure AMD wants to extend support to d3d9?
I think for that level of performance you'd be much better off with something like a discrete HD6770M (switchable of course for battery runtime), not worrying about the headaches of asymmetric CF.

Wow, that's by far the largest jump in overall-performance per memory bandwidth I've ever seen!

Not even close there are plenty of HD5570 out there with, get this, DDR2 memory!
I bet those show a larger performance increase (in percent) per memory bandwidth increase (with 10% faster GPU and 33% less memory bandwidth to start with)...

I guess the memory controller tweaks make no miracles and the Fusion APUs are, in fact, starving in bandwidth.

I don't really see any improvement in that area, the better prefetching should be specific to the cpu cores. For the gpu I think it won't make much of a difference, it'll just have slightly higher latency (likely as it's more complex) than redwood MC had.

The jump to DDR3 1333MHz for Brazos UMPCs should give them a large performance bump too. I've seen benchmarks where the CPU gets much higher scores when the fGPU is turned off.
Yes Brazos should be equally bandwidth limited. Strange though the cpu would get higher scores with gpu turned off, the GPU shouldn't need much bandwidth (when it's idle - of course running full tilt it might have a similar scheme to Llano giving gpu access higher priority).

Alexko

15-Jun-2011, 00:17

I've seen these benches but there was no proof it was due to the L3 sharing. Even if it was though there are ways to solve this (could reserve portions of the L3 cache to either cpus or gpu for instance), in any case I expect sharing L3 between cpu and gpu is going to be inevitable for next gen, the advantages are just too big.

btw anand has updated his desktop benches with ddr3-1866:
http://www.anandtech.com/show/4448/amd-llano-desktop-performance-preview/3
That gets things up to roughly gt430 performance...
So for 40% more memory bandwidth, performance improved by 20-30% (though mostly close to 20%). Definitely better than HD6450 gddr5 now (which is unavailable anyway but that's a different topic).

Now that's more like it! That's definitely enough for casual gaming, but it does raise the very interesting question (which we brushed a couple of days ago) of future scaling: how to deal with bandwidth starvation?

Lightman

15-Jun-2011, 00:28

Now that's more like it! That's definitely enough for casual gaming, but it does raise the very interesting question (which we brushed a couple of days ago) of future scaling: how to deal with bandwidth starvation?

I think Trinity will have 3 channels. Roadmaps are hinting exactly that, but they not 100% confirmed.

Extra channel assuming DDR 1600 would give 12.8GB/s more and brought total to 38.4GB/s. There is also that possibility of using L3 cache of CPU to ease memory requirements a bit. Of course we don't know yet if AMD is going to implement that for next gen APU.

Kaotik

15-Jun-2011, 00:52

So for 40% more memory bandwidth, performance improved by 20-30% (though mostly close to 20%). Definitely better than HD6450 gddr5 now (which is unavailable anyway but that's a different topic).

Crysis Warhead 1280: ~25%
Crysis Warhead 1024: ~20%
Metro 2033 1280: ~28%
Metro 2033 1024: ~29%
Dirt2 1280: ~19%
Dirt2 1024: ~22%
Mass Effect 2 1280: ~25%
Mass Effect 2 1024: ~15%

edit:
They should test with 1280x720 & 1366x768 though, rather than 1280x1024 / 1024x768, since Llano is obviously going to be used with widescreen displays, let it be TV or monitor, rather than old 5:4/4:3 monitors

ToTTenTranz

15-Jun-2011, 09:39

http://www.pcper.com/news/Graphics-Cards/AFDS11-Upcoming-Trinity-APU-will-use-VLIW4-Cayman-Architecture

Trinity will use a VLIW4 GPU.
As I suggested before, AMD is pushing harder to use the more computing-friendly VLIW4 architecture in future APUs, as the APP program is really important for the platform's success.

They've also shown Trinity in a working laptop.
Bulldozer cores + VLIW4 GPU in a laptop.. hmmm..

Would a half-Cayman (768 sps) be possible? Or something like 512sps, being almost as powerful as a Juniper Pro?

EDIT: PCPer's slides on the Fusion Developer Summit (http://www.pcper.com/news/Editorial/AMD-Fusion-Developer-Summit-2011-Live-Blog) mention a 50% increase in the APU's GFLOPs, so we're probably talking about a ~512sp VLIW4 GPU, maybe with higher base clocks for the GPU.

Kaotik

15-Jun-2011, 11:07

DarthShader

15-Jun-2011, 11:14

http://www.pcper.com/news/Graphics-Cards/AFDS11-Upcoming-Trinity-APU-will-use-VLIW4-Cayman-Architecture
Wow! I was ready to take bets there's going to be a 6570 inside Trinity! :shock:

Given how it was said the IGP in Trinity is going to be 50% faster, most likely meaning more GFLOPS, I'd say it's going to be 512 shaders, with higher clocks than now. Of course one dispatch processor with one set of fixed hardware.

That's how probably the slowest SKU in 7XXX series is going to look like, allowing for hybrid crossfire with Trinity.

ToTTenTranz

15-Jun-2011, 11:15

Though for how long? Their nextgen desktop GPUs are moving completely away from VLIW, will next gen Fusion after Trinity do the same?

I guess the fGPU will be one generation behind dedicated GPUs in architecture changes.

The PCPer's comments state the Trinity's successor will use the new Fermi-ish architecture, but that's 2013 and the next-gen graphics architecture should be available in 2012.

But yeah, one wonder's how much "effort" is worth for AMD to be investing in application acceleration while changing the graphics architecture 3 times in 3 years in a row..

And how will they achieve crossfire between the fGPU and an apparently very different discrete GPU? Maybe AMD will go with Lucid at some point?

Alexko

15-Jun-2011, 11:16

Though for how long? Their nextgen desktop GPUs are moving completely away from VLIW, will next gen Fusion after Trinity do the same?

Probably, yes. That's the "problem" with APUs: both the CPU and the GPU are likely to be always lagging about a year behind the top CPU and the discrete GPU architectures, respectively.

So yeah, expect updated Bulldozer cores + a new Scalar+SIMD-based GPU on 22nm for the 2013 APU.

Perhaps as AMD gets used to their velocity thing and streamlines the process, this lag will be reduced to just 6 months or so, we'll see.

rpg.314

15-Jun-2011, 12:20

Lightman

15-Jun-2011, 12:21

Probably, yes. That's the "problem" with APUs: both the CPU and the GPU are likely to be always lagging about a year behind the top CPU and the discrete GPU architectures, respectively.

So yeah, expect updated Bulldozer cores + a new Scalar+SIMD-based GPU on 22nm for the 2013 APU.

Perhaps as AMD gets used to their velocity thing and streamlines the process, this lag will be reduced to just 6 months or so, we'll see.

But from what I've read Trinity is based on updated BD core, so in that regard no lag to high end CPU's.
Anyway with yearly cadence this still will be huge step up compared to current situation.
I think thanks to APU's PC gaming will again gain on importance simply because nearly every new Laptop will be capable of pushing modernish games with ease.

I recon Steam user base will hit 6 million logged in users in 2012/13 :shock:

itsmydamnation

15-Jun-2011, 12:23

Probably, yes. That's the "problem" with APUs: both the CPU and the GPU are likely to be always lagging about a year behind the top CPU and the discrete GPU architectures, respectively.

So yeah, expect updated Bulldozer cores + a new Scalar+SIMD-based GPU on 22nm for the 2013 APU.

Perhaps as AMD gets used to their velocity thing and streamlines the process, this lag will be reduced to just 6 months or so, we'll see.

every single bit of info from amd about trinity has it having "Enhanced Bulldozer cores" same cores as Komodo and it appears to have a VLIW4 GPU.

so its looking like when going forward it will be current CPU -1 GPU. thats true even today as bulldzer hasn't landed yet.

Alexko

15-Jun-2011, 15:02

every single bit of info from amd about trinity has it having "Enhanced Bulldozer cores" same cores as Komodo and it appears to have a VLIW4 GPU.

so its looking like when going forward it will be current CPU -1 GPU. thats true even today as bulldzer hasn't landed yet.

Ah, I must have missed that. Pretty good news!

mczak

15-Jun-2011, 16:14

I think Trinity will have 3 channels. Roadmaps are hinting exactly that, but they not 100% confirmed.
I doubt it. I think 3 channels is a no-go for notebooks and budget desktops alike.
Maybe if ddr3-2133 got more mainstream that could be used. I've got doubts about that (well it might be supported) and it's not much of an improvement over ddr3-1866 anyway.

Extra channel assuming DDR 1600 would give 12.8GB/s more and brought total to 38.4GB/s.
Triple channel ddr3-2133 would give twice the bandwidth of dual channel ddr3-1600 indeed, solving the bandwidth issue pretty much (for now). I just don't think it's realistic.

There is also that possibility of using L3 cache of CPU to ease memory requirements a bit. Of course we don't know yet if AMD is going to implement that for next gen APU.
I'm still thinking this makes a lot of sense. Though I wonder now if AMD will really do it now. Because a 2-module BD with a 8 SIMD VLIW4 GPU is going to be seriously large if it has the same cache organization as Zambezi (2MB L2 per module, 2MB L3 chunk per module) - larger than Llano in any case, closer to 250mm² probably (though I guess that would be managable, AMD certainly did with Phenom II Danebs). They could skip L3 cache again and go for similar die size to Llano instead, though without L3 cache and only two memory channels the GPU would be very very bandwidth limited.

Trinity will use a VLIW4 GPU.

PCPer's slides on the Fusion Developer Summit[/URL] mention a 50% increase in the APU's GFLOPs, so we're probably talking about a ~512sp VLIW4 GPU, maybe with higher base clocks for the GPU.
Bringing the pieces together, with 512SP VLIW4 GPU, that'll be 8 SIMDs at ~700Mhz clock. Would also mean twice the texturing rate...
That thing is going to be quite ROP and bandwidth limited (I think there's only one other AMD GPU with such a high SIMD/ROP ratio, the RV730 though this has significantly less ALUs), but increasing ROPs isn't helpful unless bandwidth is also significantly increased (using L3 probably being the most viable option for that).

fellix

15-Jun-2011, 16:18

http://img339.imageshack.us/img339/8791/unled2uh.png

swaaye

15-Jun-2011, 19:07

According to Anandtech, Toshiba is going to be releasing a 13.3" Llano-based notebook. I think that is an excellent form factor to see this chip in.

http://www.anandtech.com/show/4449/annual-toshiba-refresh-brings-llanoand-some-style

Dave Baumann

15-Jun-2011, 19:11

http://www.hexus.net/content/item.php?item=30830

There are a number of HP ones already announced to be coming as well.

swaaye

15-Jun-2011, 19:21

They look to be 14" and larger but here here's a link with some pics of those HP notes.
http://www.ubergizmo.com/2011/06/hp-notebooks-with-amd/

mczak

15-Jun-2011, 20:54

They look to be 14" and larger but here here's a link with some pics of those HP notes.

Well I think 14" is an excellent form factor for Llano too. Only rarely do you get anything better than IGP or lowest-end chips (Caicos, GF119) in them (Llano easily beats both of them, at least the faster versions).
Larger than that though and it has to sell almost exclusively on price (as there will be options available which have faster discrete gpus).

mczak

15-Jun-2011, 22:11

Crysis Warhead 1280: ~25%
Crysis Warhead 1024: ~20%
Metro 2033 1280: ~28%
Metro 2033 1024: ~29%
Dirt2 1280: ~19%
Dirt2 1024: ~22%
Mass Effect 2 1280: ~25%
Mass Effect 2 1024: ~15%

edit:
They should test with 1280x720 & 1366x768 though, rather than 1280x1024 / 1024x768, since Llano is obviously going to be used with widescreen displays, let it be TV or monitor, rather than old 5:4/4:3 monitors
Ok my "just above 20% on average" was just guessed from looking at the scores :-).
FWIW anand has also published some ddr3 scaling numbers including ddr3-1333, ddr3-1600, ddr3-1866:
http://images.anandtech.com/reviews/cpu/amd/llano/review/desktop/ddr3scaling-16x10.png - also using widescreen res (1650x1080 though I'm wondering if that shouldn'be be 1680x1050...).

Game - ddr3-1600 vs ddr3-1333 - ddr3-1866 vs ddr3-1600:
Crysis Warhead: 114% 106%
Metro 2033: 115% 109%
Hawx: 116% 110%
Civ 5: 107% 102%
Dirt2: 113% 105%
Mass Effect 2: 117% 105%
Starcraft II: 117% 101%

So going from ddr3-1333 to ddr3-1600 scaling seems exceptional - for 20% more memory bandwidth you get 13% - 17% more performance, with only one outlier (Civ 5 only 7% - maybe even cpu limited?).
The additional 17% from ddr3-1866 doesn't nearly pay off quite that well (max another 10% only) but it's still a solid increase overall. Not sure though why it's that worse compared to the increase from ddr3-1333 to ddr3-1600 you'd expect maybe slightly worse scaling due to diminishing returns but not that much worse. Though no latencies are listed, I wouldn't expect this to be really latency sensitive but it could play some role here.
But either way at least the desktop A8 makes not much sense without at least (dual-channel, of course!) ddr3-1600, considering current memory prices.

Lightman

16-Jun-2011, 12:52

Ok my "just above 20% on average" was just guessed from looking at the scores :-).
FWIW anand has also published some ddr3 scaling numbers including ddr3-1333, ddr3-1600, ddr3-1866:
http://images.anandtech.com/reviews/cpu/amd/llano/review/desktop/ddr3scaling-16x10.png - also using widescreen res (1650x1080 though I'm wondering if that shouldn'be be 1680x1050...).

Game - ddr3-1600 vs ddr3-1333 - ddr3-1866 vs ddr3-1600:
Crysis Warhead: 114% 106%
Metro 2033: 115% 109%
Hawx: 116% 110%
Civ 5: 107% 102%
Dirt2: 113% 105%
Mass Effect 2: 117% 105%
Starcraft II: 117% 101%

So going from ddr3-1333 to ddr3-1600 scaling seems exceptional - for 20% more memory bandwidth you get 13% - 17% more performance, with only one outlier (Civ 5 only 7% - maybe even cpu limited?).
The additional 17% from ddr3-1866 doesn't nearly pay off quite that well (max another 10% only) but it's still a solid increase overall. Not sure though why it's that worse compared to the increase from ddr3-1333 to ddr3-1600 you'd expect maybe slightly worse scaling due to diminishing returns but not that much worse. Though no latencies are listed, I wouldn't expect this to be really latency sensitive but it could play some role here.
But either way at least the desktop A8 makes not much sense without at least (dual-channel, of course!) ddr3-1600, considering current memory prices.

It was CL7 for 1333 and 1600 and CL9 for 1833, so you might be right.
More tests are needed with varied CL settings and memory speeds.

But this is good already. We see that investing in decent RAM for desktop Llano will be the right thing.

ToTTenTranz

16-Jun-2011, 14:18

It was CL7 for 1333 and 1600 and CL9 for 1833, so you might be right.
More tests are needed with varied CL settings and memory speeds.

But this is good already. We see that investing in decent RAM for desktop Llano will be the right thing.

But does Llano make sense in any desktop other than HTPCs or nettops?

I have higher hopes for notebook makers to start using DDR3-1600MHz instead.

I know that if I buy a 1st-gen Llano notebook, I'll be spending the little extra to put DDR3-1600MHz in there.

BTW, I have a question.
DX9 titles require to have the graphics card's memory replicated in the system's main RAM.
Does it mean that, in an UMA environment, the graphics memory needs to be duplicated? (May sound like a stupid question.. I'm just making sure though)

And how does that differ in OpenGL, if at all?

Lightman

16-Jun-2011, 15:46

@ToTTenTranz

Llano in desktop is as valid as Phenom II. Market will force right price for Llano to be an option otherwise which mobo manufacturer would bother with full size ATX designs?

We don't know yet how well Llano might clock compared to Phenoms but indications are that even locked parts will hit 3.8-4.2GHz mark on AIR.
Considering poorer countries markets where $20 is a big difference I expect Llano to be quite popular as a good enough gaming desktop platform / computer for school, with affordable path to upgrade GPU.

Besides when GPGPU gets some traction this will be even better for desktop than now. Of course by that time we will be talking about Trinity or even next next gen APU's.

Blazkowicz

16-Jun-2011, 16:34

is there a word on IOMMU support?, AMD has been quite quiet on this. ditto on output redirection from a dedicated GPU.

mczak

16-Jun-2011, 16:39

But does Llano make sense in any desktop other than HTPCs or nettops?

Why not? The cpu doesn't quite hold up to Sandy Bridge i3 overall (certainly not in single threaded loads despite i3 doesn't turbo), though it can beat it in multithreaded workloads, just like the Phenom II of similar frequency.
And it should enable fairly cheap well equipped boxes (integrated USB3 for instance probably saves a buck or two), with decent graphics (don't forget most intel desktop chips only have HD2000, not HD3000 graphics, though there are now cheaper ones (i3-2105) which however cost a little extra).
Granted if you look at boxes with faster (discrete) graphics it loses quite some appeal, but those boxes are probably more highend than what this chip was intended for anyway - that's what BD is (hopefully...) for.

Blazkowicz

16-Jun-2011, 17:00

even the dual core llano is quite good and will feel at home in a standard and cheap micro ATX tower.
it looks old (like an athlon II X2 or a fast core2duo) but many people will be upgrading from lower (some pentium 4 or pentium D with 2x512MB ddr2, or a crumbling laptop)

for those people if you build a setup with a 2C llano, 8GB memory, 64GB ssd and 1TB hard drive that will look like an insanely fast, silent and powerful computer. (it is)

ToTTenTranz

16-Jun-2011, 17:33

Are we all assuming that Llano will replace Athlon II + 880G systems, at the same price?

I thought it'd go to a higher price point.

swaaye

16-Jun-2011, 19:17

Well I do think it is a decent replacement for Athlon II. Also, at this point I think that Phenom II is somewhat pointless unless you can make good use of the X6. Llano on the desktop makes a decent budget rig.

But I wouldn't go the Llano route for myself. I want a 4.5 GHz Core i5 or i7 setup at some point here. ;) Ivy Bridge probably. I certainly don't want something that's slower in any way than my old Q6600 @ 3.0.

mczak

16-Jun-2011, 19:23

Are we all assuming that Llano will replace Athlon II + 880G systems, at the same price?

I thought it'd go to a higher price point.
Llano goes down to dual core, with presumably 160 shader gpu. If the rumored sub 3.0Ghz cpu clocks are true (desktop Llano data is unconfirmed) cpu performance won't be any bit faster than the predecessor (could well end up slightly slower actually unless they up the clocks to same A II levels), though you get a much improved gpu (but it will be back to "better than intel" rather than "quite good" with only 160 shader units).
I can't see how that could really occupy a higher price point.
The faster Llanos should sell for more of course, I guess Phenom II X4 territory (which is still not all that much).

fehu

16-Jun-2011, 22:23

Also, at this point I think that Phenom II is somewhat pointless unless you can make good use of the X6.

Phenoms are simply faster

swaaye

17-Jun-2011, 01:33

Phenoms are simply faster
I just think that Phenom II X4 is not such a great value compared to Athlon II X4 because PII is really not that much faster and it is considerably slower than the Core i5/i7 chips.

But the X6 is a pretty great value if you can use all those cores. X6 has Turbo too so it's not bad for even less threaded tasks. Intel wants some serious money for their 6 core whereas poor AMD is shipping theirs under $200 (can't be making much money there.)

3dcgi

17-Jun-2011, 05:39

BTW, I have a question.
DX9 titles require to have the graphics card's memory replicated in the system's main RAM.
Does it mean that, in an UMA environment, the graphics memory needs to be duplicated? (May sound like a stupid question.. I'm just making sure though)

And how does that differ in OpenGL, if at all?
Are you sure DX9 requires a graphics card's memory to be replicated? That doesn't seem right.

fehu

17-Jun-2011, 11:24

...

yeah
the delta between highend and lowend is so thin that if i'm going to build a pc now i would go for phenom and a 65XX video card for better performance and about the same price
at the moment the problem for an allrange offer is the cpu clock, at least for me

Blazkowicz

17-Jun-2011, 16:19

I just think that Phenom II X4 is not such a great value compared to Athlon II X4 because PII is really not that much faster and it is considerably slower than the Core i5/i7 chips.

But the X6 is a pretty great value if you can use all those cores. X6 has Turbo too so it's not bad for even less threaded tasks. Intel wants some serious money for their 6 core whereas poor AMD is shipping theirs under $200 (can't be making much money there.)

the phenom 955BE has dropped in price, down to 99 euros. its clock multiplier is still unlocked. so you can still reasonably buy one.

for value the X3 would be the better one, as it turns into an X4 with a BIOS setting, and even if it doesn't work you have the gaming needs mostly covered

liolio

23-Jun-2011, 10:16

Techreport review of llano is up for those interested.

rpg.314

24-Jun-2011, 17:28

http://www.dailytech.com/article.aspx?newsid=22002

Intel claims it will match today's discrete graphics with Haswell. I am guessing they are shooting for 6450 level of performance.

Alexko

24-Jun-2011, 19:05

I don't know, they also said "12× improvement in 2015"… I think they clearly intend to take it up a notch, but do they hope to match AMD?

mczak

24-Jun-2011, 19:32

Intel claims it will match today's discrete graphics with Haswell. I am guessing they are shooting for 6450 level of performance.
That statement is very very vague obviously (since "today" discrete graphics would include everything from HD5450 to GTX 580). If they are shooting for HD5450 level performance, they can do that today (pretty much).
I think though Haswell aiming just for 6450 performance is a bit low. Ok depends if you're talking review-edition 6450 or for-sale-edition 6450 (quite a difference there) but given the timeframe I can't see why intel wouldn't aim for a bit more (something about as fast as Llano). Ivy Bridge might be enough already to match hd6450 ddr3.

GZ007

24-Jun-2011, 19:55

Maybe they should first begin with filtering ?

from http://techreport.com/articles.x/21099/11
http://techreport.com/r.x/llano/ansio-pattern-intel.png

Harison

24-Jun-2011, 20:04

mczak

24-Jun-2011, 20:30

Maybe they should first begin with filtering ?

Granted that's cheap. But hey at least no obvious brilinear cheating :-).
If you compare that to amd's past efforts, it's worse than r300 (higher angle dependency). It is better than r200 though, which had the same angle dependency but could do AF only with bilinear filtering...
I can understand intel though there, most people probably still don't care all that much about AF, so they used a cheap implementation, and I believe it's the same since the original Gen X chipset (i965). AF is also a patent mess, so maybe that implementation avoids some of it. I won't argue though that it's probably time for a better algorithm, wouldn't surprise me if it got improved a bit in Ivy Bridge or Haswell.

swaaye

24-Jun-2011, 20:50

This is what my X850 puts out. I ran this a few months ago. ;)
0 - 16X AF.

http://img140.imageshack.us/img140/248/d3daftesterx850x1.th.png (http://img140.imageshack.us/i/d3daftesterx850x1.png/)http://img715.imageshack.us/img715/7029/d3daftesterx850x2.th.png (http://img715.imageshack.us/i/d3daftesterx850x2.png/)http://img101.imageshack.us/img101/9610/d3daftesterx850x4.th.png (http://img101.imageshack.us/i/d3daftesterx850x4.png/)http://img715.imageshack.us/img715/6559/d3daftesterx850x8.th.png (http://img715.imageshack.us/i/d3daftesterx850x8.png/)http://img529.imageshack.us/img529/3910/d3daftesterx850.th.png (http://img529.imageshack.us/i/d3daftesterx850.png/)

The more distressing aspect to TR's comparison was the in-game shot with a clear loss of detail on the Intel IGP. He also mentions some dithering/filtering artifacts. Intel is just skimping on quality, surely because it costs transistors. In the grand scheme I suppose their hardware does get the job done though.

rpg.314

25-Jun-2011, 05:32

Considering Intel's iGPU progress and 22nm shrink, IMO they are targeting ~6650 performance, which still would be pretty decent.

By that time AMD will have APU with GCN generation. There are also similarities between Haswell and GCN, like vector/scalar design.
The vector/scalar design is upto them. But 66xx level perf in 2013 isn't exactly acceptable, considering AMD's graphics.

Matching AMD's hypothetical 86xx is plainly impossible without some kind of on package dram.

Erinyes

25-Jun-2011, 05:35

That statement is very very vague obviously (since "today" discrete graphics would include everything from HD5450 to GTX 580). If they are shooting for HD5450 level performance, they can do that today (pretty much).
I think though Haswell aiming just for 6450 performance is a bit low. Ok depends if you're talking review-edition 6450 or for-sale-edition 6450 (quite a difference there) but given the timeframe I can't see why intel wouldn't aim for a bit more (something about as fast as Llano). Ivy Bridge might be enough already to match hd6450 ddr3.

Pretty much? Intel HD 3000 beats HD 5450 performance today. And i expect Ivy bridge will be at least 50% faster. Ivy bridge will definitely be faster than 6450 and close to Llano performance already. If they really want to up the ante they could double the performance from Ivy bridge to Haswell (similar to how performance doubled from Arrandale to Sandy Bridge)

rpg.314

25-Jun-2011, 06:04

Pretty much? Intel HD 3000 beats HD 5450 performance today. And i expect Ivy bridge will be at least 50% faster. Ivy bridge will definitely be faster than 6450 and close to Llano performance already. If they really want to up the ante they could double the performance from Ivy bridge to Haswell (similar to how performance doubled from Arrandale to Sandy Bridge)

IB increases the EUs by 33%. While clocks are likely to be more as well, it appears that Intel does not intend to increase it's IGP performance beyond what moore's law already provides.

In this context, Intel's claim of 7x better graphics by Haswell sounds dubious.

Besides, IB's competitor is not Llano, it is trinity. Which will not have a lame CPU, hopefully.

Erinyes

25-Jun-2011, 08:24

IB increases the EUs by 33%. While clocks are likely to be more as well, it appears that Intel does not intend to increase it's IGP performance beyond what moore's law already provides.

In this context, Intel's claim of 7x better graphics by Haswell sounds dubious.

Besides, IB's competitor is not Llano, it is trinity. Which will not have a lame CPU, hopefully.

Yep my 50 % estimate was considering the 33% increase in EU's and say 15-20% increase in clocks(possibly also greater perf/EU as they're going for DX11 compliance as well).

Have Intel publicly mentioned that they're targeting 7x over SB? I havent come across that anywhere

I expect Llano will still be competing with IB for at least a quarter. I dont expect Trinity to be out until late Q2. IB is slated to be out in early Q2 as per the latest roadmap. And given how slow AMD have been to ramp up Llano, it remains to be seen how fast they can ramp up Trinity production. Yes the CPU performance will be far better hopefully, and with better clocks as the process matures

Lightman

25-Jun-2011, 09:21

Yep my 50 % estimate was considering the 33% increase in EU's and say 15-20% increase in clocks(possibly also greater perf/EU as they're going for DX11 compliance as well).

Have Intel publicly mentioned that they're targeting 7x over SB? I havent come across that anywhere

I expect Llano will still be competing with IB for at least a quarter. I dont expect Trinity to be out until late Q2. IB is slated to be out in early Q2 as per the latest roadmap. And given how slow AMD have been to ramp up Llano, it remains to be seen how fast they can ramp up Trinity production. Yes the CPU performance will be far better hopefully, and with better clocks as the process matures

Re: ramping up on AMD side. With Llano GF was fighting with new process for them as well as AMD designers were fighting with porting GPU logic from TSMC bulk to GF SOI with Gate First.
Path for Trinity will be a lot easier because they are staying on 32nm SOI for it. AMD already have working silicon in their labs for some time. Last year this time I doubt AMD had working Llano in their labs.

Drazick

25-Jun-2011, 09:35

Pretty much? Intel HD 3000 beats HD 5450 performance today. And i expect Ivy bridge will be at least 50% faster. Ivy bridge will definitely be faster than 6450 and close to Llano performance already. If they really want to up the ante they could double the performance from Ivy bridge to Haswell (similar to how performance doubled from Arrandale to Sandy Bridge)

Maybe I'm wrong but I think they weren't talking about the Graphics Core.
They meant the CPU performance for Vector Floating Point will exceed the performance of current discrete graphics cards.
Add that to the inherit advantages of CPU's and you'll get the best platform for the scenarios they mentioned (scientific and engineering numerical applications, visual processing, recognition, data-mining/synthesis, gaming, physics, cryptography and other areas of applications).

Alexko

25-Jun-2011, 10:03

mczak

25-Jun-2011, 13:41

Pretty much? Intel HD 3000 beats HD 5450 performance today.

That depends, my impression was it was about the same performance overall.

And i expect Ivy bridge will be at least 50% faster. Ivy bridge will definitely be faster than 6450 and close to Llano performance already.
50% faster is not enough to beat 6450 (well not the review edition). It is enough to beat the 6450 ddr3 however.
That would be close to the lower-end Llanos but not quite up there with the 400 SP versions.

rpg.314

25-Jun-2011, 13:49

no-X

26-Jun-2011, 12:04

Maybe they should first begin with filtering ?

from http://techreport.com/articles.x/21099/11
http://techreport.com/r.x/llano/ansio-pattern-intel.png
Angle dependancy is exactly the same as on ATI's DX8.1 generation hardware (~2001). The only difference seems to be support for trilinear AF :smile:

Erinyes

28-Jun-2011, 08:04

Isn't Ivy Bridge supposed to double EU count?

Anyway, I imagine Intel intends to stack some DRAM on their APUs…

Nope Ivy bridge increases EU's from 12 to 16.

DRAM Stacking according to Charlie is coming with Haswell.

That depends, my impression was it was about the same performance overall.

50% faster is not enough to beat 6450 (well not the review edition). It is enough to beat the 6450 ddr3 however.
That would be close to the lower-end Llanos but not quite up there with the 400 SP versions.

Nope, in most cases the HD 3000 is 10-20% faster than HD 5450(DDR3). And the 6450(GDDR5) is about 40-50% faster than the HD 3000. It varies but thats pretty much the case.

http://www.anandtech.com/show/4263/amds-radeon-hd-6450-uvd3-meets-htpc/5

Look at their ultrabooks presentation.

Saw that. They state 7X graphics improvement for ultrabooks in general. Given that there are a few ultrabooks already shipping this generation, do we even know what they are comparing to when they say 7X graphics improvement?

mczak

28-Jun-2011, 15:40

Nope, in most cases the HD 3000 is 10-20% faster than HD 5450(DDR3). And the 6450(GDDR5) is about 40-50% faster than the HD 3000. It varies but thats pretty much the case.

http://www.anandtech.com/show/4263/amds-radeon-hd-6450-uvd3-meets-htpc/5

Well because the performance is still all over the place compared to competition it really varies depending on what games were tested. I think though newer drivers improved it quite a bit in some titles, so an improvement of 50% of IVB over SB might really be enough to beat HD6450 gddr5, and not just draw about even.

fehu

28-Jun-2011, 16:41

swaaye

28-Jun-2011, 17:42

Well since nobody can get realtime clocks out of the CPU, it's tough to verify how their Turbo is working. Performance in single threaded apps does seem to look like it should for a CPU like a low end desktop Athlon II though. Tech Report made it sound like their Turbo is still rather rudimentary compared to Intel's however (using load instead of thermal sensors).

I have a ASUS G73 with the Core i7 720QM, the 2009 low end i7 for notebooks. It's a 1.6->2.8 GHz CPU. I wonder how it compares to Llano. I have a feeling that the 720QM is faster unfortunately.

I know there are a bunch of Llano notebooks available now but I'm having a hard time figuring out how to find all of them. I'm mainly searching out 13.3" or smaller. All that I see though are just new budget 14"+ options.

Lightman

28-Jun-2011, 18:17

I need to understand how Llano turbo core works in real world scenarios

My notebook can't boot no more and is out of warranty, and just today i have seen on a depliant the asus x53ta-sx016v equipped with A6-3400m (1.4GHz/2.3GHz) and a 6650 1GB for 499€.
My last notebook had a dual core turion2 M520 running at 2.3GHz, and was barely fine.

When the A6 will run at 2.3GHz? and how many cores? would it depend on he gpu workload?
would it be a continuous overclock or only the "some clock" one?

I've tried online but there are not two identical explanations, and even the review apperas to be made on defective pre production sample and not on retail notebook

:?::?::?:

Idea: but the laptop and educate us :grin:
I'm curious about it as well.
In case you can sustain 2.3GHz on Llano core it will easily outperform Turion M2 clock for clock. But base clock of 1.4GHz is on a low side ...
I'm on T5500 1.6GHz lappy so for me it would be massive upgrade, but I will go for 45W models once available with decent base clock.

ToTTenTranz

28-Jun-2011, 18:24

The 720QM is a quad-core, eight-trheaded i7 Clarksfield, definitely not a "low-end i7" as there are many dual-core i7 models for notebooks (no Apple laptop has a quad-core i7, for example).
Furthermore, it turboes with all cores activated to 1.86GHz.

Given the substantially higher performance-per-clock in the Nehalem architecture compared to Phenom II, it's no surprise that the i720QM will be faster than any mobile Llano.

Nonetheless, you're comparing two completely different power envelopes. 720QM's TDP is 50W for CPU alone, whereas the top performing mobile Llano (A8-3530MX) does 45W for CPU+(midrange)GPU+Northbridge.

I have no doubts we'll find Llanos in 13.3" notebooks. The killer combo of Llano + discrete GPU should be harder to find in that form factor, though.

swaaye

28-Jun-2011, 18:39

I don't really care for discrete graphics in 12-13.3" because that usually creates a lot of heat that gets poorly dissipated. I've been there. That's why Llano is interesting to me in small notebooks. The CPU is decent, it has an excellent GPU built-in, and should be cooler and more battery saving than the usual CPU + discrete GPU subnotes.

BTW, I still get a Cyrix MediaGX feeling from these CPUs with IGPs and northbridges inside. ;)

fehu

28-Jun-2011, 20:03

from what i understand reading the sildes and the (bios flawed) review, the maximum overclock happens when the gpu is at the lowest usage, and temporally based on the TC2 algorithm
then it scales down when reactivate a core or the gpu
what means "maximum overclock" it's a mistery for me, and how fast scales down too
i don't understand if in a scenario i can count on at least two cores running a 2.3GHz

i don't understand the llano positioning too
doing better research i've found an hp with N930 and switchable 5650 for 489€
it's less elegant, consumes a little more, but overall is more powerfull and cost less 0_0

swaaye

28-Jun-2011, 21:32

mczak

28-Jun-2011, 21:59

I do wish somebody would take a close look at its Turbo behavior in various scenarios. For example I don't see how a Core 2 Duo E6400 (2.13 GHz Conroe) could beat A8-3500M (2.4 GHz max) by ~80% (http://techreport.com/articles.x/21099/8) if it was clocking up properly. This test is a single core affair AFAIK.

From all the numbers I've seen, I came to the conclusion that Llano Turbo Core pretty much never reaches the Turbo Clock. Well not for significant amounts of time, not even when only one core is really used. At best it seems to give some speedup which is somewhere between the low and high clock for single-threaded benchmarks. I don't know though if that behavior could be tweaked with microcode/bios updates or if it's just that bad.

swaaye

28-Jun-2011, 22:29

I looked over the Anandtech desktop Llano performance (http://www.anandtech.com/show/4448/amd-llano-desktop-performance-preview) again. I wonder if what's happening to the mobile version is that they are trying to fit it into a TDP that's just too low for the design and as a result the Turbo Core is severely power restricted and essentially not helpful. The load-based design of Turbo Core may also just not be good enough at taking advantage of light load opportunities.

Too many possibilities...

fehu

28-Jun-2011, 23:24

if so why don't make a more aggresively clocked dual core?

cal_guy

29-Jun-2011, 00:04

Garlic operates on the Northbridge clock - up to 720MHz for notebook versions
http://www.realworldtech.com/page.cfm?ArticleID=RWT062711124854&p=2

That's much slower than the Phenom and Athlon right?

Kaotik

29-Jun-2011, 00:27

Garlic operates on the Northbridge clock - up to 720MHz for notebook versions
http://www.realworldtech.com/page.cfm?ArticleID=RWT062711124854&p=2

That's much slower than the Phenom and Athlon right?

I don't know if it's faster or slower than their internal busses, but gives far higher bandwidth than Phenom II's or Athlon II's have in their hand to memory (or the bandwidth Llano has to memory)

Phenom II X6 at Legion Hardware (first bench from google) got ~13.6GB/s measured memory bandwidth when accompanied by DDR3 PC3-12800, while according to the link you provided Llano gets around 17GB/s measured

Blazkowicz

29-Jun-2011, 06:32

Angle dependancy is exactly the same as on ATI's DX8.1 generation hardware (~2001). The only difference seems to be support for trilinear AF :smile:

I remember the radeon 9200SE, it was good enough to run the half-life 1 engine but not quite, and this is why. the slider control in the driver claiming 16x aniso filtering didn't make it look really good.
a geforce 3 or 4's 8x filtering was vastly better.

CarstenS

29-Jun-2011, 08:41

… and vastly more expensive in terms of performance hit.

Alexko

29-Jun-2011, 15:21

Nope Ivy bridge increases EU's from 12 to 16.

DRAM Stacking according to Charlie is coming with Haswell.

That's unfortunate. If Trinity does indeed have 512 VLIW4 shaders, unless it runs into severe memory bandwidth limitations, it might actually widen the graphics performance gap with Ivy Bridge.

Then again, 22nm may allow for significantly higher clocks.

Arun

29-Jun-2011, 17:00

That's unfortunate. If Trinity does indeed have 512 VLIW4 shaders, unless it runs into severe memory bandwidth limitations, it might actually widen the graphics performance gap with Ivy Bridge.But it's on the same process as Llano so that extra performance won't come for free - it will be a pretty big chip. I wouldn't be surprised if AMD kept selling Llano for quite some time.

ToTTenTranz

29-Jun-2011, 17:51

But it's on the same process as Llano so that extra performance won't come for free - it will be a pretty big chip. I wouldn't be surprised if AMD kept selling Llano for quite some time.

Agreed. Trinity probably isn't going to 110->150€ price points. When it comes out, Llano will probably lower to 80->130€ and Trinity will fill the gap between Llano and high-clocked AM3+ Bulldozers, so maybe it's more like 140->180€.

Erinyes

29-Jun-2011, 17:59

But it's on the same process as Llano so that extra performance won't come for free - it will be a pretty big chip. I wouldn't be surprised if AMD kept selling Llano for quite some time.

Also Trinity i presume is a two module, four "core" design. Im not sure they have a four module, 8 core Trinity planned as that would have an extremely large die size. Llano is available in a "true" four core design. Now while AMD Marketing might say both have four "cores", at the same clocks Llano might actually be faster than Trinity in situations that dont take advantage of BD's cores. In that case Llano might actually be a better alternative!

And as you said i expect they will continue to sell Llano for a while until Trinity production ramps up, and as the process matures they'll be able to up the clocks as well. But in Terms of compute, Trinity should be a LOT faster and given how heavily they're promoting GPGPU/Fusion, you'd expect that they'd want to ramp it up fast.

Alexko

29-Jun-2011, 18:09

But it's on the same process as Llano so that extra performance won't come for free - it will be a pretty big chip. I wouldn't be surprised if AMD kept selling Llano for quite some time.

That's a valid point, but on the other hand, Llano is the very first GPU on a GloFo, SOI process ever. So I wouldn't rule out some density improvements in Trinity.

Also Trinity i presume is a two module, four "core" design. Im not sure they have a four module, 8 core Trinity planned as that would have an extremely large die size. Llano is available in a "true" four core design. Now while AMD Marketing might say both have four "cores", at the same clocks Llano might actually be faster than Trinity in situations that dont take advantage of BD's cores. In that case Llano might actually be a better alternative!

4 Bulldozer cores > 4 Stars cores, AMD has been pretty clear about that. And clocks don't matter, only power does.

And as you said i expect they will continue to sell Llano for a while until Trinity production ramps up, and as the process matures they'll be able to up the clocks as well. But in Terms of compute, Trinity should be a LOT faster and given how heavily they're promoting GPGPU/Fusion, you'd expect that they'd want to ramp it up fast.

I think dual- and quad-core versions of Trinity should be able fill the entire notebook and mainstream desktop market without any need for Llano. Of course they'll need some time to ramp up production, but that's always the case with new chips.

swaaye

29-Jun-2011, 18:12

Lightman

29-Jun-2011, 20:09

Llano could end up being like the Core Duo - replaced by a far superior chip after a short period of time on the market. Llano does feel like a semi-failure in notebooks because of its ineffectual Turbo Core. Maybe Trinity will be much more capable in that respect.

Have you seen any review of retail laptops to back that statement up?
I know judging Llano CPU performance by SB performance will paint this picture, but no one expected Llano to win there.
If Turbo is as 'broken' in retail parts as it was in review samples then I'm with you on that for 35W parts. 45W parts are clocked much higher which gives a bit of hope for reasonable single thread performance.

Erinyes

29-Jun-2011, 21:42

4 Bulldozer cores > 4 Stars cores, AMD has been pretty clear about that. And clocks don't matter, only power does.

Source? Ive never seen a claim like that. IIRC it was upto 80% improvement for a BD module over a single Phenom core for optimal workloads. And far less for non-optimal workloads

Have you seen any review of retail laptops to back that statement up?
I know judging Llano CPU performance by SB performance will paint this picture, but no one expected Llano to win there.
If Turbo is as 'broken' in retail parts as it was in review samples then I'm with you on that for 35W parts. 45W parts are clocked much higher which gives a bit of hope for reasonable single thread performance.

Whatever hardware was shipped to reviewers is the same as the retail hardware. Its not going to magically change between the review and retail availability. Its likely been in production for months already. But all the reviews ive seen so far are for the 35W A8-3500M. Would have been better for AMD to send out the higher performance 45W A8-3510 MX CPU's to the reviewers (of course the battery life would have been worse so maybe thats why they didnt)

And the turbo cant be "broken" in either, its a hardware based turbo which measures power consumption in the chip.

swaaye

29-Jun-2011, 22:10

Yeah I don't think the Turbo is broken. I think that with the 35W chips it is so restricted that it can barely do anything. Hopefully they can get more from 32nm with new chips down the road. Maybe dual cores are the answer or perhaps Trinity will be more power efficient.

But since I haven't seen a review of a Llano with both Turbo Core and higher TDP, maybe the Turbo is not very efficient/effective. It's possible. The Anandtech desktop test uses the top end Llano that lacks Turbo Core and has a 100W TDP.

Another question is why did they stop at 2.9 GHz? And why not give the 2.9 GHz desktop chip Turbo?

AlexV

29-Jun-2011, 22:12

4 Bulldozer cores > 4 Stars cores, AMD has been pretty clear about that.

About as clear as mud. Their design targets sure as hell don't imply that (speed racer, reduced FO4, knee of the curve in IPC etc.). Some of the stuff they've unveiled would hint at BD being better in the cases in which K8L was really crap (branch-prediction, coherence mechanisms, prefetching, front-end etc.), but how that impacts the average case is unclear. So at this point, you might say...maybe it is indeed better per clock, or maybe it's not.

Lightman

29-Jun-2011, 22:50

Source? Ive never seen a claim like that. IIRC it was upto 80% improvement for a BD module over a single Phenom core for optimal workloads. And far less for non-optimal workloads

No. It was 180% of single BD core versus single BD module. AMD never gave any performance estimate of BD vs. Thuban other than on server side you will get 50% throughput with 33% more cores.
What John Fruehe said privately on one of the forums was that BD IPC > Phenom II IPC. Nothing more, nothing less.

Whatever hardware was shipped to reviewers is the same as the retail hardware. Its not going to magically change between the review and retail availability. Its likely been in production for months already. But all the reviews ive seen so far are for the 35W A8-3500M. Would have been better for AMD to send out the higher performance 45W A8-3510 MX CPU's to the reviewers (of course the battery life would have been worse so maybe thats why they didnt)

And the turbo cant be "broken" in either, its a hardware based turbo which measures power consumption in the chip.

Nitpicking, but no again :razz:.
What do you mean by hardware based turbo? That it can't be set to behave in specific way? Is it hard locked for some reason? It must be programmable as Intels solution is.
Most of the reviews openly said AMD told them test platforms are not perfect. They had bugged BIOS causing problems and corruptions with Cross Fire modes for instance.
Besides Llano Turbo behaviour is also regulated by BIOS. If for instance AMD wanted to look good in tests where it's previous mobile platforms failed miserably, battery life, they could tweak BIOS Turbo settings to conserve as much power as possible and run very conservative Turbo.
Other possibility would be that Turbo core was not considering fGPU load into calculating available thermal headroom. It can be observed from tests where fGPU is disabled and system is running on dGPU. CPU performance is not affected.
I'm not expecting all this to be the case and retail systems might behave identically turbo wise to preview systems, but simply saying what we saw in previews is what we will get from retail machines is not wise.

rpg.314

30-Jun-2011, 05:34

That's a valid point, but on the other hand, Llano is the very first GPU on a GloFo, SOI process ever. So I wouldn't rule out some density improvements in Trinity.GPUs are synthesized blocks. So I wouldn't expect any non-trivial density improvement.

rpg.314

30-Jun-2011, 05:35

LordEC911

30-Jun-2011, 08:05

That's unfortunate. If Trinity does indeed have 512 VLIW4 shaders, unless it runs into severe memory bandwidth limitations, it might actually widen the graphics performance gap with Ivy Bridge.

Then again, 22nm may allow for significantly higher clocks.

They need to figure out how to break past the bandwidth barrier with Trinity.
I'm hoping for 896VLIW4.

No one should really care if the CPU power of Trinity is barely more than Llano.

CarstenS

30-Jun-2011, 08:58

Harison

30-Jun-2011, 12:47

I wonder, if stronger graphics are really worth the hassle. I mean, at least you have to tape out different chips, plan supply accordingly and have a multitude not only of SKUs but also production lines/wafer allocations. All of that costs money - and APUs are all about reduced cost and TCO, so far.

edit:
To be more precise: You'll never catch up to discrete graphics and CPU in terms of performance, so that market is not in danger from APUs. You'll have to carter for the peeps requiring "just enough" potential in CPU/GPU to browse the web and accelerate maybe a few things like HTML5 - and that's a pretty small target group that's sitting between what is sufficiently handled by a traditional IGP and what requires a dedicated GPU (like gaming outside of facebook). Just because most customers are not (willing to be) educated enough to do product research before buying - at least that's my experience outside of tech-savvy people like mother-in-law and the like.
I think you are asking the wrong question, it should be "if stronger CPU is really worth the hassle." :wink: For quite some time CPU's (even cheap-medium class) were good enough for majority of users. However even though integrated GPUs were sufficient for a part of the market (namely office), but it was barely good enough for something more demanding than Flash games.

Its changed now, you can buy APU which covers 80+% of users needs, including gaming (outside of Crysis fan base :smile:), GPU acceleration will be handy too. Even 1st generation APU's made low-middle(!) discreet graphics obsolete, few generations later, APU's will cover 95% users needs. Discreet graphics will be only for high-end gaming and professional market.

pjbliverpool

30-Jun-2011, 12:59

Its changed now, you can buy APU which covers 80+% of users needs, including gaming (outside of Crysis fan base :smile:), GPU acceleration will be handy too. Even 1st generation APU's made low-middle(!) discreet graphics obsolete, few generations later, APU's will cover 95% users needs. Discreet graphics will be only for high-end gaming and professional market.

It depends what you mean by high-end gaming. Llano is insufficient on the graphics side for most modern console ports at higher graphical settings. Even the CPU side of it is going to struggle with a lot of games.

I would define high end PC gaming as gaming with a discrete GPU which is powerful enough to max out virtually every game available at 1080p.

There's still a big gap between that and flash gamers consisting of people that want to be able to play most PC games and get them looking/running good without maxing everything out and spending a ton on hardware.

Gipsel

30-Jun-2011, 13:46

They need to figure out how to break past the bandwidth barrier with Trinity.
I'm hoping for 896VLIW4.That's what "Trinity Devastator Duo" is for! :lol:
:runaway:

CarstenS

30-Jun-2011, 14:17

I think you are asking the wrong question, it should be "if stronger CPU is really worth the hassle." :wink: For quite some time CPU's (even cheap-medium class) were good enough for majority of users. However even though integrated GPUs were sufficient for a part of the market (namely office), but it was barely good enough for something more demanding than Flash games.

Its changed now, you can buy APU which covers 80+% of users needs, including gaming (outside of Crysis fan base :smile:), GPU acceleration will be handy too. Even 1st generation APU's made low-middle(!) discreet graphics obsolete, few generations later, APU's will cover 95% users needs. Discreet graphics will be only for high-end gaming and professional market.

80%+ of users needs or the needs of 80+% of the users? :)
Anyway, I've seen nice HTML5/WebGL-Benchmarks showing 50+ fish animated on screen, I've seen nicely animated preview-walls of pictures displayed as demos of HTML5. But when it comes to what kind of a web browsing experience real people are having, it mostly looks like facebook, youtube and your news-portal of choice.

So, while I won't argue your point of 80+ to 95+ percentages I am still questioning how large the percentage is that's already content with what a traditional, but modern IGP-style (DX9/10-capable) graphics delivers - which I reckon to be north of 80% as well. And that would really narrow the window of opportunity down for AMD if I am right.

What AMD needs for the APU to be a great product is nothing less than a massive change in the software landscape - they've realized that themselves, having established their own developers conference now for the first time. But things change slowly in software; Nvidia for example pushed GPU-Computing really hard now for - what four, five years? And still, despite heterogeneous computing now crawling into supercomputers, for the end user all those efforts are basically exhausted with your stereotypical video conversion implementation via the drivers built-in libaries.

The problem is: As long as chipzilla decides to do nothing about it, developers won't be as tempted to work on complicated heterogeneous computing models for their products as long as only one quarter of the market will profit from it.

Harison

30-Jun-2011, 14:32

It depends what you mean by high-end gaming. Llano is insufficient on the graphics side for most modern console ports at higher graphical settings. Even the CPU side of it is going to struggle with a lot of games.
Consoles themselves have weaker graphics than faster Fusion APUs, and those can run most of the games just fine at 1680x and even 1920x. Not on high-details or with cranked up AAx, but good enough for majority of casual gamers, let alone office/home needs.

I would define high end PC gaming as gaming with a discrete GPU which is powerful enough to max out virtually every game available at 1080p.

There's still a big gap between that and flash gamers consisting of people that want to be able to play most PC games and get them looking/running good without maxing everything out and spending a ton on hardware.
Few generations later APU's will make inroads in higher-end gaming too, but for the most hardcore players there will always be discreet option as I mentioned. Low-middle-end discreet GPU's will be dead.

Harison

30-Jun-2011, 15:04

80%+ of users needs or the needs of 80+% of the users? :)
Anyway, I've seen nice HTML5/WebGL-Benchmarks showing 50+ fish animated on screen, I've seen nicely animated preview-walls of pictures displayed as demos of HTML5. But when it comes to what kind of a web browsing experience real people are having, it mostly looks like facebook, youtube and your news-portal of choice.

So, while I won't argue your point of 80+ to 95+ percentages I am still questioning how large the percentage is that's already content with what a traditional, but modern IGP-style (DX9/10-capable) graphics delivers - which I reckon to be north of 80% as well. And that would really narrow the window of opportunity down for AMD if I am right.

Sorry for my not perfect English, I meant "needs of 80+% of the users" :smile:

For office/internet of course weak/average CPU + IGP was enough, even ARMs will be :twisted: Thats why I mentioned majority of users dont even need faster CPU, if anything - IGP was the weak link. Have you ever bough PC or notebook with IGP and said "thats enough for my gaming needs"? I seriously doubt it.

APUs changing that. There is a massive demand for low-middle-end graphics (actually, its the highest volume discreet GPU market), and their needs will be met by APUs from now on. If 80% of market would be satisfied by IGP, there wouldnt be such demand for discreet cards.

What AMD needs for the APU to be a great product is nothing less than a massive change in the software landscape - they've realized that themselves, having established their own developers conference now for the first time. But things change slowly in software; Nvidia for example pushed GPU-Computing really hard now for - what four, five years? And still, despite heterogeneous computing now crawling into supercomputers, for the end user all those efforts are basically exhausted with your stereotypical video conversion implementation via the drivers built-in libaries.

The problem is: As long as chipzilla decides to do nothing about it, developers won't be as tempted to work on complicated heterogeneous computing models for their products as long as only one quarter of the market will profit from it.
Thats why AMD is working closely with Microsoft, Apple, etc. to make it happen. MS already presented tools for Fusion. Nvidia failed to make CUDA mass-spread because it was closed and exclusive for NV cards, nobody of big sharks cared about it. However MS and Apple are mass pushing OpenCL/DirectCompute, and will make sure they'll become standards for the whole industry.

AMD or NV by themselves cant do that, if anything, AMD's limited endeavors will just help to transition smoother in the direction market already is moving.

ToTTenTranz

30-Jun-2011, 15:17

It depends what you mean by high-end gaming. Llano is insufficient on the graphics side for most modern console ports at higher graphical settings. Even the CPU side of it is going to struggle with a lot of games.

I would define high end PC gaming as gaming with a discrete GPU which is powerful enough to max out virtually every game available at 1080p.

There's still a big gap between that and flash gamers consisting of people that want to be able to play most PC games and get them looking/running good without maxing everything out and spending a ton on hardware.

I think Llano's main target is to be on ~1366*768 13-15" laptop screens, and not 23" 1080p desktop screens.
From what we've seen, even mid-priced 15" laptops (~750€) will bundle discrete Whistlers for Crossfire goodness.
Performance is just right for the targetted resolution, as a "HD6755G2" will be able to max out most games - especially console ports - at 720/768p (they just need to get DX9 compatibility working, fast!).

For desktops, Llano is replacing Athlon II, so the iGPU won't be of much more use than the current HD3200\HD4200 IGP on most motherboards: an integrated GPU that provides "flawless" video playback and browser acceleration. At least until APP really starts to spread some wings, of course.

AnarchX

30-Jun-2011, 15:37

I think Llano's main target is to be on ~1366*768 13-15" laptop screens, and not 23" 1080p desktop screens.
From what we've seen, even mid-priced 15" laptops (~750€) will bundle discrete Whistlers for Crossfire goodness.
Performance is just right for the targetted resolution, as a "HD6755G2" will be able to max out most games - especially console ports - at 720/768p (they just need to get DX9 compatibility working, fast!).

If you like extreme micro-stuttering: http://translate.google.de/translate?u=http%3A%2F%2Fwww.computerbase.de%2Fart ikel%2Fgrafikkarten%2F2011%2Ftest-llano-grafikleistung%2F14%2F%23abschnitt_mikroruckler&sl=de&tl=en&hl=&ie=UTF-8

ToTTenTranz

30-Jun-2011, 17:04

If you like extreme micro-stuttering: http://translate.google.de/translate?u=http%3A%2F%2Fwww.computerbase.de%2Fart ikel%2Fgrafikkarten%2F2011%2Ftest-llano-grafikleistung%2F14%2F%23abschnitt_mikroruckler&sl=de&tl=en&hl=&ie=UTF-8

They should've enabled v-sync.

Alexko

30-Jun-2011, 18:30

GPUs are synthesized blocks. So I wouldn't expect any non-trivial density improvement.

Tell that to RV770… :p

AnarchX

30-Jun-2011, 18:35

They should've enabled v-sync.

Vsync is no solution to micro-stuttering.

ToTTenTranz

30-Jun-2011, 19:34

Vsync is no solution to micro-stuttering.

It was for me, when I used dual HD3870 and then dual HD4890.

swaaye

30-Jun-2011, 19:41

It's too bad that it's probably impractical to make some GDDR5 SODIMMs. ;) I don't even know if these chips support GDDR5. The lack of volume for a specialty DIMM like that would probably make the price crazy anyway.

AlexV

30-Jun-2011, 19:48

Vsync is no solution to micro-stuttering.

Curious what this particular conclusion is based on...

AnarchX

30-Jun-2011, 20:52

Curious what this particular conclusion is based on...

Because v-sync only homogenizes the output and not the input - the produced images, which contain the temporal jumps.

Kaotik

30-Jun-2011, 21:15

Vsync is no solution to micro-stuttering.

It is for many, many people.

edit:
Clarification: It doesn't remove the problem, however it does make it less severe, even to point where it becomes unnoticeable

AlexV

30-Jun-2011, 21:48

Because v-sync only homogenizes the output and not the input - the produced images, which contain the temporal jumps.

How does this work, exactly?

AlexV

30-Jun-2011, 21:49

It doesn't remove the problem, however it does make it less severe, even to point where it becomes unnoticeable

Unless you use triple-buffering, it should. The former produces MS even with single cards.

swaaye

30-Jun-2011, 22:15

triple buffering causes micro stuttering with a single card setup?

Kaotik

30-Jun-2011, 22:57

Unless you use triple-buffering, it should. The former produces MS even with single cards.

We did benchmarks on that, the frametime still jumps up and down with vsync on

Without
http://home.akku.tv/~akku38901/Stutteringtests/veUnigineTB-30sec.png
With
http://home.akku.tv/~akku38901/Stutteringtests/veUnigineTBVsync-30sec.png

Man from Atlantis

02-Jul-2011, 15:49

Haswell with Dx11.1 igp ?

can we expect AMD's 7000 and/or nVidia's 600 series support Dx11.1? or it's pretty early you think
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=8791048&postcount=245

Blazkowicz

02-Jul-2011, 16:17

maybe that's the single address space and coherency that AMD is adopting?

AlexV

02-Jul-2011, 16:35

We did benchmarks on that, the frametime still jumps up and down with vsync on

What are we looking at? Also, some statistical analysis on the dataset always helps:smile:. What's on 0y? I presume 0x is the frame-count? What are we measuring etc.?

Kaotik

02-Jul-2011, 16:46

What are we looking at? Also, some statistical analysis on the dataset always helps:smile:. What's on 0y? I presume 0x is the frame-count? What are we measuring etc.?

Oh ye, sorry.

X-axis is frame-count, Y-axis is frametime converted into FPS equivalent ("realtime FPS instead of average over second like usually")

This particular test is 30 second period from Unigine Heaven bench

rpg.314

02-Jul-2011, 16:46

Arun

03-Jul-2011, 17:11

Oh ye, sorry.

X-axis is frame-count, Y-axis is frametime converted into FPS equivalent ("realtime FPS instead of average over second like usually")

This particular test is 30 second period from Unigine Heaven benchI think it's noteworthy that the times where the FPS increases above 60 despite vsync are basically the exact same times that the FPS decreases substantially without vsync. So with vsync the "pre-rendered frames" buffer is being (partially?) depleted as performance drops temporarily, and then it gets filled back as fast as it can afterwards.

That is an inherent effect of all frame-based rendering and the same issue would exist with any setup. But I suppose it's a problem that AFR makes sudden temporary decreases in performance more likely, but I'm not sure by how much or how that graph would look with a single GPU.

Kaotik

03-Jul-2011, 21:55

Actually, that might be single card already, it's old data and the file naming doesn't give much :oops:
It has triple buffering enabled, too.

Here's double buffered with CrossFire 100% certain

Without v-sync
http://home.akku.tv/~akku38901/Stutteringtests/veUnigineCF-30sec.png

With v-sync
http://home.akku.tv/~akku38901/Stutteringtests/veUnigineCFVsync-30sec.png

Harison

05-Jul-2011, 14:05

Interesting leaked AMD slides:

http://img.donanimhaber.com//images/haber/27368/amdbrazosthondo_1a_dh_fx57.jpg

http://img.donanimhaber.com//images/haber/27368/amdbrazosthondo_2_dh_fx57.jpg

http://img.donanimhaber.com//images/haber/27368/amdbrazosthondo_3a_dh_fx57.jpg

Npl

05-Jul-2011, 14:10

hmm, still 40nm. I`d have hope for 28nm and atleast the same performance at ~half TDP

3dilettante

05-Jul-2011, 14:20

It looks like they shaved off a few watts by removing or reducing IO interfaces, and lowering the temperature from 75 to 60.

Removing interfaces is a tradeoff, changing the temperature sounds like moving the goalpost.

no-X

05-Jul-2011, 15:37

hmm, still 40nm. I`d have hope for 28nm and atleast the same performance at ~half TDP
Hondo was presented as a ~5W product from the beginning. Typical application power around 2 W (including chipset) isn't bad.

mczak

05-Jul-2011, 16:11

Harison

05-Jul-2011, 16:52

Strange I thought by that timeframe (q2/2012) we were supposed to already get the 28nm fusion chips (krishna/wichita) rather than the exact same Ontario chip (tuned for lower power mostly by disabling some i/o).
At Computex rumour was "Krishna APUs releasing in 2012 will bring TDP down to sub-1 watt", but I guess it depends on 28nm maturity and GCN. The fact they're postponing means ether 28nm mass manufacturing wont be ready for a while, or GCN tuning will take longer than expected, or both.

AMD simply going for safer approach, not to get burned by immature node and new gen graphics. If 4-5W APUs are doing just fine in tablets and netbooks, then 2W will have even more success. They can always speed up launch if everything goes well.

swaaye

05-Jul-2011, 17:15

Have I mentioned that the GCN acronym makes me think of Gamecube every single time? ;)

mczak

05-Jul-2011, 18:21

At Computex rumour was "Krishna APUs releasing in 2012 will bring TDP down to sub-1 watt", but I guess it depends on 28nm maturity and GCN. The fact they're postponing means ether 28nm mass manufacturing wont be ready for a while, or GCN tuning will take longer than expected, or both.

Well if 28nm isn't ready by Q2/2012 that wouldn't bode well for next gen graphics chips... Unless it's just not suitable in that time frame for this kind of chip (e.g. too expensive, too high leakage or whatnot).

AMD simply going for safer approach, not to get burned by immature node and new gen graphics. If 4-5W APUs are doing just fine in tablets and netbooks, then 2W will have even more success. They can always speed up launch if everything goes well.
Well for something which more looks like a new stepping rather than a new chip Q2/2012 looks awfully late indeed.

Erinyes

05-Jul-2011, 18:50

Well if 28nm isn't ready by Q2/2012 that wouldn't bode well for next gen graphics chips... Unless it's just not suitable in that time frame for this kind of chip (e.g. too expensive, too high leakage or whatnot).

Well for something which more looks like a new stepping rather than a new chip Q2/2012 looks awfully late indeed.

I have no doubt we'll see 28nm Krishna/Witchita APU's by Q2 2012. But you guys are skipping one crucial point. Those slides are all talking about the tablet optimized versions of the APU's which will lag the netbook class APU's by a few quarters due to the need for optimized I/O. Brazos-T is all about time to market, it will be available by Q2 2012 and is right in time for the back to school season and for the Windows 8 launch.

Blazkowicz

05-Jul-2011, 19:33

in that light a mature process feels needed for the optimizing and production volume goals.

mczak

05-Jul-2011, 20:25

I have no doubt we'll see 28nm Krishna/Witchita APU's by Q2 2012. But you guys are skipping one crucial point. Those slides are all talking about the tablet optimized versions of the APU's which will lag the netbook class APU's by a few quarters due to the need for optimized I/O. Brazos-T is all about time to market, it will be available by Q2 2012 and is right in time for the back to school season and for the Windows 8 launch.
Good point but I can't see why such minor optimizations would need a full year.

Erinyes

05-Jul-2011, 20:28

Good point but I can't see why such minor optimizations would need a full year.

Me neither. In fact its taken them more than a year for Brazos. They may not have anticipated the use of Brazos for tablets though so it looks like it was a late decision. Possibly they'll be able to offer a tablet variant of Krishna/Witchita quicker

dkanter

12-Jul-2011, 11:16

Me neither. In fact its taken them more than a year for Brazos. They may not have anticipated the use of Brazos for tablets though so it looks like it was a late decision. Possibly they'll be able to offer a tablet variant of Krishna/Witchita quicker

It depends if they are disabling and fusing off I/O or actually removing it. It's easy to believe that removing power hungry I/Os like PCI-E could cost you a quarter or two. It saves leakage (compared to merely fusing off), which is probably pretty important.

The problem isn't removing some I/Os, but the subsequent testing, validation, verification, etc. or adding new I/Os (which is even more complicated).

David