AMD Llano 3x more powerful than Intel SB?

[lieutenantfrost]

My apologies if Im posting this in the wrong place but I don't think my question fit in graphics forums

Just Read This:

http://arstechnica.com/business/news/2011/05/latest-gpu-market-numbers-spell-bad-news-for-nvidia.ars

Seems like it paints a pretty grim future for nvidia, (which was news to me)


Intel's upcoming Ivy Bridge platform will feature an on-die GPU that begins to threaten the mid-range of the discrete market the way that Sandy Bridge threatens the bottom end; and the on-die GPU with AMD's Llano is rumored to be some three times the performance of Intel's Sandy Bridge.

My question is this: How accurate is this information? AMD Llano is 3 times as powerful as Sandy bridge? When is it supposed to hit the market? When is Ivy Bridge Supposed to release? Is the expectation that it will be around the same time?

 

[mdd1963]

Llano is an 'ok' cpu, with adequate laptop-grade graphics built into it; the Llano's integrated graphics is indeed superior to Sandy's HD2000/3000 integrated graphics. To say simply that the 'CPU' is superior is a bit of a stretch, unless limited to graphics workloads, and relying solely on integrated graphics solutions.

 

[cburke82]

This wont affect Nvidi much at all if any. These type of CPU/GPU combos wont replace dedicated GPU's at least for a long wile. They might take a chunk out of the on board low cost Nvidia graphics but thats all. For any gaming type of use people will still need much more powerful dedicated GPU's.

 

[Timop]

Since llano is an APU, the GPU that's integrated does count as part of it, so three times isn't a stretch, Look up HD5570 vs HD3000.

Mobile llano is out already though, Desktop versions should hit market early July.

Ivy is going to take a while however, Q2 2011? There's even a slim chance AMD's next generation of APUs "Trinity" hitting the market before Ivy does.

 

[cburke82]

Since llano is an APU, the GPU that's integrated does count as part of it, so three times isn't a stretch, Look up HD5570 vs HD3000.

Mobile llano is out already though, Desktop versions should hit market early July.

Ivy is going to take a while however, Q2 2011? There's even a slim chance AMD's next generation of APUs "Trinity" hitting the market before Ivy does.

Q2 2011? are the Ivy chips out yet? Did I miss the release? Lol you mean 2012 right?

 

[Timop]

Q2 2011? are the Ivy chips out yet? Did I miss the release? Lol you mean 2012 right?

Ah, Lemme teleport back to 2011 from 2010.

 

[cburke82]

Ah, Lemme teleport back to 2011 from 2010.

lol the other day I was going on about the 2400k Intell CPU that was so cool LMAO.

On an another note I havent heard about the trinity APU's are they the next step after bulldozer?

 

[Timop]

lol the other day I was going on about the 2400k Intell CPU that was so cool LMAO.

On an another note I havent heard about the trinity APU's are they the next step after bulldozer?

lol, you gotta pair that with crossedfired Nvidia HD560Tis.

On Trinity, Trinity is to Ivy bridge like Llano is to Sandy bridge; while Bulldozer is the CPU architecture for trinity.

From what it looks like now, Trinity is one or two of "enhanced" Bulldozer modules (2-4 "cores") plus N.I. VLIW4 (HD6000) graphics on the same 32nm process.

 

[jimmysmitty]

Since llano is an APU, the GPU that's integrated does count as part of it, so three times isn't a stretch, Look up HD5570 vs HD3000.

Mobile llano is out already though, Desktop versions should hit market early July.

Ivy is going to take a while however, Q2 2011? There's even a slim chance AMD's next generation of APUs "Trinity" hitting the market before Ivy does.

Well with that idea then HD3000 is part of the SB unit even though you don't have to use it by using a P67 mobo. If you throw a discrete GPU in with Llano, only certain ones will hybrid CF from what I have seen so its not truly part of the CPU, its just on the die just like HD3000 is on the SB die.

IB will change a lot of things. We know the GPU will get a beefing with more EUs plus the 22nm can help add even more if they use it to stack DDR on the die for fast access.

Whos to say. In GPU, Llano is faster than SB but in CPU it is much slower.

 

[Timop]

Well with that idea then HD3000 is part of the SB unit even though you don't have to use it by using a P67 mobo. If you throw a discrete GPU in with Llano, only certain ones will hybrid CF from what I have seen so its not truly part of the CPU, its just on the die just like HD3000 is on the SB die.

IB will change a lot of things. We know the GPU will get a beefing with more EUs plus the 22nm can help add even more if they use it to stack DDR on the die for fast access.

Whos to say. In GPU, Llano is faster than SB but in CPU it is much slower.

I don't get your logic here. What does hybrid CF have to do with anything?
My point in the HD5570 is that the top tier on-die GPU of Llano is approximately that, 400SP@~600Mhz and DDR3 RAM, which is about 3X the performance of the HD3000, making the statement in the original article accurate.

What will IB change though? From what I know, its simply a 22nm shrink of SB with added EUs. The closest thing to "revolution" is the stacked RAM, ie on die sideport memory.

 

[jimmysmitty]

I don't get your logic here. What does hybrid CF have to do with anything?
My point in the HD5570 is that the top tier on-die GPU of Llano is approximately that, 400SP@~600Mhz and DDR3 RAM, which is about 3X the performance of the HD3000, making the statement in the original article accurate.

What will IB change though? From what I know, its simply a 22nm shrink of SB with added EUs. The closest thing to "revolution" is the stacked RAM, ie on die sideport memory.

A little more than that. 22nm allows for stacked transistors in general. That means they can stack RAM, cache, processors or anything. And the on die RAM would be much different. Since it would be directly connected to the CPU/GPU the latency would drop tremendously and the transfer rates would jump by a lot. A current Intel CPUs L1 cache, which is on die for example, has transfer rates of about 125000MB/s. Put 1GB of memory on the die for the GPU to access at faster data rates than system RAM, such as GDDR5, and the GPU will perform better. By how much is hard to say.

As for my point, I was basically saying that just because the GPU is better does not mean the CPU itself is better. If you plan CPU intensive applications, then Llano is the wrong way to go. If you want good onboard GPU performance, then Llano is the best way to go currently.

 

[Timop]

A little more than that. 22nm allows for stacked transistors in general. That means they can stack RAM, cache, processors or anything. And the on die RAM would be much different. Since it would be directly connected to the CPU/GPU the latency would drop tremendously and the transfer rates would jump by a lot. A current Intel CPUs L1 cache, which is on die for example, has transfer rates of about 125000MB/s. Put 1GB of memory on the die for the GPU to access at faster data rates than system RAM, such as GDDR5, and the GPU will perform better. By how much is hard to say.

As for my point, I was basically saying that just because the GPU is better does not mean the CPU itself is better. If you plan CPU intensive applications, then Llano is the wrong way to go. If you want good onboard GPU performance, then Llano is the best way to go currently.

Oh, lol. I was merely proving that like the quotes article said, Llano does have ~3x the graphics performance of SB.


Since you mentioned it, is stacking is allowed with 22nm in general? Haven't read much on it, but I thought it was with the Tri-gate process intel is using WITH 22nm?

 

[jimmysmitty]

^yea its the Tri-Gate tech Intel developed. I just say 22nm since thats the first process it will be used on.

As to if AMD or anyone else will us tri-gate at 22nm is a question I want the anwser to.

 

[megamanx00]

The GPU is certainly faster, but the CPU not so much. The picture will probably look better for Trinity which will have a similar GPU but a much better Bulldozer CPU core and perhaps more integration between CPU and GPU. It will eat away a little from nVidia's low end market share, but as long as Intel is still Chipzilla nVidia will be supplying PC makers with low end GPUs as a step up from Sandy Bridge integrated graphics, which kinda takes away the quick sync advantage ^_^

 

[Guest]

22nm tri gate is going to be Intel only. TSMC should have 28nm out before too long, it remains to be seen how good that will be in comparison and how bad of a hit AMD is going to take for being behind.

As for Nvidia going down the tubes because of good integrated graphics. Yeah they might take a hit. APU's might kill discrete graphics in all but gaming laptops. Nvidia might be pushed out of the low end and midrange in mobile and low end in desktop. But that does not mean the company will go under. They are making CPUs now, their Tegra 2 chip seems quite popular with the Android phones and tablets, and they have a solid business there, and they will continue to have a solid business with high end discrete graphics as well.

And Windows 8 will be on Arm, so Nvidia could well become a huge cpu manufacturer for tablets and netbooks and low end laptops.

 

[Timop]

22nm tri gate is going to be Intel only. TSMC should have 28nm out before too long, it remains to be seen how good that will be in comparison and how bad of a hit AMD is going to take for being behind.

As for Nvidia going down the tubes because of good integrated graphics. Yeah they might take a hit. APU's might kill discrete graphics in all but gaming laptops. Nvidia might be pushed out of the low end and midrange in mobile and low end in desktop. But that does not mean the company will go under. They are making CPUs now, their Tegra 2 chip seems quite popular with the Android phones and tablets, and they have a solid business there, and they will continue to have a solid business with high end discrete graphics as well.

And Windows 8 will be on Arm, so Nvidia could well become a huge cpu manufacturer for tablets and netbooks and low end laptops.

AMD is using Glofo for 28nm IIRC, which claims to pack density higher then Intel's 22nm (?).

And with Nvidia and ARM, how many OEMs are doing ARM? Qualcomm, TI, ST-e, Samsung and that's just the big ones.

 

[mdd1963]

Summary: woe to those that read "Llano is 3x more powerful than SandyBridge", and then actually buy a MB and CPU/APU to build a gaming system around it, only to discover that most games are pathetically slow at even 1024x768. If such motherboards are available for it with a PCI-e(x16) slot to allow a nice graphics card, the system will still be slower than SB (and X4 based rigs) when both are given an actual GPU.

 

[Timop]

Summary: woe to those that read "Llano is 3x more powerful than SandyBridge", and then actually buy a MB and CPU/APU to build a gaming system around it, only to discover that most games are pathetically slow at even 1024x768. If such motherboards are available for it with a PCI-e(x16) slot to allow a nice graphics card, the system will still be slower than SB (and X4 based rigs) when both are given an actual GPU.

First of all "Llano is 3x more powerful than Sandy Bridge" is incorrect paraphrasing, "the on-die GPU with AMD's Llano is rumored to be some three times the performance of Intel's Sandy Bridge" is the reality.

Second, a quad core llano would max out most games at 1024*769 and 1366*768.

Thirdly, llano performs almost the same compared to a Phenom II X4 clock for clock, so your last point is partially incorrect.

 

[jimmysmitty]

AMD is using Glofo for 28nm IIRC, which claims to pack density higher then Intel's 22nm (?).

And with Nvidia and ARM, how many OEMs are doing ARM? Qualcomm, TI, ST-e, Samsung and that's just the big ones.

GloFlo talks a big game but 28nm packing a higher transistor count in the same space as 22nm? Even higher than 3D Tri-gate? I doubt this very much. You have to look at the technology of the process itself. If a box is 5 inches cubed can you pack more boxes into a foot cubed than boxes that are three inches cubed? No.

Now 28nm having a higher transistor density than 32nm, absolutley. But there is no way that 28nm will have a higher transistor density in the same space as 22nm. Its just not logical. Add in tri-gate and that alone means a CPU die can hold more in the same die size due to the stacking ability.

And while Intel will be the first to have 3D Tri Gate tech, it doesn/t mean they will be the only one. They were the first to HK/MG but AMD and the others are moving towards it for 32nm.

First of all "Llano is 3x more powerful than Sandy Bridge" is incorrect paraphrasing, "the on-die GPU with AMD's Llano is rumored to be some three times the performance of Intel's Sandy Bridge" is the reality.

Second, a quad core llano would max out most games at 1024*769 and 1366*768.

Thirdly, llano performs almost the same compared to a Phenom II X4 clock for clock, so your last point is partially incorrect.

The only time I saw the GPU in Llano actualy outperform HD3000 by 3x (300%) was if they were also using a discrete HD6630 GPU and doing a hybrid CF setup. Alone, the A8-3500 APU wasn't as much faster as being claimed than HD3000. It was faster, but not nearly as much as the claims were. Same with power usage. SB and Llano were about the same for power usage.

As for the CPU performance of Llano, from what I have read its actually about the same as a AthlonII X4 which is still weaker than Phenom II due to a lack of L3 cache.

 

[Timop]

GloFlo talks a big game but 28nm packing a higher transistor count in the same space as 22nm? Even higher than 3D Tri-gate? I doubt this very much. You have to look at the technology of the process itself. If a box is 5 inches cubed can you pack more boxes into a foot cubed than boxes that are three inches cubed? No.

Now 28nm having a higher transistor density than 32nm, absolutley. But there is no way that 28nm will have a higher transistor density in the same space as 22nm. Its just not logical. Add in tri-gate and that alone means a CPU die can hold more in the same die size due to the stacking ability.

And while Intel will be the first to have 3D Tri Gate tech, it doesn/t mean they will be the only one. They were the first to HK/MG but AMD and the others are moving towards it for 32nm.



The only time I saw the GPU in Llano actualy outperform HD3000 by 3x (300%) was if they were also using a discrete HD6630 GPU and doing a hybrid CF setup. Alone, the A8-3500 APU wasn't as much faster as being claimed than HD3000. It was faster, but not nearly as much as the claims were. Same with power usage. SB and Llano were about the same for power usage.

As for the CPU performance of Llano, from what I have read its actually about the same as a AthlonII X4 which is still weaker than Phenom II due to a lack of L3 cache.

It is skeptical, but I am willing to believe that it is somewhat true, there are variances in density even with apparent "identical" processes. (ie. how llano packs 1.14b in 228mm2, while SB only packs 995m in 216mm2, same goes with AMD and Intel's 45nm, how the Phenom packs a little more per mm2 than Nehalem.) Though everything is pretty much out of AMD's hands (sadly), as what Little stake they still have in Glofo means nothing more than a shareholder.

On the GPU, Desktop Llano A8 @2.9Ghz consistently doubles frames of the 2500K in anand's desktop "preview", and AMD claims a 6% IPC increase over Atlon II due to the extra L2 and optimization, it should be safe to say the GPU is ~3X powerful while the CPU is pretty much on par with Phenom II (which wah ~10% better than Athlon II X4) isn't it?

IIRC, the Compal samples AMD's sent out with the 3500M all had broken turbos, which pretty much invalidates their performance as Turbo is major for Llano. (Why would AMD do this I have no idea.)

 

[mdd1963]

Thirdly, llano performs almost the same compared to a Phenom II X4 clock for clock, so your last point is partially incorrect.

"Almost the same"....meaning....perhaps not 'faster', yet, not quite 'identical' performance, but....it is ....'slower'? (to be expected, given cache differences)

Benchmarks I saw had it more on par with an Athlon II X4/635 or 645, vice matching any of the Phenoms at/near equivalent clockspeeds...; naturally, Llano might be quicker than the slowest Phenom II X4 (910 at 2.6 GHz), if/ when the latter is at a clock deficit...

http://www.anandtech.com/show/4448/amd-llano-desktop-performance-preview/2

I think it is an impressive laptop solution, as designed, and might do just fine as a low cost dekptop for my kids' Facebook, iTunes, websurfing, etc...; but I can't imagine anyone choosing it to build a gaming system.

 

[jaguarskx]

A Llano based laptop will at least raise the minimum bar for an inexpensive gaming laptop. True, while gaming performance is not going to be breath taking, it will be a good improvement on the low end compared to previous integrated solutions.

I thought that Intel's HD 3000 graphics core was a pretty significant performance boost from their previous IGP, the 4500mHD. Although, it would have been nice if all the Sandy Bridge CPUs has the HD 3000.

 

[fazers_on_stun]

I don't get your logic here. What does hybrid CF have to do with anything?
My point in the HD5570 is that the top tier on-die GPU of Llano is approximately that, 400SP@~600Mhz and DDR3 RAM, which is about 3X the performance of the HD3000, making the statement in the original article accurate.

That article was written before anybody had actual Llano systems to bench however. When the rubber hit the road, Anandtech says Llano only 58% faster in GPU than SB, which is a far smaller amount than 300% faster:

Llano vs. Sandy Bridge: Finally, Acceptable Processor Graphics

On average the A8-3850 is 58% faster than the Core i5 2500K. If we look at peak performance in games like Modern Warfare 2, Llano delivers over twice the frame rate of Sandy Bridge. This is what processor graphics should look like. While I believe Sandy Bridge was a good start for integrated GPU performance, Llano is my ideal for 2011. Update: We've added results from the latest 2372 driver for the 2500K. Most performance results remain unchanged however a few problematic areas for Sandy Bridge have been addressed as a result. Llano still maintains a significant performance in the majority of cases.

Games that are more CPU bound however do show Llano's weakness. Both Dragon Age and Starcraft II have Sandy Bridge either outperforming or coming very close to Llano in frame rates. Those are most definitely the exception rather than the rule however; for the most part AMD is able to deliver entry-level discrete GPU performance with Llano.

What will IB change though? From what I know, its simply a 22nm shrink of SB with added EUs. The closest thing to "revolution" is the stacked RAM, ie on die sideport memory.

IIRC S/A thinks it'll do "3X" increase in GPU over SB .. Actually no numbers yet of course but the on-die dedicated (not shared with the CPU) DDR2 might actually double or triple performance, given the huge & low-latency memory pipe.. I'm thinking L2 or L3 cache speeds..

 

[fazers_on_stun]

First of all "Llano is 3x more powerful than Sandy Bridge" is incorrect paraphrasing, "the on-die GPU with AMD's Llano is rumored to be some three times the performance of Intel's Sandy Bridge" is the reality.

Second, a quad core llano would max out most games at 1024*769 and 1366*768.

Thirdly, llano performs almost the same compared to a Phenom II X4 clock for clock, so your last point is partially incorrect.

Make that an Athlon II X4 (no L3 cache) and I'd agree with youl From the same Anandtech Llano desktop preview:

CPU Performance: Pretty Much an Athlon II X4

As we found in our look at mobile Llano, the A8 isn't impressive as a general purpose x86 microprocessor. In general the chip is somewhat faster than the Athlon II X4 635 and I'd say it performs more like a 645 based on the numbers I've seen here. Again, nothing to be impressed by but if you're building a value gaming PC it may not matter.

Note that heavily-threaded applications actually favor the A8-3850 to the Core i3 2100 (its most likely target based on pricing rumors) thanks to its four cores. They may not be as efficient as the i3's cores, but you sure do have more of them. We have been discussing this tradeoff with AMD for quite a bit over the past couple of years. You lose out on single-threaded performance but you do gain better performance in heavily-threaded workloads. I had assumed that Turbo Core would partially solve this with Llano but 2.9GHz is going to be the fastest SKU AMD offers and it doesn't ship with any turbo enabled.

 

[fazers_on_stun]

A Llano based laptop will at least raise the minimum bar for an inexpensive gaming laptop. True, while gaming performance is not going to be breath taking, it will be a good improvement on the low end compared to previous integrated solutions.

I thought that Intel's HD 3000 graphics core was a pretty significant performance boost from their previous IGP, the 4500mHD. Although, it would have been nice if all the Sandy Bridge CPUs has the HD 3000.

I've heard IB will have 16 or 24 (double) the EU's of HD3000, with some optimizations possible as well, so hopefully the GPU performance will bump up nicely even if the stacked on-die DDR2 rumor from SemiAccurate turns out to be the 50% wrong info .. I'd guess it might be, seeing as how they got the tri-gate rumor correct - now they're overdue for being wrong if they wanna keep the "semi" in SemiAccurate ..

As for the GloFlo transistor density claims for 28nm, all I ever heard was their claiming 'gate first' HKMG at 28nm would be 'up to 20% more dense than the competition's (read TSMC) gate-last HKMG 28nm process'. Stands to reason since extra processing steps for gate-last (depositing a place-holder gate for self-alignment of source & drain, annealing the substrate at 1000+ degrees Centigrade, then removing the temp gate and depositing the HKMG final gate) all tend to require some extra slack in how closely you can place the features in each process layer.

But, it seems GloFlo along with the other IBM consortium members will be switching to gate last HKMG at 22nm, simply because baking the hafnium compound at 1000+ degrees requires too much compromise to survive the high temp, reducing the effectiveness compared to gate last where Intel (and TSMC) can optimize the gate composition for performance, not surviving high temps. With the even thinner gate insulation layer needed at 22nm vs. 32nm or 28nm, having the right composition becomes even more important apparently.