Bulldozers Weak/Strong points?

[LiuKangBakinPie]

Intel gave a major heads-up when it came to core2 performance. They even gave anand and other reviews sites a free x6800 extreme cpu. That was months before Intel finally released core2.

Anyways, I keep having a nagging feeling bd will be delayed one way or another. I just have a feeling that llano will be released before bd even though bd is scheduled to release first. It would still be nice to get some numbers on bd. Even a 1MB SuperPI bench would be something.

Hopefully things will go Perfectly for AMD this go round. :thumbsup:

Oemg Amd have always been awful with superpi. The idea is to promote the chip not kill it with benchmarks

 

[Joseph F]

You should read my blog on turbo core.

The net issue is that when you bin a processor you have to take worst case scenario. If at max power it reaches 3GHz for a database, 2.8GHz for java and 2.4GHz for HPC you have to fuse it as a 2.4GHz, even though it could get more.

You can't market it as a 3GHz part and say "some workloads might be lower", there would be anarchy and violence in the streets if that started happening (i.e. lawsuits).

Turbo core let's you capture the power headroom and turn it into clock speed.

Ah, so it basically throttles the processor up to near its TDP. Thanks for clearing things up for me. The blog post was very informative. I also checked out your "Tax Time" post. Too bad I'm not eligible for the contest.

 

[JFAMD]

you should really read the bulldozer "20 questions" blogs, there are 4 of them. Most importantly read all of the comments and answers as well. That should give you tons of data.

 

[formulav8]

you should really read the bulldozer "20 questions" blogs, there are 4 of them. Most importantly read all of the comments and answers as well. That should give you tons of data.

Although you may not be able to say one way or another, have you seen bd in action? Or can you say that AMD has working, shippable silicon? :sneaky: I understand if you don't want to reply.

 

[formulav8]

Oemg Amd have always been awful with superpi. The idea is to promote the chip not kill it with benchmarks

Before Core 2, AMD ruled the SuperPI roost. The P4 truly did a slower job compared to the A64.

 

[Topweasel]

Intel gave a major heads-up when it came to core2 performance. They even gave anand and other reviews sites a free x6800 extreme cpu. That was months before Intel finally released core2.

Anyways, I keep having a nagging feeling bd will be delayed one way or another. I just have a feeling that llano will be released before bd even though bd is scheduled to release first. It would still be nice to get some numbers on bd. Even a 1MB SuperPI bench would be something.

Hopefully things will go Perfectly for AMD this go round. :thumbsup:

I will say it again and probably more. Intel has the luxury of being the only manufacturer able to supply enough CPU's to OEM. They have the luxury, specially during the Prescott days where there was a measurable shift to AMD, of being able to lose even a few thousand sales. The impact on the bottom-line percentage wise is so small that any trend management they do might never have noticed it. Intel has so many fabs that they make those CPU's for several years afterward. AMD on the other hand has always quickly moved production, every wafer is precious. AMD can't afford lower sales, buying back or discounting OEM's for unsold stock.

Intel can show off specially when performance to a competitor is as close as AMD has been at times, because that Misquito bite in lost sales to them is like a Bear mauling them to Intel. To Intel, AMD's % of sales isn't the end of the world to them, but every gained sale to Intel means little, every lost sale to AMD has impact in R&D and company outlooks for the next year. Intel has always played the game of not out selling AMD, they already do that, but going out of their way to make sure AMD can't sell anything. Any time you see those benchmarks months early, remember that to Intel any loss in sales by doing so is worth it as long as AMD also loses sales.

For AMD if good or bad its the exact opposite and they have learned from two horrible examples of prerelease shows to not do it again. Its really bad if BD is good, because people will actively wait it out. Whereas most people looking into a PhII setup will know where they sit and probably not change and get a 2500k just because BD isn't great.

 

[IntelUser2000]

Topweasel: It's not that simple. Both of the manufacturers learned from their previous mistakes.

Intel: They released a lot of info about Prescott and the first Pentium 4 way before launch. It was indeed hyped quite a lot. For Prescott, some were predicting 40% performance increase over Northwood.

AMD : Before the launch of Athlon 64, fantastic SpecCPU numbers, and the confidence was shown numerous times before release. They were right then. They were at the time, behind Intel because Pentium 4 "C" chips beat the AthlonXP line. Also, I don't think I need to remind you of Barcelona.

Both of the companies are relatively quiet now.

 

[drizek]

Amd got pretty quiet about Barcelona leading up to the release.

 

[exar333]

Topweasel: It's not that simple. Both of the manufacturers learned from their previous mistakes.

Intel: They released a lot of info about Prescott and the first Pentium 4 way before launch. It was indeed hyped quite a lot. For Prescott, some were predicting 40% performance increase over Northwood.

AMD : Before the launch of Athlon 64, fantastic SpecCPU numbers, and the confidence was shown numerous times before release. They were right then. They were at the time, behind Intel because Pentium 4 "C" chips beat the AthlonXP line. Also, I don't think I need to remind you of Barcelona.

Both of the companies are relatively quiet now.

Don't forget, a lot of the hype comes from the marketing side. We (enthusiasts) are smart enough to understand real claims vs. 'benchmark data'. Seeing some synthetic numbers doesn't mean a whole lot until you get get some actual data to back that up. You can have synthetic numbers up the a$$, but who cares if it doesn't translate to real performance gains?

 

[Topweasel]

Topweasel: It's not that simple. Both of the manufacturers learned from their previous mistakes.

Intel: They released a lot of info about Prescott and the first Pentium 4 way before launch. It was indeed hyped quite a lot. For Prescott, some were predicting 40% performance increase over Northwood.

AMD : Before the launch of Athlon 64, fantastic SpecCPU numbers, and the confidence was shown numerous times before release. They were right then. They were at the time, behind Intel because Pentium 4 "C" chips beat the AthlonXP line. Also, I don't think I need to remind you of Barcelona.

Both of the companies are relatively quiet now.

Trust me you see it both ways. Just saying Intel has the luxury, they keep the older products in production, they sell enough inventory to make the impact minimal, and they profit for years in AMD lost sales. AMD doesn't have the luxury of cannibalizing and losing sales for a product people can't buy.

 

[nonameo]

So, someone riddle me this. If SOI is providing significant benefit, how come intel isn't using it?

 

[Topweasel]

So, someone riddle me this. If SOI is providing significant benefit, how come intel isn't using it?

Because it might affect their process and/or another technology they are using. There is more then one way to skin a cat. AMD and IBM had to work several year and tons of money coming up with it. Intel can almost always stay a process ahead and has had no problems keeping up, in the end its probably more then Intel wants to spend and increases the opportunity that they run into an extended delay on whatever CPU they are releasing on a new process if they went to using it.

 

[Idontcare]

So, someone riddle me this. If SOI is providing significant benefit, how come intel isn't using it?

It's a simple matter of accounting.

Node development is a fixed cost, it doesn't matter whether you are a $50B/yr company that is going to use the node to process 500,000wfrs/month or if you are a $1B/yr company that is going to use the node to process 1,000 wfrs/month.

Either way the node is going to cost you around $4B to develop.

SOI lowers the development cost, makes it "easier" to hit the leakage spec, for the node in exchange for making it more costly to produce on a per-wafer basis.

If you are that $50B/yr company you are not interested in making your 500,000 wfrs/month cost an extra $100 per wafer.

You'd much rather invest an extra $50m-$100m up-front during node development to develop a comparable node that is not dependent on SOI. Costs more to develop, but per-wafer production costs will be lower and after you make a couple million wafers with it you will have recovered your extra R&D costs.

If you are that $5B/yr company you are not looking at making 500,000 wfrs/month. You are going to be more interested in saving that $50m-$100m upfront R&D money and just take the $100 hit on your paltry low-volume 5,000 wfrs/month.

Even though your wafers cost more to produce, the added cost will never amount to $100m over the life of the node.

(PS - its the same accounting principles that come into play when IDM's decide to go immersion litho vs. dual-pattern litho...its a trade-off that makes economic sense depending on your production volumes versus R&D budget for developing the node)

 

[nonameo]

Because it might affect their process and/or another technology they are using. There is more then one way to skin a cat. AMD and IBM had to work several year and tons of money coming up with it. Intel can almost always stay a process ahead and has had no problems keeping up, in the end its probably more then Intel wants to spend and increases the opportunity that they run into an extended delay on whatever CPU they are releasing on a new process if they went to using it.

Wow, so really amd just can't win either way... at least they have options that give them a little bit of flexibility.

 

[IntelUser2000]

Don't forget, a lot of the hype comes from the marketing side. We (enthusiasts) are smart enough to understand real claims vs. 'benchmark data'.

Some of us do, some of us cling onto the numbers.

 

[formulav8]

Intel is using SOI in 22nm I think? Since AMD and IBM paved the way (Lots of r & d), Intel just needs to license the tech and can benefit from IBM and AMD's work I guess...

 

[Idontcare]

Intel is using SOI in 22nm I think? Since AMD and IBM paved the way (Lots of r & d), Intel just needs to license the tech and can benefit from IBM and AMD's work I guess...

Not all things labeled "SOI" are the same.

Intel has long stated that fully-depleted SOI provides technical benefits over bulk-Si but the production costs were not justifiable and the benefits changed depending on the node in question.

AMD/IBM use "partially depleted" SOI.

http://www.electrochem.org/dl/ma/203/pdfs/0835.pdf

http://semimd.com/wp-content/uploads/2011/02/SOIconsortium_FDSOI_QA.pdf

In addition, the vast majority of production-worthy implementations of SOI have involved planar-CMOS.

The tentative arguments made for FD-SOI are being put forth in the pursuit of non-planar CMOS (finfets and so on) where there is little IP to license.

When it comes to 3D xtors and FD-SOI, IBM and GloFo stand to be behind the curve inasmuch as was the case for HKMG.

 

[podspi]

When it comes to 3D xtors and FD-SOI, IBM and GloFo stand to be behind the curve inasmuch as was the case for HKMG.

Really? I was under the impression GloFo was working on finfet as well?! It just isn't showing up @ 22nm...


As someone not terribly versed in foundry technology, what is the significance of not switching over? I tried reading the wikipedia page but it went waay over my head

 

[Idontcare]

Really? I was under the impression GloFo was working on finfet as well?! It just isn't showing up @ 22nm...


As someone not terribly versed in foundry technology, what is the significance of not switching over? I tried reading the wikipedia page but it went waay over my head

Of course they are. Just as they were working on HKMG for 45nm...

IBM Advancement to Spawn New Generation of Chips

IBM, AMD Announces 45nm Chip High-K Gate Process

And it was true. They really were. It just turned out that their efforts were totally under-resourced (not enough R&D $) and as "out in left-field" from the rest of the industry as their efforts to put SiLK into production as BEOL low-k.

(result of IBM's efforts to say "all was well with SiLK" can be read here)

Gate-first HKMG will be a one-node solution as well. They went out to left-field on that one. Completely ditched and position reversed for 22nm. (it is gate last)

 

[AtenRa]

Not all things labeled "SOI" are the same.

Intel has long stated that fully-depleted SOI provides technical benefits over bulk-Si but the production costs were not justifiable and the benefits changed depending on the node in question.

AMD/IBM use "partially depleted" SOI.

http://www.electrochem.org/dl/ma/203/pdfs/0835.pdf

http://semimd.com/wp-content/uploads/2011/02/SOIconsortium_FDSOI_QA.pdf

In addition, the vast majority of production-worthy implementations of SOI have involved planar-CMOS.

The tentative arguments made for FD-SOI are being put forth in the pursuit of non-planar CMOS (finfets and so on) where there is little IP to license.

When it comes to 3D xtors and FD-SOI, IBM and GloFo stand to be behind the curve inasmuch as was the case for HKMG.

Nice reading thx for the links

Question,

Does the FD-SOI (Ultra thin layer) effects the Gate insulation (in parallel) with the High-K dielectric (in HKMG process) or it only effects the Source and Drain leakage of the transistor??

http://semimd.com/wp-content/uploads/2011/02/SOIconsortium_FDSOI_QA.pdf

From the drawings in the above link i would say that FD-SOI puts another layer of insulation between the Gate and the Si substrate.

If that is true then with FD-SOI we could have a thinner High-K dielectric ??

 

[podspi]

At least GloFlo has much deeper pockets now than when it was the manufacturing division of AMD...

 

[formulav8]

At least GloFlo has much deeper pockets now than when it was the manufacturing division of AMD...

True but doesn't the process transitions only go so fast? I mean money can only take you so far when it comes to process transitions?

 

[Topweasel]

True but doesn't the process transitions only go so fast? I mean money can only take you so far when it comes to process transitions?

The problem with AMD in the past and a little bit with GloFo is that the limit on objects made there. AMD had to rely on sales of their current products to afford the manufacturing plants needed more for the production of future products. It was a no win, they were poor and if they built plants based on growing production, and it doesn't grow, they die. Not only that but AMD couldn't afford to build more then 1 at a time anyways. So AMD generally built their Foundries to handle 2 processes and they could start for example on 180nm and as they transitioned to 130, start work on a new plant for 90nm and 65nm rinse and repeat.

Also they can make bigger jumps but its generally not worth it, the smaller the jumper the smaller the development time, the lower the chances of issues, shorter time to production, and gives you another target to hit.

The separation with GloFo means that A.) Glofo in the end only has to worry about their bottom line. Which means as long as they are meeting demand with AMD, that excess can be sold off to other projects with other companies (they might be limited here like no Intel). B.)They can also keep fabs up an running and profitable longer on older process's by selling production time to companies with IC's that don't need to be on the top process. C.) They are backed by a very rich entity that can afford to pay for Glofo being un-profitable as they build up new tech fabs at a much quicker rate, not just to catch up to TSMC but even Intel. D.) To get notoriety and be in demand it might even be in GloFo's best interest to upset the apple cart and make a jump or two, keep in mind they want to grow in capacity as well as technology. It might make sense for them to start building a fab now for a tech that isn't forecasted till 2015 to have it ready at the end of 2013 or start of 2014.

Intel with their 17 odd fabs or so can just play hot pocket as they lower production one series and increase on new ones.

 

[Idontcare]

True but doesn't the process transitions only go so fast? I mean money can only take you so far when it comes to process transitions?

It is true that there are practical limits on just how fast node development can occur, but having been there myself I can tell you that the node development timeline is purely money-limited down to the point of say 2-3 months.

To do a node in less than 2-3 months would require some extensive reworking of the existing societal infrastructure (education system, national priorities, etc).

But the tools and methods exist today for corporate-driven development to push the cycle time down to a mere few months if they really really really wanted to throw absurd amounts of cash at it. (by absurd I mean trillions, not practical but not impossible)

But the gap you see today between GloFo and Intel is entirely a matter of cash.

If someone gave GloFo's R&D team $8B-10B today and said "beat Intel to sub-20nm and with higher parametric performance, regardless of profitability" they could do it. It would be commercial foolishness, but not beyond the capability of the existing tools and infrastructure.

 

[formulav8]

Thanks greatly for both replys! Much clearer picture of it all