AMD Barcelona Thoughts/Questions

[CTho9305]

Originally posted by: Viditor

Originally posted by: CTho9305

Originally posted by: Viditor
CTho9305...

Beautifully written!! I can't thank you enough for all of that effort, mate! :thumbsup:

Do you mind if I reprint that at another website (it's a pay site)? There are quite a few people there who understand investing but are having a difficult time with the basics on the chips. Your post is an excellent explanation.

Many humble thanks...

I do mind. Feel free to link to it here, but I think it's wrong to increase the value of somebody else's site without compensation. If you don't want to link to this forum from another, I can put it on a page on my own site or my wikipedia user page and you can link to that.
edit: To claify, I realize that posting it here may increase the quality of AT forums and make Anand more money, but I think it's ok because everybody can read it for free.

Fair enough, and I'll comply...again, thanks for the effort! (BTW, Wikipedia would be a great idea...)

Copied to http://en.wikipedia.org/wiki/U...Out_of_Order_Execution (As a result of putting there, it becomes GFDL, which is a license I personally think is pretty crappy... oh well)

 

[bryanW1995]

nice. so an amd fanboy might not be able to say "my cpu is faster than yours", but he could at least try to justify a statment like: "my uop is wider than yours".

 

[Viditor]

Originally posted by: bryanW1995
nice. so an amd fanboy might not be able to say "my cpu is faster than yours", but he could at least try to justify a statment like: "my uop is wider than yours".

Actually, it seems to go to the point that the 4-issue feature of the C2D core isn't really that much of a factor in it's speed...it seems to me that Intel's Memory Disambiguation plays a much bigger role. It also means that whether Shanghai adopts a 4 issue core or not isn't really that important.
At least that's my impression...

 

[dmens]

Originally posted by: Viditor
Actually, it seems to go to the point that the 4-issue feature of the C2D core isn't really that much of a factor in it's speed...it seems to me that Intel's Memory Disambiguation plays a much bigger role.

i disagree, but can't give any details unfortunately.

 

[zsdersw]

Originally posted by: munky
That's simply not true. Take a look at this article (several pages comparing the difference cache size makes). Games in particular show a very significant gain from extra cache, as well as some office applications. In particular, the 1MB cache cpu suffers the biggest drop in performance going from a 2MB mode, as opposed to going from 4MB to 2MB.

Games and some office applications definitely do not rise to the level of "most applications".

 

[OneEng]

coldpower27,

I will say that it's primarily due to the FB-DIMM's that the Stoakley platform post such high power consumption.

That is a big part of it, but the IMC also makes a difference. The "other" stuff catagory on the Intel system includes their North Bridge which is likely eating some power as well.

That being said, I think that Intel would re-gain the performance/watt crown if they abandoned FBDIMM's.

I know that many believe that K10 is really not that big a deal for AMD in the light of Penryn. For JAVA application servers, database servers, and HPC applications, a 2.5Ghz K10 is a very impressive machine. These are its target markets, so I think AMD has done well with the IPC in these areas. The clock speed is another issue. It is horrendus. The 3.0Ghz demonstration shows that AMD may have some hope by Q1 2008 though.

The desktop arguement isn't as solid.

AMD's system advantages (IMC, P2P protocol, etc) don't do nearly as much for it on the desktop. That being said, a 3.0Ghz Phenom should provide a compelling desktop solution. With lower latency memory combined with significantly reduced L3 latency (due to clock speeds), K10 could be quite surprising in the desktop.

I am still not sure what is going on with the 3d rendering scores from K10. Core for core, K10 is only marginally faster than K8 in this department.

 

[Idontcare]

Originally posted by: coldpower27

Originally posted by: OneEng
It is also interesting to note that K10's power efficiency bests Intel's Clovertown and even Penryn in most cases. Intel is gaining ground here and if they ever abandon FBDIMM's, things may change quickly. The K10 is still a very cool operator showing that it can hang with an Intel 45nm part even at 65nm.

I will say that it's primarily due to the FB-DIMM's that the Stoakley platform post such high power consumption.

http://techreport.com/articles.x/13224/9

As seen there in terms of power consumed at high load the E5472 system already beats the 2360 SE system, and that is with the handicap of using FB-DIMM modules which consume considerable amounts of power.

The primary reason the AMD parts can hang in there is due to Intel's use of FB-DIMM's and not because the processors are particularly more efficient in terms of energy usage. The difference between the 2350 and the E5472 is only 43W, and there are 8 DIMM's there so even with a conservative figure like 5W per DIMM we are looking at 3 W difference which would be within a margin of error.

K10 is very competitive in terms of power usage due primarily because of the platform and not because of the processors themselves necessarily.

Any thoughts on how much performance would be lost were Intel to abandon FB-DIMM in favor of reducing system wattage?

Surely there is some performance related reason why Intel couples their server platform to FB-DIMM...if not then why would they stick with higher cost, higher power, and lower performance memory parts?

I can only conclude it is a trade-off for real-world performance, which will be lost should they back away from FB-DIMM. I also note they have zero road-map plans to back away from FB-DIMM, so Intel appears to intend to continue to rely on this performance boost for future server (and skulltrail'esque) platforms.

My point being that yes surely you can argue that migrating from FB-DIMM to standard DDR2 or DDR3 parts as AMD has done would drastically decrease Intel's server platform power consumption and ought to tip the performance/watt balance to their favor...but that argument assumes the performance decrease would not be so punitive as to negate the benefits of decreased power consumption.

Both the numerator and denominator will scale-back, and the ratio is no-doubt something Intel already knows and also no-doubt was a deciding factor in their decision to go with FB-DIMM.

I very much doubt Intel overlooked where their system watts are going in their servers when they set out to hit their performance/watt targets.

 

[Keysplayr]

FB-Dimms are a repeat of RAMBUS. It had more drawbacks than benefits, and was dismissed. The same will happen with FB-Dimms.
DDR outperformed RAMBUS, and new variations of DDR will continue to do so against FB-Dimms. This will be the second time (that I know of) that Intel tried to create a new standard in memory choice but only to have it fail. There are already rumors (or more than rumors) of FB-Dimms being abandoned. My initial thought was, why do they even bother to bring these hardwares to fruition if they KNOW ahead of time it won't make it with it's limitations? They had to have know how much power these things would draw. It is a massive increase in power consumption, when the entire industry is headed the other way.

 

[jones377]

FB-DIMM is intended to solve the memory capacity problem. Intel isn't the only company behind it. For example, Sun is using it in their Niagara 2 processor. Why? It's the only way to deliver the kind of memory capacity needed for servers into a single processor socket with an acceptable number of pins. Each FB-DIMM channel can support up to 8 DIMMs. How many do DDR2/3 support?

AMD doesn't suffer nearly as much because each processor has it's own memory controller so capacity can be expanded as you add more sockets, but they will suffer too once they get to DDR3. They have already announced their own solution by splitting up a memory channel into two using an external chip to allow more modules per memory channel.

 

[Kuzi]

Some very interesting posts here, we learn something new everyday, thanks for all the imformation guys

I still have the question about the IMC in K10, if anyone knows about it for sure. Why does AMD run it slower than the CPU speed? And is 1.8GHz the highest speed when using split planes, or will the highest speed be 200MHz lower than the CPU speed? (ex. 3GHz K10, IMC will be at 2.8GHz)

It's an important issue as higher speeds for the IMC will mean lower latency and higher performance, something AMD will really need against Penryn especially on the Desktop.

 

[Idontcare]

Originally posted by: jones377
FB-DIMM is intended to solve the memory capacity problem. Intel isn't the only company behind it. For example, Sun is using it in their Niagara 2 processor. Why? It's the only way to deliver the kind of memory capacity needed for servers into a single processor socket with an acceptable number of pins. Each FB-DIMM channel can support up to 8 DIMMs. How many do DDR2/3 support?

AMD doesn't suffer nearly as much because each processor has it's own memory controller so capacity can be expanded as you add more sockets, but they will suffer too once they get to DDR3. They have already announced their own solution by splitting up a memory channel into two using an external chip to allow more modules per memory channel.

That's my point too, why does Intel and SUN) want their customer to be able to load up a server with so much RAM that suddenly FB-DIMMs become necessary? Performance.

Take away FB-DIMM and you will no-doubt be taking away performance.

Not the kind of performance that benchmarketing fails to capture, but rather the kind that apparantely exists in real-world server applications that customers are using.

I don't care for the power consumption of FB-DIMMs but I sure hope Intel's (and SUN's) platform engineers do NOT optimize their hardware for the sake of enabling Anandtech to generate some pretty but synthetic performance/watt graphs.

I also wonder about Itanium...where does it fit into the performance/watt situation?

 

[Viditor]

Originally posted by: Idontcare

Originally posted by: jones377
FB-DIMM is intended to solve the memory capacity problem. Intel isn't the only company behind it. For example, Sun is using it in their Niagara 2 processor. Why? It's the only way to deliver the kind of memory capacity needed for servers into a single processor socket with an acceptable number of pins. Each FB-DIMM channel can support up to 8 DIMMs. How many do DDR2/3 support?

AMD doesn't suffer nearly as much because each processor has it's own memory controller so capacity can be expanded as you add more sockets, but they will suffer too once they get to DDR3. They have already announced their own solution by splitting up a memory channel into two using an external chip to allow more modules per memory channel.

That's my point too, why does Intel and SUN) want their customer to be able to load up a server with so much RAM that suddenly FB-DIMMs become necessary? Performance.

Take away FB-DIMM and you will no-doubt be taking away performance.

Not the kind of performance that benchmarketing fails to capture, but rather the kind that apparantely exists in real-world server applications that customers are using.

I don't care for the power consumption of FB-DIMMs but I sure hope Intel's (and SUN's) platform engineers do NOT optimize their hardware for the sake of enabling Anandtech to generate some pretty but synthetic performance/watt graphs.

I also wonder about Itanium...where does it fit into the performance/watt situation?

QFT

While FBDs do give a very slight performance increase, the only reason Intel uses them on their current platforms is that they are the only way Intel can compete with AMD in memory capacity. Remember that on Opteron systems (this is not the whole story, but the best visual representation), each CPU has a dual channel memory controller, because it's built into the CPU itself...on the FSB modeled system that Intel uses, there is only a single dual channel memory controller for all of the CPUs to share (even Bensley and Caneland which have 2 and 4 connections to the MCH, there is still only a single dual channel controller).
As to moving away from FBDs, this will obviously happen when Nehalem comes at the end of 2009...it too will have a dual channel MC on each chip, and FBDs will no longer be necessary.

 

[Keysplayr]

You're all wrong. It's the beginning of.............................SKYNET.

 

[Viditor]

Originally posted by: keysplayr2003
You're all wrong. It's the beginning of.............................SKYNET.

So is this the time that John Connor becomes the new CEO of AMD?

 

[Keysplayr]

Is he available? I would hope the BOD's would find him the resourceful individual that he really is..
But seriously, thanks for your response to my questions above. Hey, if I had the money to burn, I'd buy AMD, or at least 51% of it's stock. I was just thinking to myself, Bill Gates could buy AMD outright for cash about 5 times over. Is that insane or what? A single individual, nevermind a big corporation. Wow.

I think AMD will slowly come back with Barcelona and Phenom. At least be a lot more competitive against C2D/Penryn. If the platform isn't insanely expensive, and it performs on par with Intel's offerings (price per price) I may consider a Phenom system. Just for the geek in me. My server here at home can use an upgrade, so it would inherit one of my older DC intel platforms here. Maybe give my C2D system to my son and I'll use the Phenom for myself. Maybe.

 

[zsdersw]

Originally posted by: Viditor
when Nehalem comes at the end of 2009...

You mean end of 2008 or beginning of 2009.

 

[Viditor]

Originally posted by: zsdersw

Originally posted by: Viditor
when Nehalem comes at the end of 2009...

You mean end of 2008 or beginning of 2009.

My bad...thanks for the correction.

 

[coldpower27]

Originally posted by: Idontcare

Originally posted by: coldpower27

Originally posted by: OneEng
It is also interesting to note that K10's power efficiency bests Intel's Clovertown and even Penryn in most cases. Intel is gaining ground here and if they ever abandon FBDIMM's, things may change quickly. The K10 is still a very cool operator showing that it can hang with an Intel 45nm part even at 65nm.

I will say that it's primarily due to the FB-DIMM's that the Stoakley platform post such high power consumption.

http://techreport.com/articles.x/13224/9

As seen there in terms of power consumed at high load the E5472 system already beats the 2360 SE system, and that is with the handicap of using FB-DIMM modules which consume considerable amounts of power.

The primary reason the AMD parts can hang in there is due to Intel's use of FB-DIMM's and not because the processors are particularly more efficient in terms of energy usage. The difference between the 2350 and the E5472 is only 43W, and there are 8 DIMM's there so even with a conservative figure like 5W per DIMM we are looking at 3 W difference which would be within a margin of error.

K10 is very competitive in terms of power usage due primarily because of the platform and not because of the processors themselves necessarily.

Any thoughts on how much performance would be lost were Intel to abandon FB-DIMM in favor of reducing system wattage?

Surely there is some performance related reason why Intel couples their server platform to FB-DIMM...if not then why would they stick with higher cost, higher power, and lower performance memory parts?

I can only conclude it is a trade-off for real-world performance, which will be lost should they back away from FB-DIMM. I also note they have zero road-map plans to back away from FB-DIMM, so Intel appears to intend to continue to rely on this performance boost for future server (and skulltrail'esque) platforms.

My point being that yes surely you can argue that migrating from FB-DIMM to standard DDR2 or DDR3 parts as AMD has done would drastically decrease Intel's server platform power consumption and ought to tip the performance/watt balance to their favor...but that argument assumes the performance decrease would not be so punitive as to negate the benefits of decreased power consumption.

Both the numerator and denominator will scale-back, and the ratio is no-doubt something Intel already knows and also no-doubt was a deciding factor in their decision to go with FB-DIMM.

I very much doubt Intel overlooked where their system watts are going in their servers when they set out to hit their performance/watt targets.

Oh I am not arguing about the merits or not of Intel going for FB-DIMM's I was just explaining where the bulk majority of the extra power consumption comes from, in comparison to the AMD platform.

The difference in cost between ECC DDR2 and FB-DIMM isn't extreme enough to be worth mentioning.

That is assuming there is a performance decrease at all from going to FB-DIMM to ECC DDR2, which is a big unknown and remains speculation at best. It's too bad we don't have exact identical platforms to compare the 2 memory types in.

I know Intel most have some crucial reason of why they must have chosen FB-DIMM's for their server platform.

 

[OneEng]

Originally posted by: Idontcare

Originally posted by: coldpower27

Originally posted by: OneEng
It is also interesting to note that K10's power efficiency bests Intel's Clovertown and even Penryn in most cases. Intel is gaining ground here and if they ever abandon FBDIMM's, things may change quickly. The K10 is still a very cool operator showing that it can hang with an Intel 45nm part even at 65nm.

I will say that it's primarily due to the FB-DIMM's that the Stoakley platform post such high power consumption.

http://techreport.com/articles.x/13224/9

As seen there in terms of power consumed at high load the E5472 system already beats the 2360 SE system, and that is with the handicap of using FB-DIMM modules which consume considerable amounts of power.

The primary reason the AMD parts can hang in there is due to Intel's use of FB-DIMM's and not because the processors are particularly more efficient in terms of energy usage. The difference between the 2350 and the E5472 is only 43W, and there are 8 DIMM's there so even with a conservative figure like 5W per DIMM we are looking at 3 W difference which would be within a margin of error.

K10 is very competitive in terms of power usage due primarily because of the platform and not because of the processors themselves necessarily.

Any thoughts on how much performance would be lost were Intel to abandon FB-DIMM in favor of reducing system wattage?

Surely there is some performance related reason why Intel couples their server platform to FB-DIMM...if not then why would they stick with higher cost, higher power, and lower performance memory parts?

I can only conclude it is a trade-off for real-world performance, which will be lost should they back away from FB-DIMM. I also note they have zero road-map plans to back away from FB-DIMM, so Intel appears to intend to continue to rely on this performance boost for future server (and skulltrail'esque) platforms.

My point being that yes surely you can argue that migrating from FB-DIMM to standard DDR2 or DDR3 parts as AMD has done would drastically decrease Intel's server platform power consumption and ought to tip the performance/watt balance to their favor...but that argument assumes the performance decrease would not be so punitive as to negate the benefits of decreased power consumption.

Both the numerator and denominator will scale-back, and the ratio is no-doubt something Intel already knows and also no-doubt was a deciding factor in their decision to go with FB-DIMM.

I very much doubt Intel overlooked where their system watts are going in their servers when they set out to hit their performance/watt targets.

I think that Intel would gain very little performance (lower latency) by moving away from FBDimms, but would lose a great deal of power consumption.

They would lose the ability to hold lots of memory which would hurt them (in performance) in database servers tremendously.

As for AMD's memory controller running at a different clock than the core, I believe that it is the opposite of what you are thinking. It is to allow the shared memory controller to run at full speed while the various cores run at lower speeds (not to allow the memory controller to run at slower speeds while the cores run faster).

 

[formulav8]

Last I heard is only the Server versions of Nehalem is getting the OMC. The Desktop versions are NOT getting the OMC.

Did Intel change their stance on that??


Jason

 

[OneEng]

Originally posted by: formulav8
Last I heard is only the Server versions of Nehalem is getting the OMC. The Desktop versions are NOT getting the OMC.

Did Intel change their stance on that??


Jason

No. I don't think so. The server versions will get an IMC the desktop versions will still have a north bridge.

This is the best of both worlds for Intel. They get all the performance advantages in the server world, and still get to charge for both a CPU and a North Bridge chip in the desktop and mobile markets.

Since AMD can't seem to get their act together, there is little need for Intel to provide a more value oriented solution

Perhaps AMD's fusion will force such a move on Intel's part?