My thought on Intel's Quad core

newls1

Supreme [H]ardness
Joined
Sep 8, 2003
Messages
4,628
As I'm sitting here extremely bored, I thought about Intels QUAD CORE chips. Does it make sense to have 4 cores in 1 CPU that all share 1 FSB :confused: . Were as AMD's (QUAD FATHER) or rather up coming 4x4 platform will feature 2 SKT's each socket with its own independant bus (OR HTT). Dont take this post as me trying to "knock" 1 company from the other, im not, just thought about how AMD might have the better idea this time. I would think that the Intel 4 core chip is going to be bandwidth starved due to this limitation.
 
newls1 said:
As I'm sitting here extremely bored, I thought about Intels QUAD CORE chips. Does it make sense to have 4 cores in 1 CPU that all share 1 FSB :confused: . Were as AMD's (QUAD FATHER) or rather up coming 4x4 platform will feature 2 SKT's each socket with its own independant bus (OR HTT). Dont take this post as me trying to "knock" 1 company from the other, im not, just thought about how AMD might have the better idea this time. I would think that the Intel 4 core chip is going to be bandwidth starved due to this limitation.

http://www.theinquirer.net/default.aspx?article=35194

Nov 1 :eek:

Even while Quad Father will be a kick ass product, it is only meant for a few with a lot of money and bragging rights. AMD even went as far to rename the same old processor so it can jack up the prices of two cores needed. FX-72, FX74 and etc.. that are nothing more than 2.6, 2.8, and 3GHz Athlon64's. Then you'll need two Processors and a minimum of 4 sticks if RAM.

$X$ or 4 X 4 should really be compared to two Xeons on Bensley. Then bandwidth would look more like 26GB for AMD vs. 21GB for Intel. 48 PCI-E lanes is the only real thing that hurts Intel when compared to AMD's $x$ monster. AMD's might be faster but that advantage will depend on Platform, not processors.
 
newls1 said:
As I'm sitting here extremely bored, I thought about Intels QUAD CORE chips. Does it make sense to have 4 cores in 1 CPU that all share 1 FSB :confused: . Were as AMD's (QUAD FATHER) or rather up coming 4x4 platform will feature 2 SKT's each socket with its own independant bus (OR HTT). Dont take this post as me trying to "knock" 1 company from the other, im not, just thought about how AMD might have the better idea this time. I would think that the Intel 4 core chip is going to be bandwidth starved due to this limitation.

You are correct that a two-socket dual-core AMD system has some serious advantages over a single-socket 4-core 775. The best advantage is actually the integreated memory controller - two of them. Not only do the integrated memory controllers reduce latency drastically, in an AMD64 multi-socket system each processor has it's own RAM bank (NUMA). On top of that they have the fast Hyperlinks which enable talking between the CPUs much faster than Intel can even within the 4-core (remember it is two dies in one CPU casing).

However, that doesn't change the fact that the Core2 base performance is 20% or more above AMD64 at the same clockspeed, so unless you actually talk a whole lot to the outside world you will still be faster with the Intel. The Intel CPU also has the bigger cache which wouldn't be so useful of they weren't more intelligent at the same time. The Core2 cache feature speculative prefetch which can be a huge advantage when executing braindead code (most code is braindead).

Given that most multithreaded consumer applications are pretty low-quality multithreaded (dedicating cores to different tasks instead of splitting single tasks into threads sweeping over the same data) the whole AMD infrastrcture is probably of little use.

However, if you go 8 cores, and you have a tight, high-end coded multithreaded application AMD will clearly have the upper hand.

Also keep in mind that socket 775 can only have 4 RAM sockets which can be a serious problem for those applications.
 
the first kentsfield is set for $999? sounds like it's for bragging rights for those with a lot of money too. :D

price aside, there's pros and cons to both. obviously, with intel, you have the convenience of it all being in one package, whereas you gotta cool two seperate cpus for 4x4..
but with 4x4, there is more memory bandwidth... not that it really means a whole lot to K8 :p
 
About the FSB thing, Intel showed a slide at IDF (they posted it on anandtech i think) showing that with 2 cores, the FSB is barely saturated and that at four cores, the bus still isn't near saturation.
 
(cf)Eclipse said:
the first kentsfield is set for $999? sounds like it's for bragging rights for those with a lot of money too. :D

price aside, there's pros and cons to both. obviously, with intel, you have the convenience of it all being in one package, whereas you gotta cool two seperate cpus for 4x4..
but with 4x4, there is more memory bandwidth... not that it really means a whole lot to K8 :p


Hey I agree, no double standards here.

Yet, in Intel's defense, for their Quad cores you only need one ;) For AMD to be competitive that means each Dual Core FX7x will have to go for about $500 to get to Quad? Again, some of my Bud's paid $1089 to $1150 for FX-60's when they were all the rage, one also has a $899 FX-57. I'm not spending more than $400 on any Intel or AMD processor for my personal use.
 
Duke3d87 said:
About the FSB thing, Intel showed a slide at IDF (they posted it on anandtech i think) showing that with 2 cores, the FSB is barely saturated and that at four cores, the bus still isn't near saturation.

True but that's why AMD showed 4 X 4 running with a lot more activity. They saturated theirs until it did make the processors reach 100%. They know at that point, they'd have an advantage.
 
Donnie27 said:
True but that's why AMD showed 4 X 4 running with a lot more activity. They saturated theirs until it did make the processors reach 100%. They know at that point, they'd have an advantage.
That'd be the PROCs maxed out to 100%, not the proc-RAM link ;). (or am I wrong in thinking just cause a CPU is chugging, RAM isn't?)
I just started a thread in Gen Hardware about 4x4 and Kentsfield. D'oh, didn't even see this one. Oh well, go post there :p. http://www.hardforum.com/showthread.php?goto=newpost&t=1111210

-Ray.
/shamless plug
 
Looks like AMD is going to be priced out the ballpark.
You need twice as many of everything!!
2 x $50 HSF
4 sticks of ram
2 x $800 ? (FX processors)

And all they're doing is just rebranding opterons 285, 290 & 295 as FX chips. Nothing new here.

P.S. the most I'll spend on a processor is around $300-400, both camps will be out of my price range for a while. Though I know I'll be getting a Kentsfield, b/c it will be plug & play with my current Core 2 board!
 
uOpt said:
Also keep in mind that socket 775 can only have 4 RAM sockets which can be a serious problem for those applications.

Many great points, but this one isn't quite fair. First a correction. Those 4 Sockets are vailidated to work with 2GB Sticks while most of AMDs are NOT! So the 8GB or 4ea 2GB Sticks, Max figure given is real, not just a marketing Gimmick.

To be fair, the $ X $ should be compared to a two Socket Bensley Board. These feature 8 RAM Slots and can run 16GB of RAM. It's two FSB's doesn't bottleneck as bad as one would.
 
chrisf6969 said:
Looks like AMD is going to be priced out the ballpark.
You need twice as many of everything!!
2 x $50 HSF
4 sticks of ram
2 x $800 ? (FX processors)

And all they're doing is just rebranding opterons 285, 290 & 295 as FX chips. Nothing new here.

You left off the a new Power Supply if you don't already something substantial.
 
Eulogy said:
That'd be the PROCs maxed out to 100%, not the proc-RAM link ;). (or am I wrong in thinking just cause a CPU is chugging, RAM isn't?)
I just started a thread in Gen Hardware about 4x4 and Kentsfield. D'oh, didn't even see this one. Oh well, go post there :p. http://www.hardforum.com/showthread.php?goto=newpost&t=1111210

-Ray.
/shamless plug

Maybe but either way, they'd have an advantage!
 
I'm HONESTLY not posting this to get flammed, or start a war...

Does anyone have a link, detailing and proving that the FSB on Intel platforms is, or is not, a bottleneck?

Sure, we can all plug theory and calculate to our hearts contents, but that doesn't give solid, real-life performance. I just like to see the raw data showing me this one way or another.

And in 2nd thought, is there any benchmark or validation to HTT being superior/inferior to Intel in the real world, using every-day (or at least common) apps?

-a confused, Ray
 
Donnie27 said:
Maybe but either way, they'd have an advantage!
An advantage that may or may not play out, though. Sure, if we can saturate a FSB, then HTT will look better. And I'm sure Intel is looking into their own version of it for a reaon. See above post from me :p.

-Ray. Still confused, but learning.
 
Eulogy said:
An advantage that may or may not play out, though. Sure, if we can saturate a FSB, then HTT will look better. And I'm sure Intel is looking into their own version of it for a reaon. See above post from me :p.

-Ray. Still confused, but learning.

Intel is working on CSI (not the TV program) to help interprocessor and links to IMC as well. NEC and IBM have always used Multiple North Bridges to get around such bottlenecks.

Here's an example.

Tigerton

October 19, 2006 (3:42 PM EDT)
Intel To Preview Quad-Core Xeon MP
By Alexander Wolfe, TechWeb Technology News

Intel is planning to preview on Friday its ultra-high-end, quad-core Xeon MP processor. The chip, code-named Tigerton, is scheduled to ship in the second half of 2007, an Intel spokesman confirmed.

Called a "Pipedream" and "Vapourware" by AMD F@N$ all over the web. Now Intel knows dmaned well they can't just keep adding cores without giving them bandwdith.
 
Donnie27 said:
Many great points, but this one isn't quite fair. First a correction. Those 4 Sockets are vailidated to work with 2GB Sticks while most of AMDs are NOT! So the 8GB or 4ea 2GB Sticks, Max figure given is real, not just a marketing Gimmick.

But both DDR2 AMD platforms take 2 GB sticks just fine.

And DDR(1) Socket 940 also took 2 GB sticks.

Dual-socket Socket 940 and Socket F/1207 can carry at least twice as much RAM as 775 (8 slots on dual, with everybody on 2 GB sticks), except for the registered platform you can get 4 GB sticks, which makes the F/1207 system come up with 4 times as much memory - at full speed and with separate banks for each processor.

To be fair, the $ X $ should be compared to a two Socket Bensley Board. These feature 8 RAM Slots and can run 16GB of RAM. It's two FSB's doesn't bottleneck as bad as one would.

But then you have FBDIMMS. I think (not verified, would like to), that FBDIMMS should slow down quite a lot when you max out all memory slots. Accessing the farest slot should have quite a bit of additional latency due to the serial signaling through the module chain.
 
uOpt said:
You are correct that a two-socket dual-core AMD system has some serious advantages over a single-socket 4-core 775. The best advantage is actually the integreated memory controller - two of them. Not only do the integrated memory controllers reduce latency drastically, in an AMD64 multi-socket system each processor has it's own RAM bank (NUMA). On top of that they have the fast Hyperlinks which enable talking between the CPUs much faster than Intel can even within the 4-core (remember it is two dies in one CPU casing).

I'm pretty sure that NUMA is mobo dependent: most dual socket 940 boards were NUMA systems but not all (I'm pretty sure the iWill SFF kits are not).

Has there been any confirmation as to whether 4x4 will be NUMA?
 
Slartibartfast said:
I'm pretty sure that NUMA is mobo dependent: most dual socket 940 boards were NUMA systems but not all (I'm pretty sure the iWill SFF kits are not).

Yes. In particular, some normal-ATX (not EATX) size boards offered only 4 RAM slots on the first CPU.

It is not that NUMA offers a measurabl advantage in a two-single-core situation with memory at 200 MHz and CPUs at 2000 MHz. Dual-cores at 3 GHz can trash memory quite a bit more.

Has there been any confirmation as to whether 4x4 will be NUMA?

If it is 1207 based it certainly will, but again, ATX size board don't have the space for two sockets and more than one RAM bank. And even then you can't mount decent air cooling.

The board in the picture of the demo system clearly looks like a non-NUMA unregistered board, though, aka just 4 DDR2 unreg slots. It's normal ATX size, too.
 
Donnie27 said:
Intel is working on CSI (not the TV program) to help interprocessor and links to IMC as well. NEC and IBM have always used Multiple North Bridges to get around such bottlenecks.

Here's an example.

Tigerton



Called a "Pipedream" and "Vapourware" by AMD F@N$ all over the web. Now Intel knows dmaned well they can't just keep adding cores without giving them bandwdith.

I don't know if this is true but i thought i'd throw it up there and see what happens. AMD claims that the Intel NB will use a lot of power and having several GB of FBDIMM is supposed to ramp up power consumption pretty high. There's supposed to be a second generation of FBDIMM that Intel said would reduce power consumption. Anyone know when that will be out?

What i think would be cool is a quad core Xeon with a quad channel FBDIMM controller. That'd be just fun.
 
uOpt said:
Yes. In particular, some normal-ATX (not EATX) size boards offered only 4 RAM slots on the first CPU.

It is not that NUMA offers a measurabl advantage in a two-single-core situation with memory at 200 MHz and CPUs at 2000 MHz. Dual-cores at 3 GHz can trash memory quite a bit more.

Given all that then, I would bet that at some point having NUMA would be a significant advantage over not. I mean right now the difference is negligible, but if you had 8 cores sharing the memory bus the latency would be much greater. How many cores can you pack into a single socket until you have to do something about the memory bandwith?

We may someday see a system with 8+ cores in one cpu, with two IMC's: NUMA on a single socket, given that memory is still lagging so far behind cpu's.
 
Eulogy said:
I'm HONESTLY not posting this to get flammed, or start a war...

Does anyone have a link, detailing and proving that the FSB on Intel platforms is, or is not, a bottleneck?

Sure, we can all plug theory and calculate to our hearts contents, but that doesn't give solid, real-life performance. I just like to see the raw data showing me this one way or another.

And in 2nd thought, is there any benchmark or validation to HTT being superior/inferior to Intel in the real world, using every-day (or at least common) apps?

-a confused, Ray

I am interested in this question as well.
 
Slartibartfast said:
Given all that then, I would bet that at some point having NUMA would be a significant advantage over not. I mean right now the difference is negligible, but if you had 8 cores sharing the memory bus the latency would be much greater. How many cores can you pack into a single socket until you have to do something about the memory bandwith?

It is also a software problem. Coming up with the memory pages that one process or one specific CPU runs to be in that CPU's local memory is extremely non-trivial unless you are willing to severely restict the freedom of the scheduler (which so far never worked out in 5 decades of OS research) or the memory allocator (probably an even worse idea).

The only software showing a real NUMA benefit so far is on machines running literally nothing else but a single database server which locks all threads to specific CPUs and manually controls memory page placement (and thereby burning quite a bit of CPU, but since they are memory and I/O bound they can afford that). That won't work for any general-purpose workload, not even a webserver.

The Linux guys are going pretty radical in trying to exploit NUMA, they are working on great schemes of page migration to have memory pages follow their processes around through the system, migrating from one memory band to another . Whether that is such a brillant idea I don't know yet, the copying overhead is monstrous. But it allows you to let the scheduler do it's thing so maybe it'll pay off.

We may someday see a system with 8+ cores in one cpu, with two IMC's: NUMA on a single socket, given that memory is still lagging so far behind cpu's.

Personally I suppose that AMD will release it's first 8-cores the way Intel did their first 2- and 4-cores: two dies on one chip. And then have two memory interfaces out of the chip.
 
uOpt said:
But both DDR2 AMD platforms take 2 GB sticks just fine.

And DDR(1) Socket 940 also took 2 GB sticks.

Dual-socket Socket 940 and Socket F/1207 can carry at least twice as much RAM as 775 (8 slots on dual, with everybody on 2 GB sticks), except for the registered platform you can get 4 GB sticks, which makes the F/1207 system come up with 4 times as much memory - at full speed and with separate banks for each processor.

But then you have FBDIMMS. I think (not verified, would like to), that FBDIMMS should slow down quite a lot when you max out all memory slots. Accessing the farest slot should have quite a bit of additional latency due to the serial signaling through the module chain.

NOPE, one of the advantages, I'm told is that FBDIMM DON'T suffer (more than they already do) when large amounts are used.

NO, many so called 4GB max boards had Problems out the WA-Zoo, especially on the AMD side. Many had to slow the RAM's speed from 3200 to 2700 just to stay stable when all four slots were full.

NO, there is no way hell to get 4 times as much memory from 4 X 4 when compared to Bensley=P What are you talking about? I said compared to Bensley (21GB) that's Intel's Dual Socket platform:) Now, do you want to talk One socket vs. one socket or Two Sockets vs Two Sockets? That's why I said to be fair!

Oh and one thing the folks on the other side of the forum loves to point out, I'm Kentsfield ready right now LOL! No Socket change and New motherboard, no adding more RAM or Power Supply, no two Power hungry $1000 video cards and etc................. Who out there think Dual Socket motherboards will cost less than $300? Again, some one will add the needed 48 PCI-E lanes to one of Intel's server boards and wa-la, BOTH INTEL and AMD will have a Money Snatching Gimmick!
 
Donnie27 said:
NOPE, one of the advantages, I'm told is that FBDIMM DON'T suffer (more than they already do) when large amounts are used.

You can have more modules without lowering the RAM frequency but the latency for the far modules should increase. It's how the serial interface of FBDIMM works.

Donnie27 said:
NO, many so called 4GB max boards had Problems out the WA-Zoo, especially on the AMD side. Many had to slow the RAM's speed from 3200 to 2700 just to stay stable when all four slots were full.

Unregistered RAM always had this problem, and if you plug in enough RAM registered RAM has it, too. In practice I run 4 socket 939 machines right now with 4 sticks and all of them do 400MHz or better, and one socket 940 system with 8 sticks at 400 MHz.

Anyway, that's not the topic of this thread.

Donnie27 said:
NO, there is no way hell to get 4 times as much memory from 4 X 4 when compared to Bensley=P What are you talking about? I said compared to Bensley (21GB) that's Intel's Dual Socket platform:) Now, do you want to talk One socket vs. one socket or Two Sockets vs Two Sockets? That's why I said to be fair!

At that point in my quote we were comparing socket 1207 based 4x4 to socket 775 with Kentsfield, which is the main topic of this thread. There you have twice the sockets and twice the module capacity leading straight to 4 times as much memory capacity.

In the paragraph afterwards we mentioned Intel dual-socket solutions for the first time.

Donnie27 said:
Oh and one thing the folks on the other side of the forum loves to point out, I'm Kentsfield ready right now LOL! No Socket change and New motherboard, no adding more RAM or Power Supply, no two Power hungry $1000 video cards and etc................. Who out there think Dual Socket motherboards will cost less than $300? Again, some one will add the needed 48 PCI-E lanes to one of Intel's server boards and wa-la, BOTH INTEL and AMD will have a Money Snatching Gimmick!

The AMD folks using socket 940 with two dual-cores also enjoyed 4-core at reasonable prices for more than a year :)
 
Duke3d87 said:
I don't know if this is true but i thought i'd throw it up there and see what happens. AMD claims that the Intel NB will use a lot of power and having several GB of FBDIMM is supposed to ramp up power consumption pretty high. There's supposed to be a second generation of FBDIMM that Intel said would reduce power consumption. Anyone know when that will be out?

What i think would be cool is a quad core Xeon with a quad channel FBDIMM controller. That'd be just fun.

AMD is RIGHT! But there was one catch, when similar systems were tested, the *WHOLE Intel system still used less power. I heard about FBDIMM part #2 and that Some platforms will feature DDR2 and 3 but that's just speculation.

http://www.techweb.com/showArticle.jhtml?articleId=161500005&pgno=1
 
Each FBDIMM module takes about 7 watts right now.

That is quite significant given that the CPUs take something around 70-90 (the TDP is not the power consumption) and that these systems are usually used in environments where a whole lot of RAM is needed. 8 RAM modules = 56 watts extra, plus PSU efficiency, plus fans required to get rid the 56 watts.

To add insult to injury, I think the 7 watts are pretty much permanently even if the machine is idle and the CPU takes substancially less.

I don't see a reason why they wouldn't be able to just invent a bloody better register chip taking 2 watts, or 0.1 watts, though.
 
uOpt said:
You can have more modules without lowering the RAM frequency but the latency for the far modules should increase. It's how the serial interface of FBDIMM works.

BUT as I already stated smart access to that RAM is not the same=P

uOpt said:
Unregistered RAM always had this problem, and if you plug in enough RAM registered RAM has it, too. In practice I run 4 socket 939 machines right now with 4 sticks and all of them do 400MHz or better, and one socket 940 system with 8 sticks at 400 MHz.

True! But Intel didn't suffer that same problem with four slots full. My i875 was proof :)

Anyway, that's not the topic of this thread.

Excuse me?

At that point in my quote we were comparing socket 1207 based 4x4 to socket 775 with Kentsfield, which is the main topic of this thread. There you have twice the sockets and twice the module capacity leading straight to 4 times as much memory capacity.

But you posted and quoted it to me.

In the paragraph afterwards we mentioned Intel dual-socket solutions for the first time.

Sure, I know some folks like handicapped matches but that doesn't interest me at all.


uOpt said:
The AMD folks using socket 940 with two dual-cores also enjoyed 4-core at reasonable prices for more than a year :)

Reasonable is an Opinion so I have no arguement there.

Donnie27 said:
Many great points, but this one isn't quite fair. First a correction. Those 4 Sockets are vailidated to work with 2GB Sticks while most of AMDs are NOT! So the 8GB or 4ea 2GB Sticks, Max figure given is real, not just a marketing Gimmick.

To be fair, the $ X $ should be compared to a two Socket Bensley Board. These feature 8 RAM Slots and can run 16GB of RAM. Its two FSB's doesn't bottleneck as bad as one would.

Note. I said there were many great points but only disagreed with the last part of your post.
 
uOpt said:
Each FBDIMM module takes about 7 watts right now.

That is quite significant given that the CPUs take something around 70-90 (the TDP is not the power consumption) and that these systems are usually used in environments where a whole lot of RAM is needed. 8 RAM modules = 56 watts extra, plus PSU efficiency, plus fans required to get rid the 56 watts.

To add insult to injury, I think the 7 watts are pretty much permanently even if the machine is idle and the CPU takes substancially less.

I don't see a reason why they wouldn't be able to just invent a bloody better register chip taking 2 watts, or 0.1 watts, though.


I agree! This was talked about at length on Xtreme systems I think.
 
Back
Top