Dual Cores

[Gamingphreek]

Well as we all know when dual cores come out processor speeds will have to decrease. However dual cores will act as dual processors, right? SO they will really only benefit from multi tasking. In games and benchmarking and what not will we see a performance decrease before we see an increase.

-Kevin

 

[AdamK47]

Thats exactly why I wonder how people can be so excited about dual core CPUs when they will mostly benifit from multithreaded apps and multitasking.

 

[Gothgar]

Originally posted by: Gamingphreek
Well as we all know when dual cores come out processor speeds will have to decrease. However dual cores will act as dual processors, right? SO they will really only benefit from multi tasking. In games and benchmarking and what not will we see a performance decrease before we see an increase.

-Kevin

it isn't as simple as that how they will react with each other.

They will have tons and tons of bandwidth between each other, and they are supposed to load balance each other, kinda like raid 0 but on a processor...

 

[cbehnken]

Umm, Dual CPUs do this already. The "load balance" only works with multiple threads. You can't split a thread into two processors. On the Intel side of the fence they will likely have no more bandwidth than now. AMD is already using Hypertransport so they have tons of bandwidth now already.

 

[klah]

I am sure AMD/Intel (whichever gets DC out first) will have a set of 'recommended' benchmarks included with the press kit to show off where these products excel.

 

[Gothgar]

Originally posted by: cbehnken
Umm, Dual CPUs do this already. The "load balance" only works with multiple threads. You can't split a thread into two processors. On the Intel side of the fence they will likely have no more bandwidth than now. AMD is already using Hypertransport so they have tons of bandwidth now already.

but with dual cpus there is know way possible over the mobo's bus to reach the same bandwidth as another core on die... and no, you cannot split a single thread onto more than one core, but multithreaded processes will benefit hugely...

Plus it will be cheaper and closer to mainstream than setting up a dual opteron processor system...

 

[cbehnken]

They are going to use Hypertransport internally as well. The latency may be lower, but there will not be that much gained in bandwidth. Opteron's already have enough bandwidth (compared to Intel at least) as it is anyway.

I agree about the benefits of Dual core, but it is nothing like Raid 0. In the end I don't see how it will ever be more than 5-%10 faster than an equivalent Dual Cpu rig. Plus it will most likely be harder to cool and overclock. We also have to look at potential failures and production potential. If one part of a Dual core cpu fails or is buggy from the factory you either have a useless processor or some kind of Sempron type deal where they disable one core.

If AMD has a dual core chip with one core capable of 2.6 Ghz and the other only capable of 2.0 Ghz then you'll have a 2.0Ghz chip with trapped potential. If you are going true dual cpu you can trade the 2.0 Ghz for a 2.6 and have both rocking at 2.6 Ghz.

This reminds me of the original Pentium Pro. If the huge (for the time) onboard cache (in the core, a split design, but in the same chip) had a bug in it the whole chip was thrown out. This is why at the time Intel switched to the card design and place the L2 chips off core.

If you have your own Dual System and one CPU fails you can yank it and keep running. Then replace it at your own pace. If one of your cores fails your system will be useless (most likely) I suppose it this way it is like Raid 0 because your chance of failure at least is doubled.

 

[Gamingphreek]

Well we already know that Intel will have bandwidth problems. However what should performance be like. Obviously we are going to see a fairly large performance jump as we wont be multitasking, right?

Also what about heat. Imagine what the Intel cooler is going to be like, hell imagine what the AMD cooler is going to be like, both are going to increase in size by a lot.

-Kevin

 

[Wingznut]

Originally posted by: Gamingphreek
Well we already know that Intel will have bandwidth problems.

We do?

 

[Gamingphreek]

Uh yeah... Intel has always been starved for bandwidth.

-Kevin

 

[Viditor]

They are going to use Hypertransport internally as well

Not between the 2 cores, but they will between cores in a multi-chip platform. They are using what they call "Direct Connect Architecture" (significantly faster than HT...)

 

[Viditor]

Originally posted by: Gamingphreek
Well we already know that Intel will have bandwidth problems. However what should performance be like. Obviously we are going to see a fairly large performance jump as we wont be multitasking, right?

Also what about heat. Imagine what the Intel cooler is going to be like, hell imagine what the AMD cooler is going to be like, both are going to increase in size by a lot.

-Kevin

We already know that the Opteron dual cores are rated at 89W...shouldn't be too bad...

 

[iwantanewcomputer]

yeah, the 89 watt rating stas for all dual cores theyve announced sofar. the move to 90 nm and continued improvement of 90 nm manufacturing will keep chips in this rating easily at the speeds they are planning. i've seen power tests and most 130 nm athlon 64s are well below the 89 rating. can't give a link, but it think i've seen like 45-60 for all the current chips. 90 nm will cut this significantly, and i doubt we will see any 2.6 duallies soon

 

[cbehnken]

So each core won't have it's own memory controller?

This will be significantly slower than an Opteron setup that has 2 memory controllers, controlling independent memory banks. Memory bandwidth will be halved.

 

[cbehnken]

Originally posted by: Viditor

They are going to use Hypertransport internally as well

Not between the 2 cores, but they will between cores in a multi-chip platform. They are using what they call "Direct Connect Architecture" (significantly faster than HT...)

When AMD mentions that their K8 architecture was designed for multicore operation from the start, they weren't lying. Each Socket-939 or Socket-940 K8 chip, whether it's an Athlon 64, Athlon 64 FX or Opteron, features three Hyper Transport links (whether they are all operational is another question). In order to create a dual core version of a K8 based chip, you simply remove a single pair of Hyper Transport PHYs, one from each chip, and fuse the two Hyper Transport links together - thus creating a direct path of communication between the two cores, capable of transmitting data at up to 8GB/s (at 1GHz) between the two chips.

Quote from anandtech, they are indeed using the hypertransport, although it is isolated from the main one.

 

[Viditor]

Originally posted by: cbehnken

Originally posted by: Viditor

They are going to use Hypertransport internally as well

Not between the 2 cores, but they will between cores in a multi-chip platform. They are using what they call "Direct Connect Architecture" (significantly faster than HT...)

When AMD mentions that their K8 architecture was designed for multicore operation from the start, they weren't lying. Each Socket-939 or Socket-940 K8 chip, whether it's an Athlon 64, Athlon 64 FX or Opteron, features three Hyper Transport links (whether they are all operational is another question). In order to create a dual core version of a K8 based chip, you simply remove a single pair of Hyper Transport PHYs, one from each chip, and fuse the two Hyper Transport links together - thus creating a direct path of communication between the two cores, capable of transmitting data at up to 8GB/s (at 1GHz) between the two chips.

Quote from anandtech, they are indeed using the hypertransport, although it is isolated from the main one.

From AMD...
http://www.amd.com/us-en/asset...31_DC_presentation.pdf

"AMD?s evolutionary Direct Connect Architecture connects two cores on the same die"

Basically, the 2 cores are connected through what amounts to a switcher.The speed of this connection is vastly more than HT...

 

[Viditor]

Originally posted by: cbehnken
So each core won't have it's own memory controller?

This will be significantly slower than an Opteron setup that has 2 memory controllers, controlling independent memory banks. Memory bandwidth will be halved.

If memory bandwidth were an issue for the Opteron, I'd agree with you...but it's not.

 

[DrMrLordX]

Originally posted by: Viditor

Originally posted by: cbehnken
So each core won't have it's own memory controller?

This will be significantly slower than an Opteron setup that has 2 memory controllers, controlling independent memory banks. Memory bandwidth will be halved.

If memory bandwidth were an issue for the Opteron, I'd agree with you...but it's not.

But . . . but . . . think of the Sandra benchmarks! Oh, the humanity!

. . .

teehee

Aaaaanyway, in light of the fact that most/all the x86 cpu manufacturers in the world are shifting to a dual-core strategy for now, shouldn't we expect them to start applying a lot more pressure to software developers and publishers to start producing multi-threaded apps and games?

 

[Viditor]

When AMD mentions that their K8 architecture was designed for multicore operation from the start, they weren't lying

I should have mentioned something about this as well...
No, they weren't. But what they meant was that the crossbar/SRI that you see in the diagram I linked has been implemented in all of the Hammers. In fact, if you go back to even the earliest digrams, you'll see the connections there for both CPU 0 and CPU 1 (even though there IS no CPU 1 core on the die...).

 

[cbehnken]

Originally posted by: DrMrLordX

Originally posted by: Viditor

Originally posted by: cbehnken
So each core won't have it's own memory controller?

This will be significantly slower than an Opteron setup that has 2 memory controllers, controlling independent memory banks. Memory bandwidth will be halved.

If memory bandwidth were an issue for the Opteron, I'd agree with you...but it's not.

But . . . but . . . think of the Sandra benchmarks! Oh, the humanity!

. . .

teehee

Aaaaanyway, in light of the fact that most/all the x86 cpu manufacturers in the world are shifting to a dual-core strategy for now, shouldn't we expect them to start applying a lot more pressure to software developers and publishers to start producing multi-threaded apps and games?

Probably, but it is going to be difficult going for them. It may involve a lot of changes th DirectX. The big problem is what do you do about people with single core chips? Maintain two versions of the game/app or just allow them to suffer a performance loss because of the task switching overhead?

 

[Viditor]

it is going to be difficult going for them. It may involve a lot of changes th DirectX

It shouldn't...it will allow game developers to write seperate but parallel threads for both graphics and the AI parsor...

The big problem is what do you do about people with single core chips? Maintain two versions of the game/app or just allow them to suffer a performance loss because of the task switching overhead?

The mods will be the same as those for Hyperthreading. But to answer your question...yes, single cores will eventually be much lower performers. On the good side, single cores will become a thing of the past as well...

 

[imported_michaelpatrick33]

Remember, also, that AMD is releasing dualcore to the server/workstation market first where dual threaded applications abound. So these markets should say obvious performance improvements while personal use single threaded applications begin the transition to dual threaded. We probably won't see increased performance on dualcore for a year on the home computer dualcore but will immediately in the server market. That is why AMD is going that route first. Intel seems to want the personal market first. We shall see.

 

[dguy6789]

My new system 3 years from now will be dual core dual opterons(or whatever the best dual cpu out is). mmmmmmmmmmm........

 

[Gamingphreek]

With all the questions and the imminence of dual cores i thought it might be a good time to bump this thread.

Referring back to the thread title will we not see a slow down until developers effectively optimize all apps and games for multi threading?

-Kevin

 

[Rand]

Originally posted by: Gamingphreek
Uh yeah... Intel has always been starved for bandwidth.

-Kevin

I'm not sure their as desperately memory bandwidth starved as you make out at this point in time, the 1066 FSB didn't exactly help the P4 EE much at all, and while it would likely benefit current 1MB L2 Prescott's to a greater degree the Prescott won't be seeing the 1066FSB till' the Prescott-2M launches in 1Q05 if then, and it shouldn't be much more then incrementally more bandwidth hungry then the current P4EE which won't likely.

The first Intel DualCore processors on the Smithfield core, with a mere 800MHz FSB between them will likely be pretty bandwidth starved however.
Beyond that they have plans for a seperate Point to Point FSB on the Xeon, which should in time make it's way over to the DualCore P4 platform.