The end of FSBs on Intel == the end of cheap multi-core?

[VirtualLarry]

I was just thinking about how Intel's current C2D Quad-cores are put together - two C2D dual-core dice in the same package, linked by their shared FSBs.

This configuration gives Intel huge manufacturing advantages. And it's all due to the "magic" of Intel's FSB designs, which allowed for cheap dual-socket SMP designs, and eventually dual-dualcore designs.

With Nahalem, this will all be at an end. But I wonder, how will Intel create Octo-core CPUs, with two native quad-core Nahalem chips? Especially if each quad-core is expecting it's own path to DRAM, with the integrated memory controller.

As far as I understand it, it's not easily possible to have two memory-controllers sharing the same DRAM. (Since the memory controller maintains a cache of things like open pages in DRAM and whatnot.)

So does this mean that we will wait longer for Octo-core CPUs, until Intel can create "native" Octo-cores, at some future generation?

I almost wonder if the loss of the FSB design is a step backwards for Intel, seeing as how their pre-fetcher in the C2D and the size of the L2 cache shows that they don't even need the IMC for performance reasons.

What do you all think?

 

[GundamF91]

Yeah, I agree that the current "quad" core Intel are really dual-dual cores, but that's probably as good as a independent Quad core design such that AMD is doing.

 

[21stHermit]

One of your assumptions is that all cores are equal. Is that true/necessary. By the time Intel gets to Octo, wouldn't it make sense to have one or more cores do video?

This is especially true given the dearth of multi-threaded applications. What Intel really needs to have those bright Israeli engineers invent is a way for single threads to run fast on all those cores.

Just musing . . .

 

[AlabamaCajun]

I would not be surprised if they have a way to channel ram between off die cores with CSI or the new IMC. That would carry this core welding into the future.
I've seen blurbs that a Neha on 775 may be release which may be possible to have 3 welded cores by 32nm making it 6 cores. Which kicks in the question, will Neha have a monolithic C2Q?

 

[darkxknight]

i think the shift towards clock speed will start coming back into play. what honestly needs more than 4 cores.

 

[jones377]

Originally posted by: AlabamaCajun
I would not be surprised if they have a way to channel ram between off die cores with CSI or the new IMC. That would carry this core welding into the future.
I've seen blurbs that a Neha on 775 may be release which may be possible to have 3 welded cores by 32nm making it 6 cores. Which kicks in the question, will Neha have a monolithic C2Q?

There existed low-end Opteron motherboards where the memory controller was only active on one CPU socket. The other one accessed memory only through the HT. I'd imagine that same topology would work for both Opteron and Nehalem (using a QuickPath link) in an octo-MCM in the future. And yes, Nehalem is a "true-quad" as can be seen from the die-photos released.

 

[myocardia]

Originally posted by: VirtualLarry
As far as I understand it, it's not easily possible to have two memory-controllers sharing the same DRAM. (Since the memory controller maintains a cache of things like open pages in DRAM and whatnot.)

It's not possible to have two memory controllers sharing the same RAM. That's why every 2P & 4P Opteron motherboard has RAM slots dedicated to each memory controller. That's also all Intel will have to do.

 

[Cogman]

Originally posted by: darkxknight
i think the shift towards clock speed will start coming back into play. what honestly needs more than 4 cores.

ummm... You do realize that not everyone uses their computers just for games, don't you?

Encoding will use all the cores you throw at it.
Compressing will as well.
Home server could use multiple cores.
HTPC for decoding would use multiple cores if you have multiple outputs trying to display HDTV content.
For that matter, if you do a one computer 3 console setup (I think they are coming) then more cores can be quite advantageous.
Finally, while not all games use 4 cores (in fact most don't) Some do. My bet is that developers will start utilizing more cores for smarter and more AI.

 

[Phynaz]

Originally posted by: darkxknight
i think the shift towards clock speed will start coming back into play. what honestly needs more than 4 cores.

Almost everything I do.

 

[AlabamaCajun]

Originally posted by: darkxknight
i think the shift towards clock speed will start coming back into play. what honestly needs more than 4 cores.

Think of it as a 3 dimensional chart or matrix all three are constantly in play, better tech is affecting the other 2.
Cores moved in the x direction (wider).
Speed moves in the y direction (higher).
Technology moves in the z direction (forward).

We are near the point where n-cores is going to be the norm. Linux and Vista are already equipped to start taking advantage of them. Game developers are moving in that direction and MS's programming languages already support the multi-threaded world though event driven menthods. You will notice mouse clicks and other input are more fluent but it still depends on the software application.

 

[Duvie]

Originally posted by: 21stHermit
One of your assumptions is that all cores are equal. Is that true/necessary. By the time Intel gets to Octo, wouldn't it make sense to have one or more cores do video?

This is especially true given the dearth of multi-threaded applications. What Intel really needs to have those bright Israeli engineers invent is a way for single threads to run fast on all those cores.

Just musing . . .

I thought there was something like this talked about once before. Something like reverse hyperthreading where it combines the cores into one large core for single threaded apps...or at least in theory....

If someone remembers a bit more then I obviously dont about this past subject...please refresh us

 

[VirtualLarry]

There was a mysterious BIOS option called "core multiplexing", that was later removed, found on a 975X Intel mobo. Conspiracy theorists believe that it was related to the so-called "reverse hyperthreading", in which the decode and execution units (not to mention the oh so convenient shared L2 cache) for both cores would work in tandem, to execute a single thread stream of opcodes. Essentially, it turned two 4-issue cores into a single 8-issue core. The problem is, there are precious few instruction streams with enough inherent parallelism to warrant throwing that much CPU resources at it.

Edit: Thanks for the mention of the QuickPath technology, I had forgotten about that. If Intel implements something similar to coherent HT, then it could easily be possible to glue multiple cores together, and have them share a single MC.