Nehalem - Waiting Risky?

[Idontcare]

Originally posted by: Viditor
I am not positive...but I believe it to be true (I have been looking for confirmation either way, but to no avail as of yet).
The last I heard there will be 4 different sockets, with/without QPI and with/without IMC.

It would be pretty amazing if true as that would mean Nehalem cache hiearchy (L1/L2/L3) is either redesigned for the non-IMC version or the Nehalem cache hiearchy is crazy robust and can take such widely varying RAM access times (quick w/IMC versus delayed w/o IMC) and manage to not totally bottleneck the CPU.

In other words, if Nehalem's performance is not dependent on the IMC then shame on Intel for wasting shareholder profits by reducing GM's and needlessly including the IMC on the Nehalem.

If Nehalem's performance is dependent on the IMC then shame on Intel for wasting shareholder profits by increasing up-front R&D design expenses in creating a needless Nehalem SKU for a market condition where a Penryn could have readily been fielded.

 

[Viditor]

Originally posted by: Idontcare

Originally posted by: Viditor
I am not positive...but I believe it to be true (I have been looking for confirmation either way, but to no avail as of yet).
The last I heard there will be 4 different sockets, with/without QPI and with/without IMC.

It would be pretty amazing if true as that would mean Nehalem cache hiearchy (L1/L2/L3) is either redesigned for the non-IMC version or the Nehalem cache hiearchy is crazy robust and can take such widely varying RAM access times (quick w/IMC versus delayed w/o IMC) and manage to not totally bottleneck the CPU.

In other words, if Nehalem's performance is not dependent on the IMC then shame on Intel for wasting shareholder profits by reducing GM's and needlessly including the IMC on the Nehalem.

If Nehalem's performance is dependent on the IMC then shame on Intel for wasting shareholder profits by increasing up-front R&D design expenses in creating a needless Nehalem SKU for a market condition where a Penryn could have readily been fielded.

I find that I can't argue with any of your points (and you KNOW how that frustrates me!!).
That said, those are the rumours I have heard...though as always, they should be taken with the same amount of salt as found in Utah's famous wet spot...

palladium
"Oops, my bad.. it should read " some Nehalems have an integrated GPU on the same CPU package, as opposed to on the same die.""

Fair enough mate...though I believe those aren't due until 2009/2010 (about when Fusion is due)

 

[Idontcare]

Originally posted by: Viditor
I find that I can't argue with any of your points (and you KNOW how that frustrates me!!).
That said, those are the rumours I have heard...though as always, they should be taken with the same amount of salt as found in Utah's famous wet spot...

Hey I'm just stating the obvious boundary conditions that ought to exist...nothing says Intel management values those boundary conditions, Prescott wasn't exactly the hallmark of management maximizing shareholder returns.

Nehalem without IMC may very well be in the works, I was merely expressing my surprise as to some of the ramifications such a truth would bring with it. (and no doubt got half of it wrong to boot)

 

[VirtualLarry]

Nehalem without IMC would be rather stupid, no? I believe that all of them will have an IMC. What gets more interesting is the question of whether or not they can access RAM via the QPI and disregard the IMC. I'm thinking of multi-core NUMA systems like the 2xx Opterons. You could have one Nehalem with the IMC, and another Nehalem that accesses the first CPU's RAM via the QPI. If so, then there could always be an alternative design, in which the northbridge had a memory controller, and the Nehalem accessed the RAM via the QPI. Why would you want to do such a thing, and not use the IMC? Cost. You could design a chipset to run on the cheaper DDR2, instead of the DDR3, and you wouldn't have to wire up all of the IMC's pins on the socket.

 

[guptasa1]

Eh with so many changes, I think I've definitely decided to go for a system now rather than later. There's great stuff out there that's reliable (a must) and so much more of an improvement over what I've got that it's hard to pass up, plus I have a bit more time now to build and play around. Hoping to order in the next week or two!

 

[Idontcare]

Originally posted by: VirtualLarry
Nehalem without IMC would be rather stupid, no? I believe that all of them will have an IMC. What gets more interesting is the question of whether or not they can access RAM via the QPI and disregard the IMC. I'm thinking of multi-core NUMA systems like the 2xx Opterons. You could have one Nehalem with the IMC, and another Nehalem that accesses the first CPU's RAM via the QPI. If so, then there could always be an alternative design, in which the northbridge had a memory controller, and the Nehalem accessed the RAM via the QPI. Why would you want to do such a thing, and not use the IMC? Cost. You could design a chipset to run on the cheaper DDR2, instead of the DDR3, and you wouldn't have to wire up all of the IMC's pins on the socket.

I agree it does sound like a feasible engineering option, but it would be quite a low-volume niche for >1 socket systems where the purshaser is sooo budget minded that they want that $5 cheaper motherboard which has only one IMC wired up to the DIMMs and am willing to take the performance hit on the QPI hops after spending all that cash on a CPU which has an IMC.

That would be an atypical multi-socket consumer to say the least.

 

[ShadoutDH]

I'm going to buy a new PC now, and has also been wondering whether or not to wait for Nehalem, but came up with 2 reasons not to:

1) Even though it might be released late 2008, its likely it will be the same as the Penryn release, where it will be really hard to find the most popular (best price/performance) ones for 3-4 months. We might easily be in Q2 2009 before its possible to get a Nehalem for a reasonable price.

2) You kinda wait for the unknown. Who knows how Nehalem will really be when it arrives. Benchmakes seems awesome, but it might have issues etc which will take time to fix (not just the CPU, but also motherboards and whatever else). In the meantime you could have a working CPU now
Personally I think I wont jump on the Nehalem wagon before the next 'tick' (or is it 'tock'? ) with the second revision and 32nm die shrink of Nehalem (and hey, maybe an 8-core! if applications/games every start actually using multicores ^^).

 

[IntelUser2000]

Yeah, I thought all Nehalems have IMC, some Nehalems have integrated GPU on die.

Well, here's how it goes. There will be 3 versions.

Bloomfield-3 channel DDR3 IMC with QPI
Lynnfield/Clarksfield(Desktop/Mobile)-2 channel DDR3 IMC without QPI
Havendale/Auburndale(Desktop/Mobile)-Memory controller and GPU on one die connected by QPI to the CPU core

You could probably call Havendale/Auburndale as having an IMC since physically the two parts(the CPU core and GPU/memory controller core) will be very close and will be connected by QPI, which will have a lot of bandwidth(QPI bandwidth is said to be same as the memory bandwidth).

 

[Viditor]

Originally posted by: IntelUser2000

Yeah, I thought all Nehalems have IMC, some Nehalems have integrated GPU on die.

Well, here's how it goes. There will be 3 versions.

Bloomfield-3 channel DDR3 IMC with QPI
Lynnfield/Clarksfield(Desktop/Mobile)-2 channel DDR3 IMC without QPI
Havendale/Auburndale(Desktop/Mobile)-Memory controller and GPU on one die connected by QPI to the CPU core

You could probably call Havendale/Auburndale as having an IMC since physically the two parts(the CPU core and GPU/memory controller core) will be very close and will be connected by QPI, which will have a lot of bandwidth(QPI bandwidth is said to be same as the memory bandwidth).

Thanks for the post...that helps!
The only problem I foresee for Havendale/Auburndale is the latency...
At least that's the reason AMD went with direct connects rather than HT connects for their IMC.

 

[Idontcare]

Originally posted by: Viditor
The only problem I foresee for Havendale/Auburndale is the latency...

It would seem like a definite performance bottleneck, but then again I suppose these SKU's are not intended to be performance demons.

No doubt being coupled with Intel's IGP will significantly limit the scope of utlity of the systems...and let's face it if you are only buying dual-core and are willing to operate within the performance confines of Intel's IGP then you aren't likely doing something with your system which higher latency would then become a gating problem.

As with Atom, the quest for "how low can you go" performance and cost wise will apply here too. If you want a $400 laptop that gives you everything you need then this is one way to get there, as is Swift for AMD.

 

[sparkuss]

I don't plan on moving from my current S939 until all the aftermarket parts are available for Nehalem (Coolers/Fans, H20 blocks etc). Anyone have a guess at how long before the aftermarket will be up to speed with the new CPU/MBs?

 

[Dadofamunky]

Originally posted by: ShadoutDH
Personally I think I wont jump on the Nehalem wagon before the next 'tick' (or is it 'tock'? ) with the second revision and 32nm die shrink of Nehalem (and hey, maybe an 8-core! if applications/games every start actually using multicores ^^).

I know I won't go for the first wave, or possibly even the second. The software is far behind-I can't even get 64-bit versions of any of my key apps, much less-4core+ aware. I'm done upgrading for awhile. I might buy an inexpensive Penryn quad in a year or so.

 

[GundamF91]

Easy comparison. It's a no brainer unless you just absolutely need the extra Mhz that Nehalem provides. Build yourself a solid quad core or a fast dual core, and you'll enjoy it all the way to 32nm Westmere or even Sandy Bridge.

Core i7: expensive
DDR3: expensive
X58: expensive

Q9450/6600: cheap
DDR2: cheap
P35/45: cheap

 

[deadcentre]

Originally posted by: Viditor

Originally posted by: jgigz

Originally posted by: BassBomb
Does nehalem have on die memory controller?

Yup.

Correction...some Nehalem will, not all.

The only nehalems I've seen any info on have an on-die memory controller, so which ones won't and where are they on the roadmaps, as far as release date is concerned?

edit: that issue is discussed already: my bad, hadn't seen page two of the thread

 

[Idontcare]

Originally posted by: deadcentre

Originally posted by: Viditor

Originally posted by: jgigz

Originally posted by: BassBomb
Does nehalem have on die memory controller?

Yup.

Correction...some Nehalem will, not all.

The only nehalems I've seen any info on have an on-die memory controller, so which ones won't and where are they on the roadmaps, as far as release date is concerned?

edit: that issue is discussed already: my bad, hadn't seen page two of the thread

Viditor you ignorant slut. How could you not have known way back in June this year that all Nehalems would have IMC?

 

[Wuzup101]

I just built my q6600 system a little over a month ago. I was very happy with the total price (~$1100 before like $125 in rebates)... and that includes a nice new case and an OEM copy of vista 64. It was about half the price of the last system that I built for myself (about 5 years ago).... and the performance is great.

Honestly, there isn't much that the current generation of processors is going to have a problem with. Buying yourself a quad (even a q6600 which is at the bottom of the end as far as quads go) is going to be a good foundation for at least a year or two. Even a fast dual core is going to whip through everything for quite a while (better than a quad in many cases... until developers start using the power of all 4 cores... which may or may not happen in the lifetime of a current quad).

I have no need to upgrade for quite some time. I'll probably wait until the next die shrink and then purchase one of the older processors (again depending on price). Paying for DDR3 right now is just silly when DDR2 is so damn cheap. Even if DDR3 with the new tri channel boards is super snazzy... the price is probably going to be hard to justify for quite some time.

The next upgrade that I buy will probably be an additional HD4850 to run in crossfire with my current card. I probably won't do that until after the next price drop (depending on what they go down to). After that, I may explore the world of H20 cooling... as I'll be able to recycle many parts of a water cooling setup in a new/future rig.