Quick Analysis: Nehalem CPUs & Sockets

B3D News

Beyond3D News
Regular
There remains a significant amount of confusion on the derivatives based on Intel's upcoming Nehalem microarchitecture. A week ago, PC Watch released an article in japanese on upcoming sockets and CPUs, but it didn't collide perfect with previous rumours and many websites reporting on the news seemed, at best, horribly confused.

Read the full news item
 
Is there any similarly clear analysis or background on what the mobile Nehalem platform(s) will look like? I have heard conflicting statements about there being some or not being any mobile IMC parts in that generation.
 
AFAICT, those will be the same chips as in the desktop Nehalem, excluding the Socket 1366 one(s) of course. So you've basically got three possible configs:
- Socket 1a: Quad-Core;IMC;Ultra-High-End Niche.
- Socket 1b: Dual-Core;IMC;GPU MCM.
- Socket 2: Dual-Core;No IMC/MCM.

The hard part is obviously predicting the relative importances of Socket 1 and Socket 2, and I doubt anyone but Intel knows that at this point.
 
Man, I am excited about Nehalem, though I guess it is naive to think that Nehalem will be to Core 2 what Core 2 was to Netburst.

I can still hope, though. :D
 
What happened to the 8 core version? I thought bloomfield was 8 core/16 thread?

And yeah bandwidth is pretty amazing. We're looking at 32GB/s if its using DDR3 1333Mhz with 3 channels.

Bludd: Intel have stated that the jump from Penryn to Nehalem will be greater than the jump from Netburst to Core2 :oops:

Obviously that should be taken with a pinch of salt though!
 
Combining much more bandwidth and lower memory latency with SMT that hopefully works better than with Netburst it is quite possible that performance increase is as big, if not bigger. Of course that would only count for multi threaded applications as SMT won't help with single threaded stuff. I wouldn't be surprised to see SMT giving >25% performance increase.
 
Bludd: Intel have stated that the jump from Penryn to Nehalem will be greater than the jump from Netburst to Core2 :oops:

Obviously that should be taken with a pinch of salt though!
It will be pure porn if that turns out to be true! :D
 
So why do the 1160 chips have an integrated PCI-E controller, while the 1366 doesn't? Isn't integrated better for latency and performance? And wouldn't it make sense to put it in the high-end chips?
The same goes for the north and south-bridge. Wouldn't it be better for performance if they were integrated into one chip as well, just as with 1160?
Or will QPI be so fast that the performance will not be noticable?

Can't wait to see real benchmarks though, especially of the Bloomfield with 3xDDR3! :D
 
LoL @ pure porn.

I'm eyeballing that Socket 1 system myself; thinking I can wait on my current rig until Nehalem comes out. I could do a nice cheap Core 2 Duo right now, but basically everything I do is "acceptable" on my rig. I can even play Crysis at high settings! (Uh, at 640x480 :D )
 
So why do the 1160 chips have an integrated PCI-E controller, while the 1366 doesn't?
1366 might also be aimed at markets where PCI Express doesn't really matter (some servers). Also, the ultra-high-end would want more than 1x16 PCI Express, and SLI is important there. As I said, I doubt NVIDIA will be willing to let SLI run out of Intel's integrated PCI Express.

The same goes for the north and south-bridge. Wouldn't it be better for performance if they were integrated into one chip as well, just as with 1160?
There is no northbridge at all on 1160 - both the memory controller and PCI Express are on the CPU. So it's not really higher integration from the chipset POV.

Or will QPI be so fast that the performance will not be noticable?
QPI to the 1366 northbridge should be more than fast enough, yeah.

Can't wait to see real benchmarks though, especially of the Bloomfield with 3xDDR3! :D
Yup, one thing I'm worried about personally is variability. I mean, 270mm² chip on 45nm? There are rumours of 4GHz+ clock speeds (before OC), but that seems downright crazy to me.

As for performance - I'm sure results will be very impressive in servers, and even more astonishing in 2P and above. Even on desktops, I'm sure multithreaded performance will be very nice, but single-threaded performance is much more questionable. I can't see the IMC delivering substantial boosts on Intel's architecture when not bandwidth limited (which you shouldn't be in single-threaded code). Hmm.
 
Yup, one thing I'm worried about personally is variability. I mean, 270mm² chip on 45nm? There are rumours of 4GHz+ clock speeds (before OC), but that seems downright crazy to me.
I wouldn't call it that crazy. Intel could probably sell Penryn quads at 3.6-4GHz if it really wanted without getting too high power usage. Sure, they are simply two 107mm^2 chips put together but I would think that Intel can tweak its tools before it launches Nehalem.

Though I'm quite sure Intel won't launch at 4GHz as there wouldn't be anything to compete against. I'd be lucky if it launches at over 3.4GHz.
 
Sure, Penryn could be clocked a fair bit higher than 3GHz. My concerns were specific to Nehalem - I wouldn't be surprised if we only saw ~3.33GHz @ $999+ in 2008. Of course, assuming 10-30% higher IPC than Penryn, that still wouldn't be anything to sneeze at.
 
those sockets feel like they're socket 940, 939 and 754.
not sure what's the future of socket 715, it's cheaper maybe but lacks integration that would be useful on the low end and mobile markets
 
Lacking integration isn't perfectly correct - after all, the CPU+GPU MCM means it's a 3-chips/2-packages architecture. Socket 715 with a single-chip Southbridge+Northbridge+GPU (ala what NV has been doing since August 2006 with MCP61, and since September 2007 with MCP73 on Socket 775) is a 2-chips/2-packages architecture.

Also, on Socket 715, you could have a platform without any PCI Express links for a GPU (or just 8x Gen2 instead of 16x), while Socket 1160 will always have 16x PCI Express 2.0 integrated on the CPU.
 
Back
Top