Intel Core i7 - Nehalem Architecture Dive

How well will this new architecture take for Intel?

In one way Intel is crudely pushing this as "Penryn, but wider" without considering the specific pipeline alterations it smells a little of Netburst again. Why?

Because instead of saying "as long as programs are written to stream efficiently through the long pipeline..." we have "as long as programs are highly threaded..." and to date we know that this is still rarely the case for consumer applications. There will be the inevitable arguments on forums globally about what applications use four cores efficiently, let alone eight threads if we include SMT!

Well, the truth is Thread Level Parallelism is slowly getting better as developers realise that to improve performance they have to multi-thread where possible. And, depending how awesome Intel's Hyper Threading can unzip some Instruction Level Parallelism out of existing software, Nehalem/Core i7 could provide some immediate performance improvements.

We also need to understand that our computing experience will only continue get more complicated and having available resources to do more things concurrently will be ever more important - even if they are eight single-threaded applications. You only have to look at the direction of applications like Google Chrome, which opens a new instance of the app for each browser tab, to see where we're going. If you can, turn off all your cores apart from one in the BIOS and see how long you last before you fold over in frustration.

For workstation users the extra "width" will certainly benefit as there are now plenty of threaded video compression tools, rendering applications and other software for serious business' out there. If you're into virtualisation the "more at once" argument will become particularly important, in addition to the large L3 cache, fast QPI and huge memory bandwidth and size available.


The potential for mobile in the future through power saving, space saving and “performance per watt,” through to multi-socket server thanks to a complete architecture change and wider execution could be exactly what Intel is aiming for, as both these market areas are growing nicely (unlike desktop PCs). For home and power users though, Intel is relying on a lot of other important factors for its latest architecture to be successful, unlike Conroe and Penryn that were simply awesome out the box.

For gamers, we don't think there will likely be an immediate benefit because the execution cores haven't changed, and what about the latency for data from the PCI-Express bus to memory that has to now navigate through the CPU? From the few games we've run there isn't in a clock to clock evaluation, there isn't an evolution in performance - some are slightly faster, but others are actually slightly slower or show sporadic performance that's difficult to quantify.

The huge memory bandwidth and efficient memory access will make some benefit but it depends how the game engine is written and very little use more than a few threads. As we found in the Wolfdale review we did earlier this year, a very high clocked dual core still outperforms a fast quad core in gaming alone. However, luckily for Intel, most gamers are also power users and prefer to balance gaming with the "more things concurrently" argument, and so went out and bought more Q6600 G0 stepping processors than you can shake a foundry at.

The hardest call will be for general power users and overclockers contemplating the upgrade from a quad core Penryn to Core i7 - will you see a benefit? It entirely depends on the applications used, the level of multi-tasking you do (larger monitors and ever cheaper memory will continue to grow this), but we suspect unless you're very heavily multi-tasking and finding your quad-core is limiting an application you use, then an expensive total platform upgrade is an unnecessary luxury right now.

Because of the multiple socket differences, we feel obliged to compare this launch to the AMD Athlon FX on socket 940. Before AMD had launched its great value performance-mainstream socket 939, we had the choice of the socket 940 with just one or two very expensive CPUs, the much more basic socket 754, or the older socket 462 Athlon XP. Socket 940 ended up an extremely niche and expensive investment for a limited time so what makes Intel's LGA1366 different?

Well, as far as we know Intel is committed to the socket for the foreseeable future, there are multiple CPUs available from launch with a few more planned, but don't expect the wealth of choice currently available for LGA775. We don't know how well Lynnfield will perform in comparison to both LGA775 and LGA1366, but while the PCI-Express-CPU-Memory transfer will be more efficient there will be less memory bandwidth to play with. How important either of these factors are in Nehalem - compared to Penryn - will provide a key to the answer of how important the Lynnfield launch will be and whether it will be worth waiting a long time until Q309 for an overhaul upgrade.

Both LGA1366 and Socket 940 used expensive, niche memory that's specific to its platform, but the advantage Intel's new setup has is that it will work with most DDR3 already available - so if you already own a dual channel kit that will see you through for the time being. AMD's socket 940 on the other hand would only work with ECC DDR memory and there was no performance market for this.

SLI and CrossFire on a single platform will also make it somewhat more compelling than current LGA775 motherboards, especially since there's more intense competition between Nvidia and ATI right now; however that is on the presumption you care about multi-GPU of course.

All in all, the new Core i7 platform has more going for it than AMD ever did with its socket 940 thanks to greater industry support from enthusiast memory and multi-GPUs, however even the cheapest Core i7 920 processor isn't cheap if we consider what we're paying for E8400s, E5200s or the super popular Q6600s. We can only cry at the loss of extreme-value overclocking - many of the most successful overclockers in the world started out as teenagers and students learning the ropes by killing very cheap CPUs (an AMD Duron 750 for £35 here), so what will this mean for future generations? Only if you have money to play and "burn," can you be a big hit?

With Nvidia pushing its CUDA GPGPU "mass multi-core" as the way forward - such examples are its Badaboom, Folding@Home and the hugely important Adobe Creative Suite 4, a few of us doubt that Nehalem can be such a firm hit its predecessors were in the desktop market. However, then we remember it's not all about the GFLOPS - it's about ease of use with the millions of x86(-64) programs already available. Intel's competitor will continue to be AMD, not Nvidia, and we encourage digesting our performance analysis article in detail as well as keeping an eye on how well the new Core i7 920 CPUs overclock (and how the soon-to-be-released AMD Deneb CPUs perform) to see if your current setup should be retired or not. The CPU as we know it today will stick around and be very important for a while yet.