Changing for the Future - Intel's Core i7

Nehalem Becomes the Core i7

Ever since Intel adopted its relentless Tick-Tock strategy of alternating between a new CPU microarchitecture and a process shrink of the existing architecture, the company has managed to keep up with both market expectations and the undeniable pressures of Moore's Law. But while all 'Ticks' and 'Tocks' are equally important, the latest development is especially so. For this particular 'Tock' sees the introduction of a new microarchitecture that is significantly different and like all architectural shifts, it's a move that's fraught with risk.

This much awaited development is the introduction of the Nehalem microarchitecture, debuting in the market in the form of its desktop variant, Bloomfield. Intel has also gone with a new name, Core i7 to distinguish it from the current Core 2 series of processors. Three quad-core models are available as we speak (Core i7-965 Extreme Edition, Core i7-940, Core i7-920), ranging from 2.66GHz to 3.2GHz and are targeted at power-users and enthusiasts, to be used together with the Core i7 compatible Intel X58 Express chipset, which incidentally also uses a new socket - LGA1366.

What the Nehalem represents, is a shift to a NUMA architecture (Non-Uniform Memory Access) that Intel's rival, AMD has been using ever since the debut of its Opteron processors. This is a strong sign from Intel that its current architecture is insufficient - due to the bottlenecks posed by the limited bus and memory bandwidth on the Core 2 platform - to handle the growing number of cores that will be expected in the near future. To read more about the details of the Nehalem microarchitecture, please refer to our previous .

So what's really new about the Core i7 when compared to the Core 2? We list down some of the more important features that will be found on the new microarchitecture:-

QuickPath Interconnect and Integrated Memory Controller

Analogous to AMD's HyperTransport bus, Intel has come up with QuickPath Interconnect (QPI) as the high speed link between processors, memory and other internal communications (e.g I/O). To complement this, a memory controller is integrated to the processor, similar to what's found on the AMD platform. In short, say goodbye to the traditional Northbridge finally. QPI is designed to be scalable, with the number of QPI links set at two per CPU socket currently but this may increase depending on requirements.

This scalability can be seen in the newly launched Core i7 processors, with the top model, the Core i7-965 Extreme Edition having a QPI speed of 6.4 Gigatransfers/sec (up to 25.6Gb/s) while the other two models are limited to 4.8 Gigatransfers/sec.

The new integrated memory controller supports DDR3 memory of up to DDR3-1333 in a tri-channel configuration, immediately increasing the total memory bus width from the existing dual channels. Meanwhile, each channel can support up to 3 DIMMs maximum, giving a total of 9 memory slots for each processor. This also means that you'll find Core i7 motherboards to come in multiples of three for DIMM slots. As you may realize by now, the total amount of memory supported and its resultant bandwidth should become a non-issue with this new architecture.

The new memory configuration is also quite flexible. For each individual memory channel, you can have varying memory sizes. For example, one channel could have only 1GB of memory installed while the other channels could have 2GB. The catch is that within each channel (which you will remember can consist of up to three DIMM slots), the memory modules installed must be of the same size, e.g. 3 DIMMs of 1GB. To sum up, you can mix and match your memory sizes from channel to channel but within each channel, the installed DIMMs must be similar in size.

Memory Woes

A critical limitation that we found with our Core i7 review samples was that Intel states that the processor can only work with DDR3 memory rated up to 1.6V (1.65V if you're pushing it) on its reference X58 Express chipset based motherboard (this limit is in-line with the maximum recommended CPU core voltage). Any higher voltage memory may damage the processor, even if the memory is operating within its safe limits. Obviously, this means that almost all high-speed DDR3 memory kits in the market now are not supported officially.

Memory vendors have already started to address this issue by releasing newer triple-channel memory kits while some motherboard manufacturers have confirmed that their retail boards are capable of stretching this limitation somewhat, with 1.7V quoted as the upper limit on these boards at the moment. There are also other notices going around that some board manufacturers can successfully run the memory voltage completely independent of the maximum recommended for the CPU. So for the moment, it's wiser to stick with the base recommendations and wait for a couple of months when the Core i7 scene gets more prevalent and more information and findings are available.

Another minor oddity is that while the Core i7 states that it officially supports up to DDR3-1066 memory, most if not all other board vendors have much higher support, even up to DDR3-2000. Obviously since this a CPU controlled variable, Intel's official support probably means that those are the modules that would surely run within its recommended 1.6V specifications. We've successfully run the memory at higher than 1600MHz and it works swell.

Tri-Cache Level Structure

Not only has Intel gone with a very AMD-like integrated memory controller and NUMA architecture, even the new cache hierarchy on the Core i7 similar to AMD's Barcelona. Again, this move was obvious, since the addition of the memory controller meant that Intel did not need large L1 and L2 caches to bolster its Core 2 processors. Instead, a massive 8MB of L3 cache is found on the Core i7, along with a new second-level TLB to improve virtual address translations.

Talking about the new L3 cache, it is actually not recognized by Intel as part of the new 'core'. Intel has classified it under the (strangely named, if logical) 'uncore' portion of the processor. This conforms to the scalable nature of the Core i7, with the QPI, memory controller all other examples of the basic building blocks that would go into the design of a Nehalem processor and considered the 'uncore'. Intel has even hinted at adding a GPU as another such building block in a direct counter to AMD's as yet unreleased Fusion.

The Return of HyperThreading

What you should know is that beneath all the additions, the basic quad-core at the heart of the Nehalem is still mostly the same Core 2 processor. Of course, there have been some enhancements to the internal algorithms and branch prediction capabilities, but the main story here is the return of an 'old' feature from the past. Yes, the notorious HyperThreading of the Pentium 4 era is back.

Now known as Simultaneous Multi-threading (SMT), Intel claims that it has been enhanced but it is a very similar idea of having each processor core fed more than one thread simultaneously. As such, the quad-core Core i7 will be seen as 8 logical cores by the OS. With multithreaded applications more prevalent than the Pentium 4 days, and now with the staggering memory bandwidth found on the Nehalem, Intel may have a point about HyperThreading Redux having some utility this time around, though like before, it is heavily reliant on the threaded nature of the applications, which are still few and far in the mainstream area. Workstation and Server specific workloads are far more threaded and would yield a good deal more from HyperThreading.

Turbo Mode!

Finally, Intel has a Turbo Mode option for the Core i7, where the user can define the clock speeds for each core within the quad-core Core i7 individually (in the form of adjusting the clock ratio). So if your application is only using up to two cores, the Core i7 can allow for higher clock speeds for those two cores in use. It's a compromise made so that users can have the best of both worlds - better performance with higher clocks when fewer cores are used while retaining the capability of more cores when the applications needs it (scaling down the clock speeds to default or a less aggressive turbo level with all cores engaged).

** Updated on 10th December 2008 **

While the Turbo Mode option is available on all Core i7 models, only the Extreme Edition model has the core ratios completely unlocked and one can manually increase and decrease the default Turbo mode multipliers at their willing. Additionally, the Extreme Edition has no strict TDP limitation which would have otherwise throttled the CPU performance when it hits a thermal envelope (which is applicable for the other two Core i7 models).

Now that you know the important features of the Core i7, here's how it stacks up against current quad-core 'equivalents' from AMD and Intel.

Intel's X58 Express Chipset

With such a radical change as integrating the memory controller onto the CPU, a completely new chipset and a new socket, LGA1366 is required for the Core i7. The answer from Intel is the X58 Express chipset. The first thing you may notice on any X58 board is the odd number of DIMM slots, reflecting its triple channel memory support. The traditional Northbridge has been removed since there's no need for it anymore with QPI and the integrated memory controller but there is still another hub, the IOH, which primarily provides the 32 PCIe 2.0 lanes for graphics cards. The Southbridge remains and it is the now familiar ICH10R, which will support up to 6 SATA 3.0Gbps ports and 12 USB 2.0 ports as a start.

The chipset supports up to two 16-lane PCIe graphics slots (ver 2.0) and communicates with the processor via the QPI bus (routed through the IOH). Retail versions are likely to come with more than two PCIe 2.0 x 16 slots and besides the native CrossFireX support, third party vendors can also get their boards certified to work with SLI configurations, making the X58 the only widely available chipset to support both competing multi-graphics technology.

While the chipset may not seem unfamiliar to enthusiasts, the front side bus (FSB) as we know it is no more. A base clock of 133MHz is now the underlying clock that is followed by all the other components like memory and CPU, after applying a suitable multiplier to it of course. Generally, there are 4 such multipliers, though you may not get the option to tweak all 4 inside the BIOS.

First, the CPU speed is defined by multiplying the base 133MH to a clock ratio. This can be applied to each of the active CPU cores on the Core i7, e.g. 4 different ratios for a quad-core Core i7. So far, the unlocked Extreme Edition i7-965 that we have has a range of between 20 - 35 for the multiplier, giving it a resultant CPU frequency of 2.66 to 4.66GHz. This is hence the most crucial multiplier for enthusiasts.

The memory speed is also governed by its own memory multiplier, which is similarly derived from the base clock and this multiplier. Next, one can adjust the QPI rate and the uncore, which the latter limited with a distinct relationship of twice the memory multiplier value.

Overclocking the Core i7

So how does one go about changing the CPU frequency now that the familiar FSB is no more? Well it is actually much of the same and the user has a choice of either increasing the CPU multiplier or the base clock frequency (default of 133MHz) or both. Also, remember that the new Turbo Mode option gives users the flexibility of setting upper limits for the CPU multipliers for a 1-core, 2-core, 3-core setup, up to the maximum number of active CPU cores available. So if your applications are mostly dual threaded, you can opt to have a more aggressive 2-core CPU multiplier for a higher clock speed and more modest clocks for the 4-core configuration. Unfortunately the full Turbo mode goodness is restricted to the Extreme Edition while a more restricted version is available on the Core i7-940 and i7-920 models where one can't set values higher than two speed bins (each speed bin just means a 1x increase in CPU multiplier).

After this, it's a matter of trying various clock ratios and testing the system for stability. Getting the right version of CPU-Z is also useful in knowing the success of the overclock. In our example below, we're running it at the stock settings.

Of course, one can also increase the base 133MHz clock. This is done by increasing the 'Host Clock Frequency Override' setting in the reference Intel BIOS, though it could be labeled differently on retail boards. Since this base clock is used for memory clocks too, one should be careful to adjust the memory multiplier to ensure that the memory is not getting overclocked inadvertently.

As for tuning the memory, let us first repeat Intel's warning that memory exceeding 1.6V could damage the Core i7. As for the memory timings, it's quite straightforward and similar to that on the Core 2, with Intel's XMP memory profiles perhaps the easiest way to get the approved 'overclocked' settings for a particular memory kit. Again, the base 133MHz clock is key, so you can increase the multiplier slowly to get a higher resultant memory frequency or just increase the base clock. There may also be a need to increase the QPI voltage here since the UCLK multiplier may be increased (due to the requirement that it be 2x the memory multiplier).

For actual overclocking performance expectations, we'll be showing you that quite soon in a follow-up article.

Test Setup

Our review kit from Intel consisted of an Intel Core i7-965 Extreme Edition (3.2GHz) processor together with the reference Intel DX58SO motherboard. The following are the other items used in the test configuration:-

• Intel DX58SO Motherboard (Intel X58 Express chipset)
• 3 x 1GB Kingston HyperX DDR3-1333 (7-7-7-20) @1.62V
• ASUS GeForce 8800 GT 512MB (ForceWare 169.21)
• Seagate 7200.10 200GB SATA hard drive (one single NTFS partition)
• Windows XP Professional with Service Pack 2
• Intel INF 9.1.0.1007 and Matrix Storage Manager 8.6.0.1007

To examine how the revived HyperThreading (SMT) performed, we tested the Core i7-965 with and without this feature enabled. We also tweaked the amount of memory used to 2GB of DDR3-1333 at the same memory timings so that it will be comparable in terms of memory size to our older test results for the Intel Core 2 Extreme CPUs and Phenom CPUs that we will be comparing the Core i7 against. This last set also examines if the additional memory controller actually has any beneficial effect to overall performance. Hence, there will be three sets of results involving the Core i7-965 Extreme Edition.

One note regarding the memory used is that we would have loved to have run them at DDR3-1600 which would have put in direct contention with the QX9770 we've last tested. However that requires a lot more voltage than currently recommended and thus we stuck with DDR3-1333 instead.

For the Intel Core 2 Extreme (QX9770 only), the setup is as follows:-

• Gigabyte X48T-DQ6 motherboard (Intel X48 Express chipset)
• 2 x 1GB Patriot DDR3-1866@ DDR3-1600 (7-7-7-20)
• Seagate 7200.10 200GB SATA hard drive
• ASUS GeForce 8800 GT 512MB (ForceWare 169.21)
• Windows XP Professional with Service Pack 2
• Intel INF 8.3.1.1009 and Matrix Storage Manager 7.8.0.1012

For the Intel Core 2 Extreme QX9650 and Core 2 Duo E8500, the setup is:-

• ASUS P5E3 Deluxe (Intel X38 Express chipset)
• 2 x 1GB Kingston HyperX DDR3-1333 (CAS 7. 7-7-20)
• Seagate Barracuda 7200.10 200GB SATA hard drive (one single NTFS partition)
• ASUS GeForce 8800 GT 512MB - with NVIDIA ForceWare 169.21
• Microsoft Windows XP Professional with Service Pack 2
• Intel INF 8.3.1.1009 and Matrix Storage Manager 7.8.0.1012

For the AMD Phenom X4 9950, the setup is as follows:-

• ASUS M3A32-MVP Deluxe Wi-Fi (AMD 790FX chipset)
• 2 x 1GB Aeneon DDR2-1066@ DDR2-800 (CAS 5-5-5-15)
• Seagate Barracuda 7200.10 200GB SATA hard drive (one single NTFS partition)
• ASUS GeForce 8800 GT 512MB (ForceWare 169.21)
• Microsoft Windows XP Professional with Service Pack 2

Benchmarks

The following benchmarks were used in this review:-

• SYSmark 2007 Preview (Patch 3)
• SPEC CPU2000 v1.3
• Futuremark PCMark 2005 Pro
• Lightwave 3D 7.5
• 3ds Max8 (SP2)
• Cinebench 10
• XMpeg 5.0.3 (DivX 6.8 encoding)
• Futuremark 3DMark06 v1.1
• AquaMark3
• World in Conflict v1.05
• Crysis v1.1

Results - SPEC CPU2000 v1.3

HyperThreading or SMT as Intel now calls it, can be a mixed blessing for the Core i7. It looked to be heavily dependent on the workload, with it pulling back the Core i7 at the 4-user test scenario, though it turned out to be very useful when 8 users were chosen. In fact for SPECint_rate, having HT enabled produces higher scores than without, even with 1GB less memory. The catch of course is to have the 8 users test scenario and not 4 users. As we have mentioned in the past, the performance of HT is dependent on the application/OS picking the right core to send the workload (two different logical cores as seen by the OS may in fact belong to the same physical core). As such, for normal everyday usage, HT is really not necessary and in fact might be a hindrance as seen from our results.

Also notable from our testing is that the extra memory controller and extra 1GB memory made little extra impact in SPEC CPU2000 integer rate tests. However with the more aggressive floating-point workloads, the extra memory throughput gave rise to nearly 20% improvement in performance.

Looking at the peak scores from the speed tests, the Core i7 was barely ahead of the QX9770 for the integer segment, though the floating point test showed the strength of the new architecture. The same can be said from the rate tests as well. Overall, depending on workload types, the Core i7 can be anywhere 20% to 50% speedier than the similarly clocked QX9770 - which is quite a big leap. Even when discounting the effects of HyperThreading, the gains on the floating-point workloads can be up to 40%.

Results - SYSmark 2007 Preview

SYSmark 2007 will give more ammunition to those who believe that HyperThreading is a double edged sword because once we disabled it, the Core i7 scores were higher. Looking at the breakdown, 3D and video creation benefited most from having SMT enabled but this was countered with a drastic drop in scores for the productivity segment.

Having more memory too had a slight improvement for scores overall though the most significant difference between the new architecture and the older Core 2 Extreme was in the 3D segment, which saw massive gains, especially with SMT.

Results - Futuremark PCMark05 Pro

There wasn't any dramatic improvements in PCMark05's CPU score for the Core i7 as the 3.2GHz Core i7-965 Extreme Edition had only marginally higher numbers than the Core 2 Extreme QX9770. There was also negligible difference between the scores for the three Core i7-965 configurations. In the memory segment however, the new architecture flexed its muscular bandwidth with a staggering 35% gain over the QX9770. Here, there was a slight dip in performance once HyperThreading was disabled, but it was still significantly faster than the previous platforms using Core 2, which directly relates the advantages of the integrated memory controller.

Results - Lightwave 3D 7.5

The Core i7 was generally faster than its closest rival, though there were occasions when HT again seemed to be the culprit affecting its performance. For instance in the Tracer-Radiosity test, turning off HT yielded slightly better results for all scenarios, while having it enabled meant that the QX9770 was actually doing better for 4 and 8 threads, even with its 1GB RAM disadvantage.

Results - 3ds Max8 (SP2)

Once again, the 'unpredictability' of Core i7's SMT is in the limelight here. With 2 different test scenarios in 3ds Max 8, having SMT appeared to be a liability for Radiosity while the reverse was true for Light Tracer. The amount of memory (3GB vs 2GB) did not seem to have any effect on the final results for the Core i7 and in both tests, with or without SMT enabled, the Core i7 was ahead of the Core 2 Extreme. We also had a rather anomalous result in Radiosity, with the Core 2 Duo E8500 outperforming every other CPU.

Results - Cinebench 10 and XMpeg 5.0.3 (DivX 6.8 encoding)

With Cinebench 10 capable of using multiple cores and threads to improve its scores, the SMT enabled Core i7 was primed to top the charts. Simulating 8 cores, the Core i7 took a 24% lead over the QX9770. However, when it came to a more realistic application like DivX encoding using XMpeg, the Core i7 was just about equal with the QX9770, with HyperThreading again having a say on the results.

Results - Futuremark 3DMark06

The CPU segment in 3DMark06 benefited significantly from having SMT enabled but in terms of the overall score, there was hardly any difference between the new Core i7 and the Core 2 Extreme QX9770.

Results - AquaMark3

CPU scores were as expected higher in AquaMark3 for the Core i7, with this lead extending into the overall FPS scores.

Results - Crysis and World in Conflict

Onto our gaming benchmarks, where HyperThreading had a slight impact in World in Conflict, though Crysis was unaffected. There was also not much of a boost in frame rates, rather Crysis saw a slight drop for the Core i7, putting it on par with the dual-core E8500. This was despite the fact that Crysis was able to take advantage of 4 or more threads but we didn't see any of that happening. We'll be trying these tests again shortly with a more powerful graphics card than the GeForce 8800 GT we're using here, but we doubt we'll see a lot of difference because the resolutions and settings used are quite tame and shouldn't bottleneck the GPU that soon. Stay Tuned for an update on this.

Power Consumption

Before we talk about the power consumption results, we have to say that we were unable to get the Core i7 and the X58 motherboard to enter the idle state. We didn't notice any frequency or power throttling when idling in desktop like the other Core 2 Extreme CPUs. As a result, the 'idle' results returned by the Core i7 were all higher than the other Intel processors, though the AMD Phenom still took the unwanted tag of being the most power hungry CPU.

When running at peak in both 3DMark06 and SPECviewperf 10, the Core i7 consumed marginally less power than the QX9770, though turning off SMT had rather contradictory effects on the power consumption numbers. Again, it does look like SMT has quite the impact, even here for the power consumption testing and it depends very much on the actual workload, with the SPECviewperf test probably utilizing the SMT more fully.

Overall, it certainly looks like the Core i7 has higher performance potential at similar or less power draw than the previous Core 2 architecture and platform.

Conclusion

The Nehalem architecture is one for the future. With its heavy slant towards multi-core and multi-threaded applications, expanded and dedicated memory bandwidth for the processors and the re-introduction of HyperThreading, the Core i7 launching now may be meant for the high-end desktop but its true intentions lie with the server front.

While Intel may still retain its performance and thermal envelope advantage over AMD, there's no doubt that the company has recognized that its present course needs to be adjusted, particularly for its server class chips. Adding increasing amount of L2 cache is after all not a very scalable nor inexpensive move. Hence the shift to an architecture that's similar in a sense to AMD while designing it with modularity in mind that gives Intel the option of customizing the CPU to fit market requirements, with an integrated GPU a very feasible addition in the future.

Of course, Intel has kept with the superior basic internal core design of the Core 2 processor, which then makes the Core i7 a Core 2 remade with some drastic, new trappings. For sure, we welcome QPI and the integrated memory controller, but not all of those design decisions play off, at least not from the applications and benchmarks that we tested. Some of this could be due to the inherent nature of these applications. Others could be the limitations of the enhanced 'HyperThreading' or Simultaneous Multi-threading that Intel has brought back. We are somewhat ambivalent about the use of SMT at this moment, since its performance can vary quite significantly depending on your workload. However, it's no doubt a technology that will only see more mileage as applications get even more multi-threaded. In our opinion, most enthusiasts are better off disabling HT to maximize the gains we noted in more general enthusiast oriented workloads, while specific highly threaded applications will no doubt reap the benefits of HT that maximized unused core resources.

Although Intel has said that the Core i7 is meant for digital media and gaming enthusiasts, it would involve quite the switch for some of these intended users. Firstly, the cost of switching over to the Core i7 platform would require a substantial investment. The Core i7-965 Extreme Edition at just under US$1000 may be relatively cheap compared to the Core 2 Extreme QX9770 (~US$1500) but the added cost of changing the entire platform (LGA1366 compatible motherboard and suitable 1.6V DDR3 memory) will give users some pause.

Fortunately, the Core i7-940 is priced at roughly half the price of the 965 (US$562), with the 920 model even lower still at US$284, and should be a relatively affordable alternative to getting a high-end Core 2 Extreme, even after factoring in the platform cost. While gamers hoping to get the fastest CPU on the planet may not find much benefits from upgrading their processor to a Core i7 as much as getting the latest high-end graphics card, those involved in digital media creation and editing may see a reason to upgrade.

Considering the general enthusiast point of view, we're not entirely sold on it yet, but with an eye towards the future, it's probably the best single modular architecture Intel could come up with that would suit the home, workstation and server segments needs now and in the near future. Clearly some of the workloads bring in phenomenal gains and when used with the right applications now or when applications get even better threaded, users can expect to benefit from the Core i7. For now, if Intel wants the hardcore gamers to move with the Core i7, they really should work a lot closer with the game engine developers to find new ways to leverage more out of the processor.

Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.