Lynnfield i7-870 CPU Review
Intel’s Lynnfield processors have been
available for several weeks now and extensively reviewed in the online press.
Our aim today is not just to look at the performance of an i7-870 and its
architectural efficiency against a similarly clocked AMD processor but also (and
more importantly) to run tests to predict the performance of forthcoming
processor ranges such as the i5 and i9. We will do this by running tests with 1,
2, 3 and all 4 cores enabled which will give us accurate results for a dual-core
i5 Clarkfield processor and use extrapolation to approximate the performance of
a 6-core i9 Gulftown processor.
By spending a long time duplicating our tests four times we are able to see how various applications perform with differing numbers of cores allowing us to establish the multi-core efficiency of games such as Far Cry 2 and benchmarking tools like 3D Mark Vantage. The testing is by no means comprehensive and if we had 2-3 weeks to spare we could have tested every recent game and application for completeness so our apologies in advance if your favorite application is not included in our representative sampling.
Of more universal interest is comparing the efficiencies of the latest Intel and AMD architecture to compare current and future efficiencies and predict how future trends and architectures (such as the reduction of floating point capability in AMDs new Bulldozer architecture and moves toward CPU/GPU convergence) will affect performance.
First, a recap of the i5/i7 Lynnfield
architecture and layout:
The section on the right is the PCIe controller and represents a sizeable amount of real-estate. The two QPI links that provide high bandwidth communication with other processors and chips on the motherboard are gone as is one of the three memory channels. This means that the socket 1156 platform is not suitable for multiple socket motherboards (no Skulltrail etc.) and these seem to be the preserve of the socket 1366 “bigger brothers”. If you need a multiple CPU system then socket 1366 is the way to go, especially with the forthcoming “Gulftown” 6-core processors launching soon exclusively in the socket 1366 format.
The GPU connects directly to the CPU but those 16 lanes will be divided amongst PCIe cards if you have more than a single GPU installed.
HyperThreading is not supported on the lower
range of Lynnfield processors (i5) but is thankfully present in the i7 series.
It seems to us that for the extra real-estate required (about 5%) HyperThreading
is a bargain given the performance increase in many situations and should be
included as standard on all new processors. Intel, however, have to apply
business strategy to their offering to increase product differentiation and
it’s just not possible to do this on speed alone hence the lack of universal
HyperThreading and forthcoming 2-core processors for the low end and 6-core
processors for the high end.
On the left is the i7-870 and on the right is an
AMD Phenom 2 X4 630 of a similar speed. The 45nm Nehalem core is smaller than its AMD
counterparts or perhaps the AMD heat spreader is just larger on purpose.
Lynnfield is slightly larger than it older Bloomfield predecessor due to the
inclusion of an onboard PCIe controller. We will not be examining this aspect in
detail other than to say that numerous tests have shown a slight benefit for the
Lynnfield processors against equivalent Bloomfield ones with a single PCIe
graphics cards but a slight disadvantage in SLI and CrossFire modes.
One of the features of the socket 1156 (and socket 1366) design is that the pins are on the motherboard socket and not the processor shell. This has the benefit of not risking bent/broken pins during transportation and handling (we testers at The Hardware Review are known for being somewhat ham-fisted with large fingers but have never had any problems handling CPUs over the last 20 years). The drawback of Intel’s design is that extra pressure is placed on motherboard manufacturing – an area where components are usually selected for their low cost. A recent example is the spate of damaged FoxCon sockets due to poor contact with the CPU. It is difficult for motherboard manufacturers to adhere to Intel’s high standards of quality control when they source components from many suppliers. The FoxCon problem has been resolved and all socket 1156 boards using FoxCon sockets now on the market should be using the revised socket (the version number is listed on the back of the PCB so there is no way to know for sure before purchase so it may be worth contacting the manufacturer to allay fears if necessary).
The retail heatsink is not recommended for overclocking, in fact, it’s smaller than the retail Bloomfield socket 1366 cooler so a decent custom cooler will be required for any significant overclocking.
We used a Corsair H50 which gives the benefits of water cooling with the ease of installation of an air-cooled heatsink. Please not that due to a small reservoir on this sealed system it should not be used for extreme overclocking and if the processor temperature gets above 70 degrees Celsius it should be brought back down immediately to prevent water turning to steam and permanently “unsealing” the system.
Without increasing the cpu core voltage we
achieved these results:
3.5GHz and perfectly stable after an hour of 100% cpu usage.
still at stock voltage! Unfortunately not stable during the torture test until we
increased core voltage to 1.35v, which is by no means a big increase and well
within safety parameters.
Since the total absolute maximum power dissipated by a processor needs to be kept below a certain maximum (130W it turns out) that means that all the cores on a die must not total more the processor allows. This figure is 95W for the socket 1156 Lynnfield range.
To achieve this the processor is clever enough to throttle back when things get too hot and to power down inactive cores as demand varies. This is done through a power control unit and saves a significant amount of energy in light use. The opposite is also true and if only one or two cores are being used (as with many current games) the speed can actually be increased by 2 speed categories. This is a free overclock and shows the frequency headroom for forthcoming 2-core Clarkdale processors which should run at higher frequencies due to greater thermal headroom. If you’re a serious overclocker you may want to turn this feature off as any sudden burst of speed may result in instability although in practice, if a processor is running flat out, the thermal envelope will probably stop turbo boost from kicking in. The greatest speed boost is when only one core is active and can be as much as 17%.
Since this review is primarily about multi-core efficiency it worth explaining the inherent problems with multi-tasking. This may surprise some readers as we already have supercomputers made up of thousands of Intel or AMD processors and if they did not scale well then research institutions would not buy them to predict climate change, where minerals are buried and so on. The reason they work so well is that it is easy to split millions of operations among thousands of cores. Splitting one thread across multiple cores is actually quite difficult.
The problem involves concurrency, monitors and semaphores and is too involved to go into here although interested readers are encouraged to read the Wikipedia article on “Dining Philosophers” which explains the whole problem in easy to visualize terms. It can be found here.
Until Quantum Computing is viable we will have to rely on programmers making allowances for multiple cores and programming accordingly. Some games and applications are already optimized to a limited degree for multiple cores and theoretically every application will get a boost with a second core, even if just by offloading the usual Windows background processes to the other unused core.
It has been clear for some years that frequencies cannot continue to increase due to manufacturing limits and have remained roughly constant around the 3GHz mark for about 6 years. Instead it seems that the future gains will be attained by increasing the number of cores in a CPU, whether physical or also virtual (as with HyperThreading). Our test will aim to show which architectures are most suited to getting the best out of extra cores, where the bottlenecks are and, hopefully, give an indication of how the architecture will scale in the future as number of cores increase.
All games are tested at the maximum available settings and initially at 1280x1024 so we can be sure of hitting CPU limitations before bandwidth or fillrate ones related to the GPU. We selected Far Cry 2 (first person shooter), HAWX (air combat) and Resident Evil 5 (horror) for our tests as they are newer titles that are suited to benchmarking and make most systems struggle.
The results show fairly linear scaling as we go
up in cores. It should be noted that synthetic tests such as SiSoft Sandra will
scale quite well and are mainly useful as an indication of bottlenecks and to
see what programmers can achieve if they overcome the hurdles they face.
The processor multimedia results also scale well
although real-life differences will not be as pronounced as this chart
Interestingly, the memory bandwidth results show
that a single core cannot make full use of available capacity and is
particularly the case for the AMD Phenom 2 architecture. Dual core or higher is
required to overcome this limitation.
Despite this test favouring processors with
HyperThreading (i.e. Intel ones) there is a huge difference in performance
between the two architectures. While two cores are fine for the i7-870 here, the
AMD X4 630 needs at least 3 to put in a reasonable showing. Since graphics
performance is similar (same GPU after all) the limitations lie with the
processor. This bodes well for the forthcoming Clarkdale dual-core processors but
it will be necessary to see this repeated in real-world benchmarks to draw any
Far Cry 2 has a very useful built-in benchmarking tool with many configurable parameters.
First thing to note is that this game is playable with 8x AA on any number of cores (fortunately a single-core Phenom 2 does not exist). We will test at varying resolutions later on.
HAWX is a bit of a strange game but provides a
consistent benchmarking function. At 1280x1024 with 8x AA on the highest
settings we can see that a 2-core i7-870 outperforms a Phenom 2 X4 630 with all
4 cores at maximum. There does however appear to be a bottleneck that could be
resolved with more efficient programming.
Two things are noteworthy. There is a bottleneck on the i7-870 performance but its high enough to not be an issue. More importantly it takes the Phenom 2 X4 630 at least 3 cores to match the performance of a single i7-870 core.
Now we have compared differing numbers of cores,
it’s worth showing the performance of the above games with all 4 cores active
but at differing resolutions to show the maximum performance that can be
Suddenly, things are not so bad and both
processors can run at good speeds at all resolutions. If we had not tested with
different numbers of cores we would not be able to tell from the above results
that a 2-core Lynnfield runs this game just as well as a 4-core one and that the
AMD processor needs at least 3 cores to keep up.
Performance is virtually identical across
differing resolutions hiding the issue with a single AMD core. This is a game
that will not tax even basic systems.
Here the AMD Phenom 2 X4 630 outperforms the Intel i7-870 slightly at higher resolutions but hides the previous results showing poor performance with 1 and 2 cores.
We’ve done something not seen in other reviews and looked at the multi-core efficiency of the latest architectures from Intel and AMD (these architectures will change next year and will need re-appraisal) and looked beyond the simple results of just running benchmarks at default (and sometimes overclocked) speeds.
By using the motherboard BIOS to selectively disable cores we can look at the per-core performance which gives us a much greater insight into the architecture’s potential than just interpreting the results from the more traditional benchmarks.
It’s clear that in many cases that a 2-core Lynnfield processor will be perfectly adequate for most consumers and that AMD really need a 3-core processor to get reasonable performance. Of course Intel and AMD are aware of this which is why AMD released their X3 range and why Intel is about to launch it i3 range.
As things stand, this may be the choice for consumers on a budget. The 2-core Clarkdale processors may turn out to be a huge success at the low end – it all depends on the price. Entry level quad core AMD processors are so cheap that there is really no reason to not spend an extra $10 or so and get one instead of a X3 cpu.
For reasons of fairness it is important to mention that the Phenom 2 X4 630 we are using for comparisons sells for about 3 times less than the Intel i7-870 so we are not criticising it directly as it would make for a good budget system and did well (with all cores active) in our tests.
We do however recommend the i7-870 Lynnfield processor (if readers find the price of it outside their budget then the i7-860 is only fractionally slower (2.8GHz instead of 2.93GHz) and around half the price). It has managed to cope admirably with all our tests and we suspect that in dual SLI/Crossfire mode (or even quad GPU configurations) we may find the bottlenecks gone and all four cores being stressed thoroughly.
Another reason we would recommend the i7 870/860 over waiting for the 2-core Clarkdale despite our observations is that, having spoken to some developers, future games are being designed to stress all 4 (or more) cores. Then there are the applications that will always max out multi-core processors such as video editing, media encoding and other specialist segments that will always benefit from greater parallelism and those users will inevitably be drawn to the triple memory support and capacity to take 6-core processors of the socket 1366 platform. But increasingly, especially given the high price of putting together a socket 1366 system, a Lynnfield system will do fine for home users editing DVDs or HD Video of birthdays etc. Then again the Clarkdale may be really cheap and/or have a very fast stock speed….
As a general rule though, if you are looking to purchase a high end processor then the i7-870/860 is it assuming that a socket 1366 solution (with its potential to upgrade to 6-core) is out of budget. There is no planned path for 6-core processors on the socket 1156 platform but our results show no benefit for 6-cores over 4-cores with current applications except for the specialist segments mentioned above.
It would appear to be a decisive victory for the Lynnfield architecture but the AMD “Bulldozer” is expected soon and it will be interesting to see how things develop over the next few months.
All trademarks are the property of their respective owners.