A week ago we gave you a short preview and some basics about the upcoming Core i7 series CPUs, here. Today we will show you our benchmark results and find out if this CPU is worth spending your hard earned money on.
Intel iOH X58/iCH10Ri
6-phase analog VRM
ON Semiconductor ADP4000JCPZ 6-phase VRD11.1 controller
ti TSB43AB22A 1394a Firewire controller
Intel 82567 PCIe Gb network
Marvel 88SE6121 dual-port (e)SATA controller
Winbond WPCD3761AUFG Super I/O controller
passive cooling of chipsets/VRM
16Mb BIOS, version 2624
2x PCIe 2.0 x16
1x PCIe x4
2x PCIe x1
4x Triple-Channel DDR3-slots for PC3-8500U memory up to 8GB (only three work for triple channel)
6-Port SATA II featuring RAID 0, 1, 0+1, 5
1x Gb LAN
8x USB 2.0
7.1 analog output
1x SP/DIF optical out
6x SATA cable with clips
1x 4cm LED fan
Beta BIOS with OC functions, final features not set for sure yet.
Intel DX58SO "Smackover" (provided by Intel)
Intel Core i7 920 (provided by Intel)
Intel Core i7 965XE (provided by Intel)
Thermalright Ultra 120 eXtreme 1366 (provided by Thermalright)
Qimonda 3GB Kit PC3-8500U (provided by Qimonda)
Kingston 3GB Kit PC3-10600U KHX1600D3K3/2GX (provided by Kingston)
CL9-9-9-27 CR1T @1.65V
Jetway Radeon HD3870 (provided by mec-electronics)
Seasonic S12-II 500W
Samsung Spinpoint F1 500GB
SilenX iXtrema Pro 14dB(A) (provided by PC-Cooling.at)
Cooler Master Stacker 831 Lite (provided by Cooler Master)
The Intel DX58SO "Smackover" Mainboard
As Intel has now followed AMDs lead, including integrating a memory-controller into the CPU and doing a native four-core, the board resembles many Athlon boards. It's quite odd that the memory slots are beside the CPU instead of underneath. The first impression was quite disappointing, because this is a €300 board and it's not an all solid capacitor design. You also notice the Northbridge cooler, now called hub, features a cheapo aluminium cooler which will exceed 60°C when under heavy load and overclocked. Intel did indeed include a ridiculous 4cm fan which does lighten up your case inside, but also annoys you with unbearable noise. We hope they won't include such a toy in the retail product, otherwise just forget about it and buy something better.
(Click here for a bigger picture)
The VRM is a six-phase analog design, which due to the high current the i7 needs, should be mandatory for any board you consider buying. Of course, we would have liked to see a digital solution, but it did fine. For reasons beyond our imagination they still stuck with the ICH10R Southbridge, which could have been easily integrated into the new Northbridge, but we guess they have a lot of them in stock and we will have to wait a bit longer to see a true single chip chipset from Intel.
Two PCIe 2.0 x16 slots gives you Crossfire compatibility but this board is not SLI certified, because Intel has no reason to pay nVidia. Luckily, fanboys will get plenty of boards with this feature, of course with a price premium. An additional PCIe x4 slot and PCI slot limits your flexibility for additional cards when running this board with two dual-slot graphic-cards, otherwise it is enough.
Intel finally managed to put a tiny power-on button on the board, but "forgot" to addreset and CLR-cmos buttons. At least, in the majority of cases the board was able to boot up even when overclocked too high or setting the memory to too tight latencies.
While we hoped Intel would get rid of the push-pins which are a pain in the you-know-what, they managed to disappoint us again. The holes are now even wider and this will result in some problems to put enough pressure to the CPU-cooler. The funny thing is, the Northbridge cooler features a bracket and screws were used.
Overall, it's an okay board, but we expected more. For 965 users it will be better, because there are limits while overclocking the 920.
Intel Core i7 CPUs:
We have two CPUs, the 920 and 965XE. The biggest difference besides the free overclockable Turbo-Mode of the 965XE is the L3-cache multiplicator. While the 965XE has a multiplicator of 20, resulting in 2.66GHz, the 920 has only 16, resulting in 2.13GHz, quite a huge difference. The BIOS let us change the multiplicator for the 965XE but didn't with the 920. So, when overclocking consider you can't lower the L3 cache multiplicator.
Both CPU-z screenshots were taken with Turbo-Mode disabled. Turbo Mode with a 965XE will result in 3.46GHz under load. CPUz reports slightly reduced numbers when the CPU is idling but energy savings are turned off:
When idling the multiplicator is reduced to 12:
As already mentioned in our preview overclocking the new CPUs is a mess. Intel managed to put all settings in three different BIOS screens and on top of it, the BIOS does not calculate any frequencies for you; we feel like we are back in the stone ages when we were happy to overclock a P90 to 120MHz.
Nevertheless overclocking the 965XE is quite simple, because you do it only with the Turbo-Mode. We adjusted the VCore +275mV with the dynamic overvoltage feature, which avoids increasing the stock VCore and set the Turbo multiplicator to 30, voila 4GHz. If you increase the stock voltage it will increase power consumption in idle mode, too.
We tried even further, but with a multiplicator of 32 the board did boot up, but wasn't stable. As you can see we did decrease the QPI link frequency, which did help with overclocking, otherwise the board would not have booted up at all.
With the 920 it's quite more difficult. While you have a Turbo-Mode which increases the multiplier by 2, you can't set it freely. So you have to set the virtual host clock - in case you forgot, we don't have FSB anymore. For some reason our board wouldn't like frequencies over 166 and it did not matter if we enabled or disabled the Turbo-Mode. Furthermore, now the BIOS has settings for maximum current and TDP. We can't change them, but you will on non-Intel boards, for sure. So, our results were limited to 166MHz. Also, the Turbo-Mode kicked in only up to 21, because the TDP maximum - which we couldn't change - was hit. At least we got nearly 3.5Ghz without adding any VCore, which is quite impressive. To have only two multiplicators for memory, six and eight, is not a problem at all, even if you can hit a virtual host clock of 200, with 8 you get 1600MHz, which is the highest speed with affordable prices yet.
Of course, we resorted to our standard benchmarks, first which is lame. While benching we noticed the Hyperthreading of the CPU does increase the performance, so the CPU does shift loads on it's own, even if you run only single-threaded applications. On some applications this will backfire and will be slower compared to Hyperthreading disabled.
Because lame is a small program which should fit at least in the L3 cache, we checked out if triple or dual interface does make a difference. Also we reduced the QPI Link frequency to 4.8Mbps. Surprisingly, a dual channel interface with reduced QPI Link comes close or does overtake the triple channel high bandwidth QPI settings. SMT stands for Hyperthreading enabled.
Just to check if there is any real difference, we used the Everest benchmark to confirm our suspicion that it does not matter much:
While benching x264 we noticed enabling Hyperthreading does hurt the results considerably. So this is the only benchmark where we have disabled it by default. Because we need comparable results we do use a rather old binary, we will have a follow up article concentrating on gaming where we will check out, if this is still a problem or just related to this old build. For some reason the x264 bench will crash the QX9650 4GHz test-setup, just believe us, when we tell you, it's much faster.
Of course, we also checked if triple or dual channel memory interface makes a difference:
Because x264 is also very demanding on the memory interface, we thought it's the best benchmark suited to point out the differences in memory speed. We did get the first 3GB Kingston kit just at the end of last week and would have liked to bench more, but we were not able to achieve low latencies with lower speeds. This kit is aimed for high speeds, not low latencies.
As you can see, there is only a 1 FPS difference between 1066MHz and 1600MHz, so it's better for your purse to buy the 1066MHz modules, as they will get cheaper because demand from OEMs will increase. To invest in high-speed memory modules is just a waste of money, in our opinion.
There is not much to say about these benches; you can clearly see that rendering does benefit from Hypertreading, so Hollywood CGI studios will love the new CPU for sure. Hyperthreading will hurt results for single CPU benches, but who cares when you have eight CPUs?
In 3DMark03 Hyperthreading does hurt the results, it seems, while 3DMark06 gets much higher CPU-scores.
While in most cases the new Core i7 CPUs are faster, they keep power consumption in check. Calculated by performance by wattage the i7 wins. The CPU does load more cores due to Hyperthreading enabled.
For unknown reasons the 920 needs a bit more in idle, compared to the 965XE which does great on idle-power through all speeds because we used only the dynamic overvoltage feature of the Intel board which keeps stock voltages nominal. We have no idea if any other vendor has the same feature, but we guess time will tell.
The "full" load does include two cores for x264, one core for superpi and the rest for 3DMark06. Of course we could have loaded the i7 more, but consider there are not many applications available which will force all "eight" CPUs with high graphics-card load. So, we think this is quite suitable. This results do not reflect the speed advantage of the CPU, we will do a more comprehensive test in our Games Special, where we will have FPS vs power consumption.
While Intel had no need to bring a faster CPU in the desktop market, the new CPUs are quite promising and the speed increase impressive. They are the fastest CPU you can buy, but while the 920 is quite "cheap" at under €300, you have to change your whole platform. Board prices start now at about €220 up to €350, so you have to even pay more for a board than your CPU, which is quite odd.
Also the cheapest 3GB PC3-8500 CL7 kits will cost you about €120; 6GB starts at €200 where a 4GB DDR2 kit can be bought for roughly €35, which is insane.
So even when you consider the smallest of the Core i7 CPUs "cheap", the whole system will cost you a fortune and the better performance does not justify the cost for everyday users or gamers. If you use your computer for commercial rendering the cost will be justified by the time-savings you can achive. We doubt a professional Photoshop user can do it, because most plugins are only optimized for two cores, Otherwise, there is no need to upgrade as long as your system costs three times more, compared to a standard DDR2 system.
We will have to wait for next year, when the Core i7 Dual Core CPUs in the LGA1156 socket arrive. The boards for it should be cheaper because you won't need a six-layer PCB due to fact this CPUs will "only" feature a dual-channel memory interface, besides OEM demand will reduce DDR3 costs.
Sadly, with the two-socket policy there is no question of upgrading, either you want two cores with four threads or you want four cores with eight threads. In our opinion Intel could have achived the differentiation between workstation and normal user by introducing new Xeons first and skipping the "core" madness. We predict that even in two years there will be hardly any application for the masses which will benefit from eight threads at all.