My thoughts on a higher-model i7

[Randall Leong]

Recently, I replaced an i7-920 processor in my main system with an i7-950 because the price of $230 was too good to pass up. I then put the system through the PPBM5 benchmark procedure (with the i7-950 overclocked to 3.72GHz), and came up with the following results:

1) There was no change in the timeline rendering times, with MPE on or off.

2) The MPEG-2 and H.264 encoding times increased by a second or two - not enough to be of consequence.

3) The Disk test time, however, increased by eight seconds with the i7-950 @ 3.72GHz compared to the 86-second result of an i7-920 @ 3.675GHz. I suspect that this is due to the lower BCLK speed needed on the i7-950, and thus lower QPI bandwidth. In this particular case, the i7-950 needed only a 155MHz BCLK to reach its 3.72GHz while the i7-920 needed a BCLK of 175MHz to reach 3.675GHz (in both cases, at their Turbo multiplier of 24x for the i7-950 and 21x for the i7-920).

So, with every single i7 CPU family, the higher-model CPU will be faster at stock speed, but won't perform quite as well as a lower-model CPU with the exact same type of core at the same overclocked clock speed. Therefore my recommendation is if you're going to overclock, go with the lowest available CPU in that series unless you are using really cheap, lower-quality memory.

Do you agree or disagree with this?

[Randall Leong]

If I did not see a response to this thread, please note that my results are with the Turbo Boost enabled (and thus the CPUs are running at their maximum Turbo multiplier at all times).

I think I will play around some more with the maximum non-Turbo multiplier on the i7-950 (with Turbo disabled), This will force the BCLK to a higher speed. My current Gigabyte mobo tends to peter out at just above a 180MHz BCLK regardless of the CPU (running Prime95 resulted in a BSOD regardless of the memory used). And I deliberately kept the CPU clock speed to around 3.7GHz maximum because even with my current NH-U12P SE2 cooler, the CPU temperatures still reached 90°C or higher at 3.8GHz with this mobo. After a new round of tests, I will check back with the results.

UPDATE: I just retested, and there was no significant difference between 155x24 and 175x21 in any of the tests. One thing is for sure: That i7-950 CPU actually runs more stably than my previous i7-920 (which is now in my auxiliary rig) ever did.

[Peter Chung]

Is there a preferred CPU for overclocking? There is the i7 875K, which is unlocked for overclocking but uses the 1156 socket. Is it better for overclocking than the 930 or 950? They are all in the same price range and it seems overclocking gives some pretty good improvements in CS5.

[Randall Leong]

Originally Posted by Peter Chung

Is there a preferred CPU for overclocking? There is the i7 875K, which is unlocked for overclocking but uses the 1156 socket. Is it better for overclocking than the 930 or 950? They are all in the same price range and it seems overclocking gives some pretty good improvements in CS5.

Actually, LGA 1156 is not as well suited to CS5 as LGA 1366. Namely due to the lower bandwidth between the LGA 1156 CPU and the motherboard's PCH compared to the bandwidth between the X58's IOH and the ICH10R. In addition, the dual-channel memory controller on LGA 1156 CPUs has a theoretically lower maximum bandwidth compared to the triple-channel memory controller on LGA 1366 CPUs.

[Harm Millaard]

Randall,

You are forgetting the major drawback of the chipset IMO: the limited number of PCIe lanes, which prevents one to use a raid controller or other PCIe card.

[Randall Leong]

Originally Posted by Harm Millaard

Randall,

You are forgetting the major drawback of the chipset IMO: the limited number of PCIe lanes, which prevents one to use a raid controller or other PCIe card.

I was aware of this limitation, but I did not mention it. Although the 1156/P55 platform technically has 24 PCIe lanes, only 16 of them operate at full PCIe 2.0 spec. And a RAID controller needs the full PCIe 2.0 bandwidth to function at its best. The eight PCIe lanes provided by the P55 PCH itself are technically also PCIe 2.0 but are artificially restricted to PCIe 1.x bandwidth.

So, the 1156 platform does not have enough full-bandwidth PCIe lanes to fully support the use of an add-on RAID controller especially since the graphics cards would have already consumed all of the CPU's integrated full-bandwidth PCIe lanes. Thus, only the half-bandwidth PCIe lanes are available for expansion.

I am just clarifying this misconception that all 1156 platforms have only 16 total PCI-e lanes. If this were the case, then all expansion would have been available only externally (via USB) or via the legacy PCI bus - and one would not have been able to even use the onboard NIC (on those motherboards that use a NIC controller from a company other than Intel; for example, the typical onboard Realtek or Marvell NIC uses up one additional PCIe lane).

By a similar token, most X58 motherboards for the LGA 1366 CPUs do not fully utilise the X58 IOH's 36 PCIe lanes. A few utilise only 16 PCIe lanes, leaving 20 PCIe lanes unusable. And on those X58 motherboards that run three graphics cards in an x16/x16/x4 configuration, the x4 slot is run off of the ICH10R instead of the X58 IOH - and that x4 slot is of PCIe 1.x bandwidth.

The forthcoming LGA 1155 Sandy Bridge platform is slated to replace the LGA 1156 platform around the beginning of next year. Unfortunately, the CPU will still be saddled with the same 16 PCIe lanes as the current LGA 1156 CPUs but the eight PCIe lanes on the P67 and H67 chipsets that will be used with the new platform will at least be of full PCIe 2.0 bandwidth. So, all 24 of the PCIe lanes on higher-end implementations of the LGA 1155 platform will be full PCIe 2.0. However, early published reports indicated that the LGA 1155 platforms will have all of the buses tied to the BCLK, thus hindering overclocking capability via the BCLK (even a BCLK setting of 105MHz is enough to cause instability, where the stock BCLK is 100MHz). Also, the new P- and H-series chipsets for LGA 1155 will not natively support the legacy PCI bus; however, most motherboards will incorporate a PCIe-to-PCI bridge chip. PCIe 3.0 will not arrive until some time after the introduction of the LGA 2011 platform towards the end of next year. Unfortunately for budget performance seekers, the least-expensive "non-crippled" LGA 2011 CPU is expected to cost $1,000 by itself -- the same price as the current going price of an i7-980X Extreme.

As for one of my two i7 rigs needing all of 140 seconds to complete the AVI disk test (I will be submitting that result shortly), it was due to the use of a 1TB WD Black drive as the project/media/rendering drive. At least with the version of the Black that I have, it was not one of the faster 7200 rpm SATA hard drives in sequential transfer speed in part due to its relatively low areal density; in fact, this particular Black is only about as fast as the latest revisions of the 5405 rpm WD Green. On the plus side, with only 6GB of RAM installed on that auxiliary i7 rig (overclocked to 3.36GHz), the MPEG-2 encode time in PPBM5 got shortened from the 190-ish seconds in 5.0.1 to roughly 120 seconds in 5.0.2 though even 5.0.2 would have performed better still with 12GB or more RAM.

[Randall Leong]

I just looked over the LGA 1156 landscape again, and discovered that not all such platforms are suitable for a video editing system. Currently available hardware RAID controller cards cannot take full advantage of more than four PCI-e 1.0 lanes (or 10 Gbps, or 1 GB/s) even though most have electrical provisions for a PCI-e x8 slot. So, theoretically an LGA 1156 system can be used without much if any degradation of performance. However, many LGA 1156 motherboards cannot accommodate both a graphics card and a hardware RAID controller card at the same time not because of the platform's shortage of available PCI-e lanes, but because those particular motherboards simply have only one PCI-e x16-length slot (and thus, the only remaining PCI-e slots available are all x1 slots, which would have been slower than a single onboard SATA port on the motherboard). Many P55 and H57 motherboards have an x4 slot in addition to the x16 slot or split the two x16-length slots electrically into two x8 slots when both slots are occupied; those motherboards are better suited to a video editing system if one must use an LGA 1156 system. (And as I have proved recently, not even a GTX 480 takes anywhere near full advantage of the full 16 PCI-e 2.0 lanes in terms of bandwidth.)

Which takes this back to the following quote:

Originally Posted by Peter Chung

Is there a preferred CPU for overclocking? There is the i7 875K, which is unlocked for overclocking but uses the 1156 socket. Is it better for overclocking than the 930 or 950? They are all in the same price range and it seems overclocking gives some pretty good improvements in CS5.

As stated in the previous paragraph, it depends on the motherboard. Many LGA 1156 motherboards are ill-suited to a video editing system due to their lack of a sufficient number of full-length PCI-e slots running at least x4, not because of their shortage of available PCI-e lanes.

Hope this clarifies things.

[Peter Chung]

I won't be using any RAID controllers so it shouldn't be too much a problem.

I'm just trying to find the best value for my investment so was wondering if the unlocked CPUs (i7 875K) allow more or better overclocking vs the standard locked i7 processors (i7 930, 950). It is my understanding that the i7 CPUs can be easily and safely overclocked and give good improvements.

[Harm Millaard]

Peter,

You are comparing two different platforms, P55 and X58, and all 9xx are perfectly capable of high overclocking with sufficient cooling.

[Randall Leong]

Originally Posted by Harm Millaard

Peter,

You are comparing two different platforms, P55 and X58, and all 9xx are perfectly capable of high overclocking with sufficient cooling.

I am going to try my hand at lapping the HSF. My CPU temps are reaching the 80s C (at full load) at a relatively modest overclock to 3.7GHz.

[Peter Chung]

Originally Posted by Harm Millaard

Peter,

You are comparing two different platforms, P55 and X58, and all 9xx are perfectly capable of high overclocking with sufficient cooling.

That's good to know.

[Randall Leong]

Originally Posted by Randall Leong

Recently, I replaced an i7-920 processor in my main system with an i7-950 because the price of $230 was too good to pass up. I then put the system through the PPBM5 benchmark procedure (with the i7-950 overclocked to 3.72GHz), and came up with the following results:

1) There was no change in the timeline rendering times, with MPE on or off.

2) The MPEG-2 and H.264 encoding times increased by a second or two - not enough to be of consequence.

3) The Disk test time, however, increased by eight seconds with the i7-950 @ 3.72GHz compared to the 86-second result of an i7-920 @ 3.675GHz. I suspect that this is due to the lower BCLK speed needed on the i7-950, and thus lower QPI bandwidth. In this particular case, the i7-950 needed only a 155MHz BCLK to reach its 3.72GHz while the i7-920 needed a BCLK of 175MHz to reach 3.675GHz (in both cases, at their Turbo multiplier of 24x for the i7-950 and 21x for the i7-920).

So, with every single i7 CPU family, the higher-model CPU will be faster at stock speed, but won't perform quite as well as a lower-model CPU with the exact same type of core at the same overclocked clock speed. Therefore my recommendation is if you're going to overclock, go with the lowest available CPU in that series unless you are using really cheap, lower-quality memory.

It turned out that I was using the maximum Turbo multiplier on both of those CPUs in my testing, and that some of the discrepency is due to the fact that the memory was running at DDR3-1400 with the i7-920 but only DDR3-1240 with the i7-950. Unfortunately, the next-higher memory speed at 155MHz BCLK is DDR3-1550, which did not work stably in my system at even 9-9-9-24 timings. I have yet to test it at DDR3-1550 and 10-11-11-28 timings - but based on my tests at DDR3-1600 @ 10-11-11-29 timings with my i7-950 at stock CPU speeds, it shaved just five seconds off the total PPBM5 benchmark time (297 seconds) compared to the 302 seconds that I achieved with the same memory running at DDR3-1333 and 9-9-9-24 timings.

UPDATE: I did a test with the i7-950 @ 3.72GHz with the memory set at DDR3-1550 speed and 10-11-11-28 timings, I got just about the same result (in terms of the PPBM5 test times) as I had achieved with an i7-920 @ 3.675GHz with the memory running at DDR3-1400 and 9-9-9-24 timings. Between these two CPUs, the performance is a toss-up.

[Harm Millaard]

Peter,

I just updated the PPBM5 Benchmark with the latest submissions and added some information on where we are intending to go with the benchmark: Latest News

Looking at the BFTB chart, I feel pretty comfortable with my top position, although I do not like to have only the 9-th / 7-th position in total performance in time and RPI, but that is of course due to an almost 2 year old system with last generation disks, a outdated CPU and mediocre memory.

Today your best bet is the 950 CPU and the GTX-470 video card IMO. BFTB-wise that is.

[Panagiotis Raris]

i had a TON of stability problems overclocking my (3!) i7 920's. i RMA'd 2 of them and finally just sold the 3rd to a friend. The 950 has been wonderfully stable, even at 3.8Ghz.

How comfortable are all of you with temperatures above say 70 celcius? i have been somewhat militant, but ive kept all of mine below 50 celcius at full utilization on my personal microATX machine with some clever fan work and a corsair H50 with a second 2200RPM fan (2200RPM at 20000RPM push, 1400RPM, came with H50 fan at 1400RPM pull has been best configuration for temps). I also have a 250MM fan blowing on the BACK of the motherboard about 2 inches off the surface (cut a large section of sheet steel out and soldered in stainless steel mesh) and have a PSU 140mm exhaust fan and a few other fans crammed in, on, over, under, or around the case as it runs 24/7. This is my home editing PC, media PC, movie/DVR storage, etc machine.

I noticed the southbridge and northbridge on my Asus got very hot, VERY quickly. 75 celcius with 5 minutes of prime95; so i threw some 80mm 3400RPM intake fans on them and made some clever air inlet and exhaust outlet ducting. I did the same to my 4 disk RAID0 array, and graphics card. Now even at 100% for 2+ hours, temps never reach over 50 celcius at 21 celcius ambient. I had to add a fan controller deck, that came with temp monitoring.

My Q6600 died after a prolonged export (7+ hours) and it was hovering around 88 celcius, temps also killed my P4 3.06Ghz HT laptop processor (85 celcius for 13+ hours) so i strive to keep temps WELL below 70; my CPU idles at 31, ram at 38 (no fan), northbridge and southbridge at low 40's, HDD's at 38, PC case ambient at 25 with room temp of 21. Under load, nothing exceeds 50 celcius.

[Harm Millaard]

Did you have a look here: Adobe Forums: Overclocking the i7, a beginners guide