iXBT Labs - Computer Hardware in Detail






Intel Core i7 LGA1366 Processors

<< Previous page

     Next page >>

Alas, we don't see the future for Core i7 processors in their present design. A temporary solution that was born out of a contradiction between PR promises and real solution designed by engineers (they promised to launch Nehalem at the end of 2008, but they designed neither server nor desktop modifications by that time) cannot live for long. It had a bright, but short past -- the first processors of this series with indices 920, 940, and 965 (the last digit 5 in this series stands for extreme modifications, not cut-downs) entered the market in November, 2008. And they had been the only Nehalem processors for entire six months, because Xeon with this core was completed only in March, 2009. By the way, Series 5500 differs from i7 only in two QPI links instead of one. Besides, Northbridges in the i5520 and X58 are in the same situation, so it's not hard to guess which problems Intel faced last year. Yep, QPI bus is a new solution. It's similar to some of the existing solutions, but this bus is much faster than all of them. So no wonder that its 'double' modification was ready only six months after its 'single' model. And launching Xeon 3500 alone did not look like the right decision -- these processors exist as cheaper modifications of Xeon 5500 now, thus raising no questions. But a single-socket Xeon family would have raised them.

A tad later the Core i7 family was overhauled, as the manufacturer was not satisfied with its initial design. For one, there appeared Xeons that immediately captured interests of workstation makers. For two (which is more important for most users), performance of the old Core 2 series has grown as well -- Core 2 Quad Q9650 (3.0 GHz) and Core 2 Duo E8600/E7600 (3.33 GHz) come with relatively low price tags. They are not sterling competitors for Core i7, of course. But they can try a fall with the cheapest model in a number of cases, keeping the price of an entire system to minimum. LGA1156 processors will debut very soon, and they can cover mid- and high-performance segments for reasonable money, leaving only top systems for Core i7 LGA1366. For obvious reasons, users of the latter processors should get the best there is, not almost the best. That's why the company has made some changes, replacing the 940 with the 950, and the 965 with the 975. The clock rate has grown by 133 MHz, while the price remains the same. The top non-extreme Core 2 Quad processor has the nominal clock rate of 3 GHz, while the top model of Lynnfield (which will belong to Core i7, even though it will have a different socket) will work at 2.93 GHz at first. That is Core i7 950 with its official 3.06 GHz (in fact, it will be a bit higher, we'll talk about it later) looks much better in all parameters (even in terms of megahertz) than Core i7 940. And what about the 920? Formally, this processor will be manufactured a little longer, it's been upgraded to the new stepping, but... it has already competed its mission. So we cannot expect an inexpensive processor for the top platform. The company needs it so far, but it will cease to exist right after the appearance of popular Nehalems. However, you can think about adding these three hundred dollars of the difference between the 920 and the 950 and get a noticeably faster processor even now. The power of this platform is revealed, only when you don't save on the other components either.

So, this family includes five models so far. Two of them are going to disappear soon. Another model will probably be discontinued later. We've heard some rumors that the Core i7 950 won't live through the end of the year either: it may be replaced with the 960 (its clock rate is identical to that of the old eXtreme 965 processor, but it cannot be tweaked). But we decided to test all five processors using our new test procedure in order to evaluate performance gains of the new models versus old ones. Besides, these results will come in handy to compare with other processors from Intel and its competitors. But first of all, let's find out the frequency Core i7 processors work at in actual fact.

Turbo throttling

It's a provocative title. However, this issue is even more provocative: is Turbo Boost a cynical version of throttling? It's an appropriate question -- there are different angles of view for each phenomenon. For example, frequency raises when possible or vice versa -- it drops when necessary. If we say that the official clock rate of these processors is indeed their main frequency, the first is correct. But is this assumption correct?

Let's recall the initial kind of throttling. This mechanism made Pentium 4 processors insert idle cycles, when critical temperature was reached. It reduced the effective clock rate as well as heat release, so this procedure kept CPU temperature within safe limits even in 'bad' conditions. It was a rude mechanism, but it was easy to implement and efficient. It's not used now, even though processors still support it. The fact is, both companies (Intel and AMD) have mastered a much more attractive and effective power saving technology. It changes the multiplier (and consequently the clock rate) as well as voltage to save power in idle mode and effectively solve the problem with overheating. Any modern processor supports EIST or Cool'n'Quiet, and this fact rendered the initial mechanism of throttling outdated.

EIST is supported by Core i7 as well. It differs from the previous Intel solutions only in the following: it can increase CPU frequency, not only drop it. What concerns Turbo Boost, we initially knew only that it worked, when CPU power consumption was within limits specified by the company (not disclosed at the time). The clock rate was considered to depend on CPU load. That is when all four cores are loaded, their clock rate will be equal to the nominal value. When two or three, it will be raised by 133 MHz, when only one -- by 266 MHz. The reality turns out to be much more interesting.

For one, CPU frequency grows even when all four cores are loaded. In this case -- by one step of the multiplier, in all the other -- by two steps. So you should overclock CPUs with a bus very carefully -- you may get curious results: when an already overclocked processor tries to accelerate by another 400 MHz under load, nothing good will come out of it. So in this case you should disable this new technology. Extreme modifications of processors for enthusiasts allow a subtler approach to overclocking. For one, they can be overclocked without increasing the reference clock rate to limit Turbo Boost. For two, it's not necessary. Moreover, Intel removed the option to increase the CPU multiplier, when boost is enabled, in the latest BIOS versions for our testbed. In return, you can adjust the boost amount -- just select multipliers for one, two, three, or four cores under load. It's clear why these situations are different: idle cores do not contribute to heat release, but they take part in channeling the heat away. So the 'base' multiplier can be left unchanged -- just increase boost ratios in return. In this case, the 965 processor can be considered as a quad-core 3.7-GHz CPU (Multiplier 28), 2- or 3-core 3.85-GHz CPU (29), or a single-core 4-GHz CPU (30). This scenario preserves all power saving technologies and protection from overheating.

When does this boost works? There are two more parameters. They are implemented in all processors, but they can be changed only in eXtreme editions. They are maximum power consumption in the boost mode and usual maximum power consumption. The latter is the TDP Core i7 that equals 130 W. Overheating protection snaps into action above this threshold. If you are confident about your cooling system, you can raise this threshold to 150 W. Or even to 200 W, if you want to break a new benchmarking record. Bad news for such record breakers -- it cannot be done without increasing this parameter anymore! Processors of the previous generation allowed to control only one parameter, namely temperature (straight forward: overheating -- throttling). But now PSUs always monitor power consumption of each core. And then temperature and power consumption are used to adjust CPU frequency and voltage for each core. In other words, good CPU cooling had been the only requirement for excellent overclocking results, but now it's practically impossible to exceed the limits of a given 'thermal package'. Current capacity equals current strength multiplied by voltage. So when you try to exceed this value (which is accompanied by growing current strength), voltage will be decreased. As a result, you will have to reduce the multiplier (that is CPU frequency) to preserve stability. It's an excellent arrangement for nominal modes. It will ensure that a processor works even if you try to use a 'dumb-ass' mode. But if you actually need this mode, this procedure will only stand in your way -- you will overclock a processor, but it won't get faster.

It concerns processors working in the usual mode, but Turbo Boost actually does not differ from it! The upper threshold here is 110 W. It's hardcoded for popular models and adjustable for eXtreme editions. For example, both thresholds can be set to the identical value (130 W or 150 W). So what have we got? Turbo Boost will be always enabled, until the threshold power consumption is reached, when we start to reduce the frequency to cool the processor. However, the processor may stop working in the nominal mode at all. It happens rarely even without any adjustments -- 110-130 W is a narrower range than the 'normal' 50-90 W range. Note that it usually takes special software to warm processors above 100 W. They do not reach that level in the normal mode. Thus, the base frequency of Core i7 processors in normal conditions is the boost frequency, but not the nominal mode. And the nominal mode is a special case of thermal protection mechanism -- it snaps into action when the first threshold is reached. It's just that Core 2 has a different threshold than Core i7 and the new Xeon processors. Core 2 processors controlled power consumption using only indirect signs (namely, temperature), while Nehalems measure and control power consumption directly. So, it's not thermal protection, but sterling power and frequency management.

How should we interpret that? There are two ways. It's either a brilliant engineering solution, a significant step forward. Or it's a cynical version of throttling and a stumbling block on the way to overclocking. It's up to you to decide, which definition is better. We like the first one. However, it does not cancel the fact that such mechanisms make it impossible to determine at what frequency the Core i7 or the new Xeon processor work. We can measure the instant clock rate at any time (the only problem is that all software methods of measuring it affect CPU load, so they introduce a certain measurement error). We know the so-called base frequency of a processor specified in its manual. We know the starting operating frequency selected with the multiplier for the 'non-turbo' mode. We also know the maximum clock rate of a processor core. But we cannot 'guess' the clock rate of each core at a given time. For this purpose we have to know the maximum and starting operating frequencies as well as the amount and type of computational load. Just because the real clock rate now depends on them, and it may grow above or fall below the starting value. Processors could do only reduction before this technology appeared (when idle or overheated), boost has been added only now. Perhaps, it was advanced power management technologies that suggested the idea to give up marking processors by their clock rate, not impeachment of high frequencies by the NetBurst family. Because who knows what it is now. The clock rate was indeed a clear-cut physical parameter for Pentium III or old Pentium 4 processors. Later on, there was added an option to decrease it. And now various units in a processor can work at different frequencies, and they can change it dynamically at that. 'Core i7 920' is a simple name, but 'Core i7 2660' does not characterize a processor: it's 'Core i7 2800' or even '2.93' most of the time. And LGA 1156 processors promise to increase their clock rate up to five steps, so the situation will grow even more complex.

By the way, here's a myth about the new processors. There is an opinion that they are fast, so they are hot. An attentive reader must have already cleared up that it's vice versa -- the cause and effect are swapped in this case. Core i7 processors are fast, because they are hot. Indeed, this power management scheme can easily fit into smaller heat release. For example, into 110 W, when the frequency can still go up. But it was not done so that the i7 LGA1366 could demonstrate maximum performance. Xeon processors are not just cooler, they can be slower under heavy load. If the load is so heavy that a processor starts to consume 95 W, Xeon X55x0 will decrease its frequency even below the base level. And Core i7 or Xeon 3500 will still raise it, that is two processors with different nominal frequencies will work in the same conditions at different physical frequencies. The same will hold true for the new Core i7 LGA1156: they will aggressively raise their clock rates under medium and relatively high loads; but in case of heavy loads they will have to fit into the 95-W heat release with all that it implies. The only fact that stops modern i7 processors to demonstrate their absolute advantage over their colleagues is that it's not that easy to warm processors even to 90 W in real applications.

And the last interesting piece of information from the life of processors (I know that my introduction is unusually long, but be tolerant, please -- today's tests will be a little boring) has to do with an opinion that PCU mechanisms do not depend on a platform. That is Turbo Boost as well as protection from excessive power consumption are implemented on the hardware level in a processor. It has always considered to be the main difference from AMD processors, which also have individual voltage/frequency control for each core (taking into account the current strength), and dynamic multipliers. But Cool'n'Quiet 2.0 interacts with the on-board power converter, while the latter should be supported by software. We cannot say for sure whether Intel processors do all the job, but we have some food for thought. The fact is, if you install an eXtreme processor after a regular model without clearing CMOS... you will get absolutely the same settings that were true for the previous die. This way we found out an interesting fact about Xeon: top boost for these processors equals TDP of the previous cooler family. That is the X-series has the 'protection' level set to 80/95, while the E55x0 -- to 60/80. But that's not the main point -- what's important, power protection parameters (it's probably a better term than 'thermal protection' for this CPU series) can be not only adjusted for some processors, they are also stored by a motherboard, and they are changed to defaults only when the specified values don't work. It's up to you to decide whether it's the proof that it's more than just about an Intel die, but a motherboard-processor interaction.

Testbed configurations

Processor Core i7 920 Core i7 940 Core i7 950 Core i7 Extreme 965 Core i7 Extreme 975
Core name Bloomfield Bloomfield Bloomfield Bloomfield Bloomfield
Process technology, nm 45 45 45 45 45
Core clock, GHz 2.66* 2.93* 3.06* 3.2* 3.33*
Number of cores 4 4 4 4 4
L1 Cache, I/D, KB 32/32 32/32 32/32 32/32 32/32
L2 cache, KB 4 x 256 4 x 256 4 x 256 4 x 256 4 x 256
L3 cache, KB 8192 8192 8192 8192 8192
Memory** 3 x DDR3-1066 3 x DDR3-1066 3 x DDR3-1066 3 x DDR3-1066 3 x DDR3-1066
Multiplier 20 22 23 24*** 25***
QPI, GT/s 4.8 4.8 4.8 6.4 6.4
Socket LGA1366 LGA1366 LGA1366 LGA1366 LGA1366
TDP, W 130 130 130 130 130

* When Turbo Boost is enabled (by default), the real clock rate of separate cores grows by 133-266MHz depending on the current load.
** The maximum frequency officially supported by CPU memory controller.
*** Unlocked for overclocking needs.

  Motherboard Memory
Core i7 Intel DX58SO (X58) Kingston KVR1333D3N9K3/6G (1066, 8-8-8-19)
Core i7 Extreme Intel DX58SO (X58) Kingston KVR1333D3N9K3/6G (1333, 9-9-9-24)

Several comments about our tests. First of all, as you all know, Core i7 processors still officially support only DDR3 1066. Memory multiplier for the non-eXtreme editions is limited, while eXtreme processors allow to change it in a wide range. That's exactly what our Intel motherboard with old BIOS versions did: for all 1066 processors. What concerns Extreme Edition, memory modes can be adjusted manually. But the approach has been changed in the latest versions -- now eXtreme processors set memory frequency by SPD even in the automatic mode. That is our memory kit worked at 1333 MHz. We decided to leave it as it was. Especially as we used non-extreme memory (albeit DDR3 1333) that works with safer timings for the top mode: 9-9-9-24 versus 8-8-8-19. Our Corsar modules that we used in older reviews worked with 8-8-8-24, which apparently provided a greater advantage to processors working with memory as DDR3 1333. Let's see whether this advantage is preserved with cheaper memory modules. This question will add some interest to our boring tests (or they risk to degenerate into examination of five identical processors that differ only in their clock rates).

Write a comment below. No registration needed!

Next page >>

Article navigation:

Page 1: Introduction, testbeds

Page 2: Tests, conclusions

blog comments powered by Disqus

  Most Popular Reviews More    RSS  

AMD Phenom II X4 955, Phenom II X4 960T, Phenom II X6 1075T, and Intel Pentium G2120, Core i3-3220, Core i5-3330 Processors

Comparing old, cheap solutions from AMD with new, budget offerings from Intel.
February 1, 2013 · Processor Roundups

Inno3D GeForce GTX 670 iChill, Inno3D GeForce GTX 660 Ti Graphics Cards

A couple of mid-range adapters with original cooling systems.
January 30, 2013 · Video cards: NVIDIA GPUs

Creative Sound Blaster X-Fi Surround 5.1

An external X-Fi solution in tests.
September 9, 2008 · Sound Cards

AMD FX-8350 Processor

The first worthwhile Piledriver CPU.
September 11, 2012 · Processors: AMD

Consumed Power, Energy Consumption: Ivy Bridge vs. Sandy Bridge

Trying out the new method.
September 18, 2012 · Processors: Intel
  Latest Reviews More    RSS  

i3DSpeed, September 2013

Retested all graphics cards with the new drivers.
Oct 18, 2013 · 3Digests

i3DSpeed, August 2013

Added new benchmarks: BioShock Infinite and Metro: Last Light.
Sep 06, 2013 · 3Digests

i3DSpeed, July 2013

Added the test results of NVIDIA GeForce GTX 760 and AMD Radeon HD 7730.
Aug 05, 2013 · 3Digests

Gainward GeForce GTX 650 Ti BOOST 2GB Golden Sample Graphics Card

An excellent hybrid of GeForce GTX 650 Ti and GeForce GTX 660.
Jun 24, 2013 · Video cards: NVIDIA GPUs

i3DSpeed, May 2013

Added the test results of NVIDIA GeForce GTX 770/780.
Jun 03, 2013 · 3Digests
  Latest News More    RSS  

Platform  ·  Video  ·  Multimedia  ·  Mobile  ·  Other  ||  About us & Privacy policy  ·  Twitter  ·  Facebook

Copyright © Byrds Research & Publishing, Ltd., 1997–2011. All rights reserved.