Huawei API v4.0

Sunday, March 7, 2010

Maximum Efficiency: Build A 25W Performance PC Using Core i5

Back in late 2008, we ran an article showing how an efficient power supply, low-end motherboard, and entry-level Core 2 Duo could reach less than 40W system idle power. Then we took the test config and proved that a Core 2 system provides power consumption similar to that of a cheap, Atom-based desktop solution, even busting Atom in performance per watt. The introduction of the Core i5/i3 Clarkdale dual-cores with the H55/H57 chipset and on-package integrated graphics marks the next step. Systems at 30W and less are now within reach for everyone. We wanted to go even lower this time around and set 25W as the target-but at zero compromise

Why Do I Care?

I should clarify that (apart from a discussion about global environment concerns) there's no urgent need to achieve the lowest power consumption possible on your desktop PC. This is is an optional and voluntary item. An average desktop PC requires between 45W and 60W, as long as it's employing integrated graphics and mainstream components. If you add powerful graphics and more capable hardware, you'll be at 80W and up. In this context, 50W is perfectly okay for a modern PC without serious 3D graphics. But why not pay attention to system power if it doesn't cost anything or jeopardize performance or features?

A 30W (versus 50W or higher) system idle power draw will mean only small savings on your energy bill, but it will introduce some added advantages. The most important of these certainly is the drop in required cooling. A 30W idle translates into slower-spinning fans, both on the processor and throughout the system. Another direct advantage is lower noise levels, as temperature controlled coolers have less work to do. Even passively-cooled PCs are clearly within range at 30W, although such systems typically come at a price premium.

25W PC Idle Power

Reaching almost 30W system-wide idle power was possible with Core 2 Duo when using a motherboard with integrated graphics and an efficient power supply. Of course, such a system could only deliver mainstream performance, and almost always entailed a reduced feature set. For example, G31 integrated graphics might only sip a few watts, but the IGP was already old fashioned at the time of our original article.

Today's 25W idle power target is obtainable thanks to the new Core i3/i5 Clardale-based processors and the H55/H57 chipset. This platform offers the best overall efficiency today. With additional measures, such as slight undervolting or smart component choices, we could actually go even lower. However, it's amazing that we can reach such low system power without jeopardizing performance or features.

Naturally, you won't get quad-core performance and 3D graphics horsepower; we're talking about mainstream components, the performance you'd expect from a dual-core CPU, and premium features. Let's have a look at no-compromise computing at unbelievably low power.

Still fresh out of Intel's starting blocks, the Clarkdale-based 32nm dual-core CPU with an on-package 45nm graphics/memory controller has made quite a good impression with regard to performance. The family from which the -661 hails remains controversial though, mainly because of its steep pricing. AMD offers much better value per dollar, and the direct comparison of AMD's existing dual-core CPUs against Intel's new ones was perceived as largely unfair. We'll be following up with more articles and value comparisons of AMD and Intel in the following weeks.

We again used the Core i5-661 here, since it is the premium graphics-focused flagship in the i5 lineup, running the IGP at 900 MHz instead of the 733 MHz found on all other Core i5 and Core i3 models. The result is a TDP of 87W, rather than 73W (although we found that the processor typically doesn't get too close to its maximum thermal ratings). Yet the CPU, combined with MSI's H57M-ED65 motherboard was low enough on power to reach the 25W idle limit we desired.

The H57M-EG65 is a mainstream H57 motherboard in microATX format. Important enthusiast features include a dynamic eight-phase voltage regulator, a massive heatpipe, twin x16 PCI Express slots (though remember, Clarkdale-based processors cannot divide their 16 PCIe lanes on an H55/57 motherboard), unofficial DDR3 support up to 2,133 MT/s, RAID 5 support, and MSI's overclocking assistant, OC Genie. MSI exposes HDMI, DVI, and D-Sub display outputs, as well as an eSATA port for external high-speed storage devices.

Still, we missed the ability to reduce component voltages in the board's BIOS. This would be necessary to reduce power consumption at stock speeds. Half a year ago, we reduced the core voltage on AMD and Intel systems and achieved impressive power savings. The same could be done on the modern Clarkdale-based chips as well. Nevertheless, we took the easier path of using MSI's ControlCenter to reduce the processor voltage.

Step 1: An Efficient Power Supply

We've been using the 750W Silencer from PC Power and Cooling for a long time because it offers enough output for a majority of the configurations we test on a daily basis here in the lab. It is an 80 PLUS device, but it's not the right choice for a low-power computer for a very simple reason: power switching circuits are typically only efficient within a certain output range.

For this reason, the 80 PLUS specifications define efficiency for three different load scenarios. In a simple example, 80 PLUS Gold requires 87% efficiency at 20% of the nominal load, 90% efficiency at 50% load, and 87% efficiency at full output. In the case of our 750W Silencer, 20% load equals 150W-a level we want nothing to do with on our test system. Clearly, we needed a different PSU for this project.

This small 220W power supply by FSP/Fortron is also an 80 PLUS device. Its small capacity may not make it seem attractive at first, but a 20% load equals 44W-much closer to where we want to land. Check out the results of this simple component exchange in the results section on the following pages. This measure alone was the most important step in power consumption reduction, moving the system from 33W idle to only 26W idle.

Step 2: Undervolting

We decided to tweak the CPU core voltage and introduce a minor Vcore reduction to see if we could make another impact on power consumption. Experience tells us that a few percent voltage reduction typically doesn't lead to stability issues. However, excessive undervolting can certainly be problematic, and insufficient voltage supply can be complex and tricky to identify. If you aren't sure, we recommend staying close to the default voltage to avoid unneccessary issues.

MSI's ControlCenter tells us that the default voltage for our test system is 0.95V (idle). CPU-Z confirms the setting and tells us that it's effectively 0.944V.

We decided to adjust the voltage from 0.95V to 0.90V in the ControlCenter software.

The result was an effective CPU voltage of 0.888V at idle.

Load Vcore Readings

Peak Vcore at default auto settings.

Here's the peak Vcore at 0.9V manual setting.

We measured a decrease in system power at the reduced settings, but the impact was negligible at idle. At peak load, we had a decrease from 80W to 76W total system power.

Step 3: An Efficient Hard Drive

There are still a few components left to optimize: the main memory and the hard drive. We tried removing one of the two DIMMs, but since DDR3 memory uses less power than DDR2, the impact was extremely small-doubly predictable, since we were running the DDR3 at 1,333 MT/s and default voltage. However, we have room to drop consumption versus the Western Digital VelociRaptor in our reference system, since it was never designed to be a low-power component.

We received Toshiba's newest 640GB 2.5" hard drive, the MK6465GSX, at about the same time we completed this review. Since its power consumption is among the lowest seen in today's hard drives, we decided to use it instead of a faster 3.5" desktop drive.

The 5,400 RPM Toshiba helped deliver additional power savings. System idle power went from 26W with the undervolted system on the 220W FSP power supply to as little as 23W. That put us under the 25W target for which we were aiming. The difference at peak load is only two watts, but we're getting close to the accuracy limits of our measuring equipment at this point. The MX6465GSX is also available in 160, 250, 320, and 500 gigabyte capacities. A detailed review of the new 640GB drives will follow soon.

Of course, a solid state drive is also an option here, but the limited capacity would be an issue for most folks, and the increased cost would push the value of this build in the wrong direction. For that reason, a 2.5" mobile drive with more room for storage makes the best sense.

Test Setup And Power Consumption Results

System Hardware



Motherboard (LGA 1156)

MSI H57M-ED65 (Rev. 1.0)

Chipset: Intel H57, BIOS: 1.0B28 (04/01/20100)

CPU Intel

Intel Core i5-661 (32nm, 3.33 GHz, 2 x 256KB L2 and 4MB L3 Cache, TDP 87W, Rev. B1)

RAM DDR3 (dual)

2 x 2GB DDR3-1600 (Corsair CM3X2G1600C9DHX)


Intel HD Graphics

Hard Drive I

Western Digital VelociRaptor, 300GB (WD3000HLFS), 10,000 RPM, SATA 3 Gb/s, 16MB Cache

Hard Drive I

Toshiba MK6465GSX, 640GB, 2.5," 5,400 RPM, SATA 3 Gb/s, 8MB Cache

Power Supply I

PC Power & Cooling, Silencer 750EPS12V 750W

Power Supply II

Fortron FSP220-60LE 220W

System Software & Drivers

Operating System

Windows Vista Enterprise Version 6.0 x64

Service Pack 2 (Build 6000)

Drivers and Settings

Intel Chipset Drivers

Chipset Installation Utility Ver.

Intel Storage Drivers

Matrix Storage Drivers Ver.

Results: Idle and Peak Power

We started off at 33W idle power, which is low but not revolutionary. We then exchanged the high-power PSU with a more efficient one.

The most important thing in this measurement is the fact that we aligned the PSU's output to the expected system power consumption as much as possible. A 750W PSU running at only 40W load would be pretty inefficient. For this reason, the idle power decreased from 33W to only 26W. Even our small voltage decrease didn't make much of a difference . The only other step that had an impact was the use of Toshiba's 640GB 5,400 RPM hard drive instead of the high-performance WD VelociRaptor. Ultimately, we dropped from 33W to 23W with only two component changes. This is a 30% reduction in idle power.

Peak power doesn't differ that much, as even the 750W Silencer power supply is operating at much more efficient levels. This is why its replacement doesn't make as much of a difference under a heavy load. The voltage tweak, however, did make a difference of 4W. Another 2W could be saved through the more efficient hard drive.

Overall, the total impact on peak power is less significant than the change in idle power. This is where a desktop-class build will be significantly more power-hungry than a machine centering on Intel's Atom CPU. However, most desktops will spend little time using this much juice. In return, they also finish the same tasks in less time.

Benchmark Results: Efficiency (Performance Per Watt)

Once again, we ran our efficiency suite, which consists of several scripted benchmarks.

Since we keep tracking power in one-second intervals, we can calculate the average power used for the entire workload. The difference is significant considering that the performance does not change. All we're doing is modifying components and voltage settings in an effort to decrease power consumption.

The results on total power used for this test run are similar to the average readings.

The total runtime is one form of a performance result. As you can see, the results are very similar for the first three configurations. The last test run, which was conducted with the slower but more efficient Toshiba 2.5" hard drive, delivers less overall performance due to the decreased disk performance.

We divided the performance result by the power used to get our efficiency score. As expected, efficiency increases quite a bit.

The efficiency diagram visualizes power consumption at every point of the benchmark run and also shows differences in runtime.


We expected that it would be hard to get the Core i5-661 system to run at only 25W idle power, but it turned out that the entire project was pretty easy to execute. Our test results show impressively that matching the power supply to your expected output (with some power reserves, of course) is imperative. Replacing the high-output 750W PC Power and Cooling PSU, which is a neat piece of hardware for a high-end gaming rig, with a low-power 220W PSU made the largest difference in our efforts to lower system power. Everything else is secondary. VoilĂ -26W reached with zero performance impact.

22-28W Idle Power is Realistic

We tried running the system with only one DDR3 DIMM module instead of two, but found that the savings are below 1W. At the same time, this measure would have had a noticeable impact on performance. This also applies to the hard drive, but we were able to save an additional 2W to 3W there. Still, even if you only pick a suitable power supply and optimize component voltage, it's absolutely possible for anyone to recreate our 25W build-assuming that you go for the latest mainstream 32nm Intel hardware and a motherboard with the H55/H57 chipset. The voltage reduction reduced peak power, but had little impact on idle power.

And Now?

Let me come back to the question of why this is relevant. For power users, it really isn't. You'll probably go straight to a quad-core machine and not bother with power consumption questions. For average users, it becomes relevant if you want to go for a low noise system, or if you intend to pair high performance with the lowest possible acoustic footprint. This is typical of an HTPC, living room machine, or home server application. Keep in mind that a 3+ GHz Core i5 dual-core system is significantly faster than any other dual-core configurations around. It's finally possible to get strong performance at the power consumption levels of a dreadfully-slow Atom machine.

Intel's Low Power Platform Means Higher Cost

However, there is a significant catch: you have to be aware that the new Intel platform is not a bargain. The processors are expensive, especially if you look at AMD's pricing in comparison. Obviously, AMD has to stay aggressive on pricing to remain competitive from a performance perspective, but the result really makes you wonder. Do you really need to spend several hundred bucks on a Core i5 solution to reach the lowest power, or will an AMD solution at half the cost, but higher power consumption, be more suitable? We've made it clear that energy cost isn't really an issue at this low level. If you want to go for the lowest power, for whatever reason, Clarkdale-based Core i5 processors paired with suitable components are unbeatable. They're just expensive relative to the competition, which is why we continue to recommend the i5-750 to enthusiasts at the $200-ish price point where you find the i5-661.

Powered by Zoundry Raven

No comments:

Post a Comment