Last articles

Intel Xeon L5640

Tue, 24 Aug 2010 17:01:01 +0200

Introduction

Intel has recently updated its line of Xeon processors, and as such we present to you our test of some of the new low-power Xeons.

For our test we have a pair of Xeon L5640s. These are hexcore processors with a default clock speed of 2.2GHz, though this can rise to 2.5GHz if TurboBoost is active. The processors have a TDP of 60W each, which is just 10W per core, and are also fitted with Intel's HyperThreading technology, allowing each processor to present 12 cores to the OS (24 in total).

Test setup

Machines used in the test :

Our tests were performed on two different machines, as the SuperMicro board we used for the previous generation processors had a bug which meant it could not fully support the L5640's turbo mode.
The benchmarks were set by the following setups:

Dual Quad-core :
Supermicro x8DAi
2 Xeon X5560 ES (2.8 Ghz - 3.06 Ghz with Turbo mode)

Dual Hexcore :
EVGA SuperRecord 2
2 Xeon L5640 ES (2.2 Ghz - 2.5 Ghz with Turbo mode)

The other components in the test machine were the same for both systems:
6 x 2 GB Crucial ECC DDR3-1333
2 Noctua NH-U12P SE2
Nvidia 9800 GT
Windows 7 Professional 64-bit

Test methods:

The tests were carried out using one unit from project 6014 for standard SMP benchmarking and one unit from project 2685 for -bigadv benchmarking.
To measure power consumption, we used a watt meter manufactured by Watt&co. The values given are for total system power consumption, not just the CPUs. For Folding@home power consumption, only the SMP client was running.

Where possible, we will only give a single value for each WU type (currently sorted by how many points the relevant unit is worth). Measurements have been made using projects currently being assigned at the time of writing the article. For ease of use, all performance comparisons are given in PPD. The value we have used is the effective time per frame over the whole unit as measured by FahMon.
We use the site http://sites.google.com/site/mrjellyorg/folding-home/calculator to calculate the true PPD including bonus for SMP and bigadv units

The overclocked frequency of the L5640s was obtained with turbo mode enabled and a QPI of 195MHz. This resulted in a clock speed of 3.7GHz for the 12 cores with the boost from Turbo mode.

Folding@home performance (PPD)

The values are expressed in PPD (points per day)

Here we see a huge difference between the two systems. It is not surprising to see the overclocked L5640 has the performance advantage, but the performance gap is so large that the overclocked hexcore system is more than twice as fast when running standard SMP as the quadcore system was when running bigadv.

Power consumption

All readings are in Watts.

During the test with both machines at their standard settings the difference in consumption between them was only around 20 watts. The dual Quad predictably uses less energy but this is probably attributable to it having fewer cores and less cache than the dual Hexcore. The EVGA SR2 motherboard used for the hexcore also has more features requiring more power, for example an SLI controller chip (NF200), and USB3 and SATA 6Gbps controllers.

Energy efficiency (PPD/W)

All values are in PPD/W

We have recently seen graphics cards capable of 73 PPD/W (see the GT240 review). Here the efficiency may surprise some people: the dual Quad already reports very good results for energy efficiency but the dual L5640 takes it to a new level.
The most surprising result is for the overclocked L5640s. Even with energy consumption at the wall approaching 500W, performance has increased such that PPD/W actually increases with the overclock. In all cases the L5640 system is more efficient than any graphics card we have seen. This could of course change depending on any modification or cessation of the quick-return bonus system, or its extension to more client types.

Conclusion

With every new wave of processors we get new hopes for increased performance, better efficiency, etc. Quite often we are rewarded for our belief but occasionally our hopes are dashed. With the new Xeons, Intel has delivered in style, setting the benchmark for energy-efficient folding.

We have seen that the hype is backed up by the numbers, particularly when overclocked, and for that reason anyone looking to acquire a dual-processor Intel system is well advised to buy the EVGA SR2 as the heart of their system.
The initial investment may appear to be high but the performance has to be seen to be believed. We can only regret the seemingly random availability of the board in Europe.

The justification for paying so much upfront is the reduced outlay when your electricity supplier sends you the bill. Assuming continuous usage, each watt used for one year will cost about 1€. As a result the total operating costs for the overclocked system would come to around 500€ a year, which is extremely cheap for the performance available (though still not for the faint hearted )
One last note about the motherboard: we have already said that it is a must to unlock the true potential of Xeon processors, of this there can be no doubt. However the board will not be accessible to everyone as it is much more complicated to use and overclock than most motherboards. Part of the problem stems from the board's own heatsinks: any QPI baseclock higher than 200 will cause the system to overheat and crash. The use of waterblocks is thus almost mandatory for stable overclocks, though the performance we achieved in this test was done with air cooling only. A member of the site is trying the board out with waterblocks and a pair of X5560s, we will post a news when his tests are complete.

Finally, we must thank M. Moreau of Intel France who had the confidence in us to allow us to test Intel's hardware for you.

ATI Radeon HD5770

Mon, 26 Apr 2010 01:15:01 +0200

Overview

After previously reviewing two nVidia graphics cards, we will now take a look on the other side of the fence with Connect3D's ATI Radeon HD5770.

Click image to enlarge.

In time-honoured tradition, we shall begin by making the obligatory useless observations about the packaging. Said packaging is larger than that of the GT240, but smaller than the 9800GTX+, which we previously reviewed. We find a truly magnifiqu� and dominant shade of blood-red emblazoning the box, in obvious homage to the Art Nouveau masters of old. The female character and dragon theme are also surprisingly arousing.

Click image to enlarge.

The included peripheral bundle is also similarly reduced, a reminder that the HD5770 is a mid-range card.

The graphics card itself, thankfully
A CD containing drivers for the card
A Quick Installation Guide leaflet, containing instructions in several languages

No display adaptors are provided, a choice that is probably justified by the fact that the card already has connectors for all the standard formats (1x VGA, 1x DVI, and 1x HDMI), but it may still inconvenience those that wish to connect multiple monitors of identical connector type. We experienced a similar situation with the GT240.

The CD, dated February 9 2010, is particularly spartan. It contains merely:

The ATI drivers
BumpTop, a fairly useless 3D desktop environment

The HD5770's GPU is codenamed Juniper, but it may also be recognized as a RV840. Its specifications are very similar to those of the RV770, which can be found in HD48xx cards, but it is in actual fact a new, albeit similar, GPU. It shares with its predecessor 800 stream processors (more precisely, 160 processors working on a 5-dimensional vector, often denoted as vect5 160, which is equivalent to 800 scalar processors). The first change of note is that the Juniper is manufactured on a 40nm process, with the end result of a card that operates with a far more reasonable power consumption and heat level. The GPU has some 1.04 billion transistors, which is considerably fewer than the RV870, which powers the HD58xx (however, the RV870 has twice the number of stream processors, while the RV770 contained 956 million transistors). Due to the increased efficiency of the smaller circuits, the stock frequency of the GPU has been increased from 750 to 850MHz. All of this helps to deliver an increased processing power from 1.2 to 1.36 TFLOPS. Juniper also uses a simpler form of GDDR5 memory, with a 128-bit bus running at 1200 MHz.

The card's theoretical power consumption as announced by ATI is 18W at idle, and 108W maximum load. This corresponds to a reduction of 72W while idle and 52W while at load, compared to the 55nm RV770. We can therefore say that the 40nm process on this card achieves its first goal, but we will see if these figures hold up in practice.

Click image to enlarge.

The architecture of the Connect3D is based around an RV840, as recommended by ATI. The cooling system used, however, is not part of the ATI specification. The heatsink and fan are both imposing, and occupy two slots between them. In addition, there is no facilitation of direct heat removal from the case or air vents.

The card is at least a reasonable size in terms of length, and so should not pose any problems, but the placement of the PCIe power connector is a little dodgy, and may pose some problems to a very few guys that lack the proper wrist flexibility when trying to jack said connector in. Note that the reduction in power consumption is evident here to the naked eye; while the HD4870 uses 2 PCIe 6-pin connectors, the HD5770 will be more than happy with just one.

Click image to enlarge.

As previously stated, the card is equipped with 1x VGA, 1x HDMI, and 1x DVI outputs. The HDMI and DVI outputs are capable of carrying sound, with the aid of the ATI chip.

Installation

The physical installation of the card is fairly routine and should pose no real problem. The card is short, and can be installed in all standard form factor cases and motherboards without too many problems, provided that the board has a PCIe slot capable of accommodating cards with two-slot cooling systems. You may, however, need to re-situate your hard disk if it interferes with wiring up the power connector.

Given the card's overall physical design, the existence of multiple cards in one machine might pose some logistical issues. This is covered in the cooling and noise section of the review, which contains details of problems encountered during the course of the test.

Following the installation of the card, the machine booted without issue, and the card was fully recognized by overclocking utilities, including MSI Afterburner, which we used in the GT240 review.

Benchmarks

The card was tested and benchmarked on a machine with the following specifications:

Intel Core i7 920 CPU (Turbo mode and power saving disabled, clocked to 2.8 GHz)
Asus P6T Motherboard
3GB DDR3-1600 triple-channel G.Skill brand RAM
320GB Seagate and 1TB Samsung SATA HDDs
Windows XP Pro SP3
F@H GPU client - systray 6.23
Fahcore 11 v1.24

Where possible, I will give a single value for the same type of WUs. Benchmarks were made on projects in progress at the time of the article's writing. For convenience in the comparison between cards or clients, the results are given in points per day (PPD) form.

Before we continue, please remember that if you want to optimize the folding performance of your ATI card, you must set certain environmental variables in Windows. In order to do this, I would invite you to read our article on this subject. For these tests, and after various tests in order to obtain the best performance across all types of units, the following variables were used:

BROOK_YIELD = 2
CAL_NO_FLUSH = 1
CAL_PRE_FLUSH = 1
FLUSH_INTERVAL = 256

These are the results:

Click to enlarge image.

Since by this point we have tested several cards on FAH-Addict, we have handily placed the average performance scores of each card in a table for your viewing pleasure:

Click image to enlarge.

Our HD5770 is clearly not the front-runner in the competition, but that isn't to say it's a useless card per se. Nonetheless, we will hopefully see a new version of the ATI core in the near future, that will allow ATI cards to catch up in the folding race.

If you plan to use your ATI card in conjunction with an SMP client (or even a uniprocessor client), please take into account the fact that the GPU core consumes a significant amount of CPU time as well. The use of the BROOK_YIELD variable with a value of 2 aims to free up the CPU to some degree, but it also has an unpleasant side effect, namely when there is a single application using a great deal of CPU power (eg an SMP client), the variable disrupts the GPU core's functioning slightly, which results in less than 100% use of it.

Also, be aware that standard day-to-day use of the machine (surfing the web, pirating movies, downloading porn...) in such conditions increases the CPU activity at the price of reducing the GPU usage.

If you want to use an SMP client on your machine, I would encourage you to start it with the -smp 7 flag (for i7 CPUs), or the -smp 3 flag (for quad core CPUs). With this configuration, the performance loss that affects the SMP client will be mitigated as far as it can be, and your ATI GPU client will run at its fullest potential.

Energy consumption

For these measurements, a Watt & Co. Wattmeter was used. The measured values do include the efficiency losses of other machine components, but for the test, only the GPU client was launched. The following values only show the power consumption of the card.

Click image to enlarge.

When folding, the power consumption of the card turns out to be less than the theoretical 108W maximum announced by ATI. After some investigation, it would seem that this is due to a portion of the stream processors remaining unused, even during intense card usage. This phenomenon is evident in the increase in power consumption along with the increase in the number of atoms in the protein being processed. The case of the p5732-5739 projects remains a mystery; they are the projects that have the largest number of atoms, but the performance and usage of the chip (and thus power consumption) appear to decrease without any logical explanation.

To conclude these tests, the following charts show the effectiveness of the card, expressed in PPD per Watt.

Click image to enlarge.

Again, as we have now tested several cards on FAH-Addict, here is a chart showing the average power consumption of each card tested to date:

Click image to enlarge.

Despite economic power consumption comparable to a 9800GTX+, the poor performance of the HD5770 also strikes at its overall effectiveness.

Overclocking

Here, we have sought to achieve the maximum sustainable clock frequency in Folding@Home by limiting ourselves to using the Overdrive overclocking tool in the Catalyst Control Center. We then checked the performance and power consumption of each WU.

The clock frequencies used for testing were:

Core: 960 MHz (the maximum attainable for this card in Overdrive)
Memory: 1350 MHz

The GPU is perfectly stable at this frequency, irrespective of the software used to benchmark it, be it Folding@Home or the most intensive games. The memory was, however, at the limit of stability, and if pushed further would eventually generate artifacts after several minutes of operation, probably owing to the fact that the memory chips are not cooled directly.

p5740-5747 / 384 points

Click image to enlarge.

p5651 / 388 points

Click image to enlarge.

p4754-4756 / 477 points

Click image to enlarge.

p5732-5739 / 511 points

Click image to enlarge.

Another observation: it seems that with this card, overclocking the memory was completely ineffectual. No performance increase was noticed, and the overall power consumption increases along with a decrease in efficiency.

One final point: the HD5770's GDDR5 has exactly the same behavior as that exhibited by the GT240, and does not appear to have the advanced error correction capabilities which could help it to achieve higher stable clock frequencies.

Cooling and noise

The cooling system remains impressive, and coupled with the card's 40nm manufacture, it is rather effective. When idle, the card maintains a temperature of around 50�C. When in use, the fan control is very linear; an increase of one degree causes a 1% increase in fan speed. In the worst case (GPU overclocked to 960MHz, and a FurMark stability test run) the GPU reached 85�C with the fan spinning at 80% speed. Under Folding@Home, 5-10�C less and a fan speed of 70% should be expected.

As for the noise level, the fan's size enables it to emit only a low hum, which is quite bearable and shouldn't irritate many users. It only really becomes audible when spinning at 80% speed. The only drawback is the sometimes unpleasant vibration produced when the fan changes speed abruptly. This is particularly noticeable at around 55% speed. In comparison, the 5770 fan is quieter than the glorified 747 turbines on the 9800GTX+, while still being louder than the small fan of the GT240.

However, I have a negative comment to make when the card is mounted in the following configuration:

Click on the image to enlarge.

In this situation, the lower card blocks the air inlet fan of the 5770. This has several consequences: firstly, the very thin gap between the two cards (fan chassis might be in contact with the screws or the components of the lower card) may amplify the noise caused by the fan vibration. This is even more problematic in these conditions because the temperatures hit the roof; the GPU temperature reaches 70�C with the fan running between 65 and 70% of its maximum speed. As soon as the card is at full load, it becomes unusable. GPU temperature goes over 100�C, with the fan running at full speed, and the thermal protection comes into play (frequency throttles: 850 then 600, 533, 300 and finally 157 MHz). The card is not suitable for use in such conditions, either for GPGPU or video games.

Conclusion

The Connect3D 5770 is in all good stores, and priced at around €150. This price, although reasonable, does not allow ATI to compete very well with nVidia's cards of equivalent performance. We pay more, however, for the decisions made by ATI in the architectural design of their GPUs, which presents performance troubles in Folding@Home, where relatively simple mathematical calculations can be needlessly intensive and slow for their cards to perform.

At the cost of a sometimes complex configuration process, involving the environmental variables we set early on in the review, daily use of ATI cards on Windows is generally simpler and easier than with nVidia's; you can easily use your machine when FAH is running or while watching videos, or even while playing games, without having to pause the client. I still strongly recommend halting the client, however (the systray client is required in order to pause rather than halt) when you are using an intensive application such as a modern 3D game, in order not to lose the current WU.

Pros:

Low power consumption and only one PCIe power connector
Good overclocking potential
Reasonable temperature levels
Quiet running

Cons:

Relatively poor performance
Excessive CPU utilisation can interfere with SMP2 clients
Excessive configuration in terms of environmental variables
Cooling system makes coexistence with other GPUs impractical

To conclude this test, I think that this card perfectly embodies the spirit of the original Folding@Home: if you don't choose your equipment solely based on whether or not it is perfect for folding, then the HD 5770 is a very good choice for you, based on its general 3D performance and its compliance with Folding@Home.

We hope that ATI's transition to OpenCL is not too protracted, and that they can get back on track and provide some competition with nVidia on the Folding@Home performance front once again.

nVidia GeForce GT240

Fri, 05 Feb 2010 17:23:01 +0100

Overview

In this test, we will be taking a look at a mainstay of many GPU folders due to its power-to-price ratio: the GeForce GT240 from nVidia. The card we'll be using here was manufactured by MSI.

Click image to enlarge.

The box is relatively small compared to the those of high-end cards, suggesting a fairly compact form-factor. Among the many logos and stickers on the box is one stating "Military Class". Don't worry, this card isn't a secret army weapon designed to sterilise your prize rubber cactus; it's just a name used by MSI to designate cards especially suited to overclocking, due to higher-quality components and power management capabilities.

Click image to enlarge.

The contents of the box are kept to a minimum in keeping with the military theme:

The graphics card
A CD containing the graphics drivers
Two leaflets containing truly invaluable installation tips

No cable adaptors are provided, though you're provided with a choice of three connectors: 1x VGA, 1x DVI, and 1x HDMI. This can however be an annoyance for those wishing to connect two or more monitors using the same cable type. For this, you'll have to buy the adaptor(s).

The drivers CD is fairly extensively loaded up with useful content:

The nVidia drivers
The Afterburner overclocking utility from MSI
The MSI Live Update utilities
The CUDA-powered Badaboom video conversion tool
Links to nVidia and MSI's websites
A demo version of PowerDVD 7
A 60-day trial version of Norton Internet Security 2010
A trial version of TMPGEnc 4.0 Xpress

The GPU powering this card is interesting in more ways than one. nVidia has finally decided to phase out its old G92 card and the variations derived from it. The chip powering the GT240 is named the GT215. This card is part of the same wave as the G210 and GT220, and is therefore placed near the top of the mid-range market. The GT240 is positioned as the successor to the 9600GSO and 9600GT.

Click image to enlarge.

The GT215 heralds the 40nm manufacturing process that will be used for the next generation GF100 chip (also known as the GTX4xx). We will see during this test if the process is delivering tangible benefits. nVidia did not see fit to impose a standard manufacturing specification on the memory cards used in the GT 240, meaning one can find versions that have GDDR3 or GDDR5, both in 512 MB and 1 GB capacities. The memory frequencies are not fixed, and similarly vary from manufacturer to manufacturer. This GPU supports Compute Capabilities 1.2, as with the G210 and GT220. CC 1.2 is very close to 1.3, which the GTX2xx series utilise, with the sole exception that the double precision calculations are not supported in 1.2. The GT215 is clocked at 550 MHz and offers 96 CUDA compute units clocked at 1340 MHz. It is capable of delivering 255 GFLOPS in its mandatory single precision mode.

Click image to enlarge.

MSI has opted to equip this card with GDDR5 128-bit memory, which is clocked at 1800 MHz. The cooling system consumes two slots - however, it does not directly expel air from the computer case, even if an air vent is directly facing the outputs.

Click image to enlarge.

Note that the GT240 is the first nVidia card to support audio output over HDMI. Combined with its relatively small size, this card should fit easily into most machines provided they have space for the dual-slot cooler. Finally, the card is fed entirely by the PCIe slot, meaning it is designed to use up to 75W.

Installation

The physical installation of the card into the machine poses no problem. So long as you have a PCIe slot that can accommodate cards utilising dual-slot cooling.

The machine starts up without problems - however, the card is not supported by the version 190.38 drivers that are currently distributed in the CUDA Zone. You'll need to either use the drivers supplied on the CD (non-WHQL 191.07), or grab the latest driver version from nVidia's site. Version 195.62 WHQL was used in this test; and as of writing, the latest available version is 196.21 WHQL). The 191.07 driver was the first to implement CUDA 2.3 support, while introducing support for version 3.0 and the OpenCL 1.0 beta in version 195.62.

The card and/or the 195.62 drivers were not recognized by the current version of Riva Tuner 2.24, so I used the program provided by MSI on the CD to overclock the card during testing. The program is called MSI Afterburner, and runs successfully for most current- and last-generation MSI cards. It is based on Riva Tuner, so you may well find some elements of the program to be familiar. The program has the advantage of being able to change the voltage of GPU between -0.1V (if you want to underclock, or limit power consumption) and +0.1 V. We'll see if this feature proves useful.

Benchmarks

The benchmarks to follow were conducted on a machine with the following software and hardware configuration:

AMD Athlon X2 4400+ processor
DFI LanParty UT nForce 4 Ultra motherboard
1GB dual-channel OCZ DDR memory
160GB Hitachi IDE hard drive
Windows XP Pro SP3
FAH GPU console client, v6.23
Fahcore 11 v1.31
Fahcore 14 v1.26

I will avoid giving single values for the same WU types (they are currently grouped by value of points on the GPU projects page). Measurements are made on projects in progress, at the time of writing the article. For ease of comparison between cards and clients, the results are given in points per day (PPD) form. The value given in all cases is the actual value on the entire unit (Eff. Time / Frame under Fahmon), so you should not take into account variations in time frame for projects using the core 14.

We will begin by studying the influence of the drivers on performance:

Click image to enlarge.

The latest drivers seem to be slightly slower than those supplied with the card, but the performance difference is not great enough to be a significant anomaly. It could indeed fall into the category of natural deviation in the measurements.

For further testing, the 195.62 drivers are used.

Power consumption

For these measurements, I used a Watt & Co Wattmeter. The measured values naturally include the power consumption of other system components, and not just that of the card. For the following measurements, the only application running is one GPU client.

Caution: while running the Core 14 (p59xx) WUs, I did not have a meter capable of calculating the average power consumption over a given period, so I therefore used the formula (max value + min value) / 2 to get an approximation. It is by definition inaccurate, but nonetheless reflects the reduced power consumption that these units afford (the machine spends more time in minimum power consumption here than in maximum).

Click image to enlarge.

The power consumption of the card is very low compared to cards of the previous generation, as the card is capable of vastly reducing its operating speeds when idle (down to 135MHz core/270MHz shader/135MHz RAM). This complicates comparisons with previous generation cards (G9x) and may lead to spectacular results for the idle power consumption. The card also has a medium-power mode (405MHz core/810MHz shader/435MHz RAM) which is used when decoding GPU accelerated video using PureVideo.

When folding, the maximum frequencies will be used, so these low power modes are of little relevance to folders. However we observed some cases where the card did not correctly return to it's full-power mode after certain errors. If your card is abnormally slow, check this. A reboot resolved the problem, which may suggest a driver issue.

To complete these tests at the original frequencies, the following charts show the PPD/W figures for the card, calculating using the consumption of the card.

Click image to enlarge.

Overclocking

Here we have attempted to achieve the maximum speed the card can sustain running Folding@home without producing errors. The power consumption has been measured for each individual component overclock, and various combinations of overclocks.

The frequencies used for the test are as follows:

Shaders: 1836 MHz
Shaders: 1900 MHz with an additional 0.1V
Core: 667 MHz
Core: 667 MHz and Shaders: 1836 MHz
Memory: 2106 MHz
Core: 667 MHz, Shaders: 1836 MHz and Memory: 2106 MHz.

Most of these frequencies would be unstable on cards used 24/7 but they were stable enough for the duration of the test run, running Folding@home. As a precaution for intensive use, we advise that you aim for slightly lower figures than these to avoid damage to your card (based on our observations, lower clocks of 100 to 150MHz for the shaders, 20 to 50MHz for the core and 50 to 100MHz for the memory).

During the overclocking attempts, the MSI overclocking program limited the core frequency to 667MHz whilst the shaders remained at their stock clocks. When overclocking the shaders, this limiter was removed.

p5765-5772 / 353 points

Click image to enlarge.

p5787-5798 / 787 points

Click image to enlarge.

p5912-5915 / 1888 points

Click image to enlarge.

First observation: the 40nm technology works wonders, as the card overclocks very well. Shaders can be overclocked to the same high levels as the previous generation's high-end cards (9800GTX+). Hopefully, the more complex chips of the next generation (GF100) will be capable of similar feats.

Second observation: during our search for the highest frequencies, we noticed a peculiar phenomenon for performance related to overclocking the core and the memory. After an initial gain, we return to the widely held theory that overclocking these sections of the GPU increases power usage with little performance gain for Folding@home, so once again it should be just the shaders that we overclock.

Last point: we were disappointed by GDDR5. We had hoped that it would increase the reliability and performance of the card at higher speeds, but found no evidence of this when folding.

After inquiring (we thank Imran Haque for the detailed explanations), it has transpired that the ECC fitted to consumer cards in GDDR5 is rather more basic than anticipated. It is content with delays for information to be received to the memory, so if it is supposed to arrive in X clock cycles, the ECC will allow X+Y clock cycles before reporting an error. There is no real error correction, just a certain flexibility that should allow higher clock speeds.

This is different to the ECC built into Tesla cards based on Fermi, which have an embedded feature built into the chip to track data through the entire bus to ensure it is correct.

Cooling and noise levels

Despite a fairly lightweight cooling system, this card does not overheat. This can probably be explained by the 40nm technology used in production. At stock frequencies the fan was never observed to go higher than 35% and the GPU never warmed beyond 56�C.

During the overclocked periods, the fan speed increased slightly to 40%, with the GPU at a maximum of 60�C. Even when testing with an extreme voltage increase to the GPU, the chip did not go higher than 74�C with a fan speed of 65%.

As for noise, the cooling system is pleasingly quiet: it was not discernible over the general sound produced by the system in use, even at 65% of its maximum speed.Unfortunately, manual fan control in the MSI Afterburner program did not work, so we were unable to test the noise at 100% fan speed.

Although the cooling system is very discrete, the card sadly is not. Despite using so-called "Military Class" components, this card is another that is plagued by whistling from the capacitors when folding. The whistle is quiet and barely audible at stock frequencies, but it is a shame that a card so obviously marketed at overclockers has this issue when overclocking. The noise increases with frequency, and if the voltage needs to be increased to keep an overclock stable then the whistling quickly becomes very irritating, and unbearable after a few minutes near the machine.

The exact origin of the whistle could not be traced, but it does not appear to be entirely down to the card. When the card was tested in a second machine with a Corsair TX650W power supply and Asus P5E3 Deluxe motherboard, the whistle was inaudible at stock frequencies and much quieter (though definitely still there) when overclocking/overvolting.

This could be attributed to the capability of the motherboard to power the card. When overvolting the card, power consumption increased to 86W which exceeds the theoretical limits of what PCIe 1.1 can provide without requiring a 6-pin PCIe power cable to power the card too. The P5E3 Deluxe supports PCIe 2.0 which can supply 150W to the card, and this extra headroom may have caused the power delivery to the card to be more stable, thereby reducing the whistling.

Conclusion

The GT240 is a perfect card for folders who wish to fold cheaply and effectively, without using too much power or sacrificing performance. This card replaces the 9600GT and 9600GSO in nVidia's pricing structure, at about 85€, and is slightly superior to its predecessors in terms of performance. This isn't including the fact that with a simple overclock the card can produce over 5000PPD, matching the best cards of the previous generation.

Pros:

Performance
Overclocking
Price
Cool running
Easily fitted into the vast majority of cases

Cons:

Very annoying whistling in some scenarios
Stock frequencies that should be revised upwards to utilise the capabilities of the chip more effectively
A 6-pin PCIe power plug may be sorely missed, particularly by overclockers with PCIe 1.1 motherboards, to quieten the whistling.

To conclude the test, I gave the card a go in my main machine, to see what kind of performance it gives in more general day-to-day use.

As in most cases when GPU folding, I strongly recommend stopping your client when you start a GPU-intensive application. This not only reduces the risk WU miscalculations and therefore loss, but in the case of our GT240, an interface lag is present in Windows (XP SP3 being used here). Even if this lag doesn't annoy you, you will probably still encounter problems when watching video, running Flash applications, or playing 3D games.

nVidia GeForce 9800 GTX+

Sat, 12 Dec 2009 18:59:01 +0100

Overview

We now turn our attention to a Geforce 9800GTX+ from Twintech.

The box is nothing revolutionary; par for the course for most Twintech products. The only thing that attracts our attention is the 2-year warranty. Since GPU folding involves far more intensive card utilisation than video games, this guarantee can be quite useful in case of premature card failure.

Click image to enlarge.

After opening the box, we find several peripherals:

the graphics card
a multi-language manual
a CD containing graphics drivers
a DVI –> VGA adaptor
an S-Video -> composite adaptor
an S-Video –> RGB adaptor
a 2-part Molex -> 6-pin PCIe adaptor

Firstly, the user manual is very light on detail, with only 4 pages per language, and contains such sage advice as only installing the card into a PCIe interface. If you'd rather flip through installation diagrams for whatever reason, you'll be disappointed, as there are none present.

The CD is also sparse on content; no software is provided, and all you will find is a set of drivers for the card. Three versions are provided: 175.19, 175.31 and 177.42. In order to use the card for running CUDA code, I suggest you opt to download the updated drivers available in the CUDA Zone.

In terms of connection adaptors, you're spoiled for choice, though we note the conspicuous lack of a DVI -> HDMI adaptor (or indeed any dedicated HDMI on the card).

Before looking at the card itself, we will first study the GPU: an nVidia G92b manufactured on a 55nm process.

The GPU is the latest evolution in a lineage that was begun by the G92 65nm GeForce 8800 GTS 512MB at a core frequency of 650MHz and 1625 MHz for the shaders. The same G92 was then used on the GeForce 9800 GTX, only with a slight overclock (675 MHz / 1688 MHz). Incidentally, the GPU is gaining in frequency, and has reached 738 MHz / 1836 MHz. The same GPU has been used with the same frequencies in the current GeForce GTS 250. All of these cards utilise on board GDDR3 memory on a 256-bit bus.

Click image to enlarge.

As a side note, the G92b has 128 units (also known as CUDA and Shader Processor Cores), programmable using the NVIDIA CUDA API. It supports Compute Capability 1.1 (for reference, the GT200 supports Compute Capability 1.3). In terms of academic performance, these 128 units are clocked to 1836MHz, and deliver 468 GFLOPS single precision, and do not support double precision.

The card is based on the G92b architecture, with the standard frequencies recommended by nVidia, paired with 512MB of GDDR3 memory clocked at 1100MHz (2200MHz effective). Again, the manufacturer complies with nVidia's recommendations. The PCB and two-slot cooling system are also used in compliance with nVidia's reference design. You'll need two 6-pin PCIe connectors to power the card. You have two DVI and one S-Video output on the back of the card. The dual slot fan removes heat through the rear of the box. This will limit the temperature of the housing and improve the cooling of the card itself.

Installation

The card is very long, and may pose problems in some cases. This was the case with my tower; I was forced to attack the hard drive cages with a hammer to bend them enough to allow me to bring the card fully into its slot.

Once installed, the card may also block access to certain connectors, or hinder the cooling of some components of the motherboard (such as in the case of the SATA connectors on the DFI LanParty UT nForce 4 Ultra, which was used for this test).

Click image to enlarge.

In the picture above, one can see the two 6-pin PCIe connectors needed to power the card, as well as the slight adjustment made to the hard drive cage. One can imagine also the chipset fan, which is almost completely blocked by the card.

I also encountered another problem: a rather capricious boot sequence. I previously did some (incomplete) research into graphics card incompatibilities with some motherboards and chipsets. The machine boot got stuck during the system memory test, and one or two resets proved necessary to start the machine. This problem is not present with an ATI Radeon X800 or an HD 4870 installed in an identical configuration. Fortunately, once started, the machine does not pose any more problems, and the card is very stable.

Benchmarks

The benchmarks were performed on the following machine and software:

AMD Athlon X2 4400+ processor
DFI LanParty UT nForce 4 Ultra motherboard
1GB OCZ DDR memory, dual channel enabled
Hitachi 160GB IDE HDD
Windows XP Pro SP3
CUDA 190.38 drivers
FAH GPU 6.23 console client
Fahcore 11 v 1.31
Fahcore 14 v 1.26

I will also include the benchmark reference for Stanford's sake, should we encounter another stats server crash, as well as a reminder that 1500 PPD is the average on an ATI Radeon HD 3850. Wherever possible, I will not give a single value for the same type of WUs (currently grouped by value of points on the current projects GPU). Measurements are made on projects in progress at the time of writing the article. For ease of comparison, the results are given in points per day (PPD). The value found in all cases is the actual value on the entire unit (Eff. Time/Frame under Fahmon), so we do not take into account variations in time frame for projects using the core 14.

Click image to enlarge.

Energy consumption

The following values are taken with one GPU client launched.

Note: when processing the Core 14 (p59xx) WUs, I did not have a meter capable of calculating the average consumption over a given period, so I resorted to using a formula (min value + max value)/2. This formula is probably inaccurate to say the least, but it still reflects the reduced power consumption of these units (the machine spends more time in a state of minimum power consumption than in maximum).

Click image to enlarge.

The graph above shows the energy consumption of the card, without the energy used by the rest of the machine.

To complete these tests in the original frequencies, the following graph shows the effectiveness of the card in terms of PPD per Watt.

Click image to enlarge.

Overclocking

For this test, I'll take measurements with the following frequencies:

Shaders : 1944MHz / 740 MHz core
Shaders : 1836MHz / 775 MHz core
Shaders : 1944MHz / 775 MHz core

Then I will return the chip to its original frequencies (1836 / 740) to test overclocking the RAM to 1200MHz (1100MHz being the original clock frequency) then I will test all three, overclocked (RAM, shaders and core).

I am not trying to push the card to the max, for the simple reason that these cards are deemed structurally vulnerable. I myself have a card that I ended up RMA'ing back to the manufacturer; it wasn't overclocked at all, and deteriorated over time to the point where it refused to fold. In order to avoid risking the same fate with this card, I have deliberately limited the rise in frequency.

The unit used for this test can be classified as intermediate in terms of performance and use of the chip: a p5793 for 787 points.

Click image to enlarge.

The first thing that we see is that overclocking the core does not provide additional performance (or only very little). Shaders are the key elements of GPU folding. Contrary to popular belief, the benefits of overclocking the RAM are not negligible. Beware, however, that on most cards the flexibility of the RAM frequency is often limited, so do not try to go too high.

Click image to enlarge.

The consumption measurements show what we had noticed previously: the core is not too extensively utilised by FAH; consumption is not affected by overclocking this part of the chip. By contrast, it seems that overclocking the RAM allows a performance gain with no negative impact on consumption.

Click image to enlarge.

This confirms that overclocking the RAM is the most interesting in terms of efficiency. Interestingly, in all cases, the efficiency increases with overclocking.

Cooling and noise

During the various tests with the original frequencies, the temperature of the chip remained reasonable, and did not exceed 70�C. In terms of noise, the card is discreet, and the fan did not exceed 65% of its maximum speed. Moreover, the model tested here does not hiss, which is a characteristic of most nVidia cards when running Folding@Home.

However, in a configuration with two of these cards, the fan will spin to 100%. The noise is bearable if the machine is in a living room, but I strongly warn against such a configuration if it is located in a bedroom, unless the user is an insomniac already.

Conclusion

The 9800GTX+ is a very good card, and nVidia has understood this, as shown by the number of versions of the chip that have emerged on the market. This is true also in terms of Folding@Home; with a 9800GTX+ you have a card capable of producing between 5500 and 6000 PPD for a reasonable energy consumption. When I originally bought it a year ago, it was sold new for €189, though these days you will certainly find fairly cheap prices, especially on the secondhand market. If you prefer new equipment, you will find the current version of the card, called the GTS250, for about €100.

The good:

Performance
Silent operation of the fan
Absence in annoying GPU folding noises

The bad:

Size of the card can be a problem when installing
Noise when in a multi-GPU configuration
Some reliability problems in some series

Vince Voelz

Wed, 11 Nov 2009 21:54:01 +0100

Dr. Vince Voelz is a relatively recent addition to the Folding@home team (in 2007). His interests involve studying protein folding in biophysical settings as well as folding in the presence of the ribosome. He is part of the multi-institutional grant from NSF's Frontiers In Biological Research program. While it's a bit early to talk about his work with FAH, Vince has been one of the key driving force behind a major new protein folding initiative at Folding@home (more on that later) as well as working with FIBR experimentalists to see what we can do to predict their experiments and build a tighter connection in general.

Below is a link to a YouTube movie where Vince explains about electrostatics in proteins, a key aspect of protein stability.

Vince's video

Copied from: Meet FAH team member: Vince Voelz (Vijay Pande's blog)

Edgar Luttmann

Wed, 11 Nov 2009 21:51:01 +0100

Dr. Edgar Luttmann has been a FAH group member for several years. His primary areas of interest are Alzheimer's Disease and new methods for describing water. In studying Alzheimer's disease, Edgar has contributed to FAH in several ways. He's been performing simulations of protein aggregation involved in Alzheimer's disease as well as helping understand the experimental work coming from the group (primarily NMR).

In his second research area, Edgar has been working with Jason Wagoner in the Pande group to develop better models of water. These models were originally conceived in order to have a model which was more computationally efficient, but turns out that they may be both faster (by about 10x perhaps in some cases, maybe more) but also it looks like they may be in fact more accurate as well. Finally, these new models are also very well suited to GPU's and PS3's. Bottom line, we're very excited about this and hopefully will be submitting our first results for peer review soon.

Copied from: Meet FAH team member: Edgar Luttmann (Vijay Pande's blog)

Adam Beberg

Wed, 11 Nov 2009 21:50:01 +0100

While Adam Beberg is currently a grad student at Stanford in the Computer Science department and a member of the FAH team and Pande group, Adam has been collaborating with FAH from the very beginning. Adam has experience with distributed computing beyond FAH, as he was one of he founders of distributed.net and from his experiences with that, wrote the Cosm library to help with many areas, including writing distributed computing code.

Adam Beberg

In his thesis work, Adam has worked in many areas, including distributed storage (more on that later), distributed computing code, as well as the GPU code. For now, I'll highlight on his work with GPU's (and will comment on the rest at a later date). Adam has helped make major steps forward in our FAH code for GPUs, taking the code from "academic quality" to something which is very robust and broadly useful. This was a major undertaking, requiring knowledge in many areas including both the science, internals of GPUs, and the whole tool chain. This work has been performed in conjunction with Mark Friedrichs (a programmer in the Simbios center and Pande Group).

Copied from: Meet FAH team member: Adam Beberg (Vijay Pande's blog)

Comparison between linux kernels

Sun, 01 Nov 2009 13:24:01 +0100

Kernel, Kernel on the wall, who is the fairest of them all?

Sadly kernels cannot talk so we shall have to find out for ourselves which is the best!

Frodo decided to test one or two kernels after a fair number of users recently reported that our modified kernel did not improve performance. Frodo also decided to run the test on standard a2 WUs rather than bigadv WUs.

The following test kit was used for the test:

SuperMicro X8DAi
2x Xeon X5560 ES (2.8 Ghz)
6x2GB DDR3 ECC Crucial PC13333
Western Digital Raptor 150 GB
8800 GTS 512 MB

The tests were performed on a single unit from Project 2669. The same unit was used for all tests, and the best time from the first five frames was used to establish a time reference.

Ubuntu 9.04 Kernel 2.6.28-16-generic

So here we have the recently-compiled 2.6.28-16-generic kernel from ubuntu after a shaky start...

Click images to enlarge.

We have thus obtained the following results:

-smp 8: 2m 15s per frame for 12,288PPD
-smp 16: 1m 56s per frame for 14,325PPD

FAHAddictXeon64 kernel 2.6.28-10

So, now we turn to the FAHAddict kernel 2.6.18-10 as modified by Frodo.

Click images to enlarge.

From this, we obtained the following results:

-smp 8: 2m 13s per frame for 12,454PPD
-smp 16: 1m 52s per frame for 14, 785PPD

The difference is small but it is there...
This represents a gain of 1.35% for -smp 8, but a bigger gain of 3.2% for -smp 16 (more than double). These kernels are clearly not designed for dual- or quad-core machines. However in the case of the bigadv units with the bonus, those few percentage points can earn you an extra 2000PPD without extra cost in terms of power consumption.

Ubuntu 9.10 kernel 2.6.31.5

Click images to enlarge.

The time per frame for this core was catastrophic...

-smp 8: 3m 41s per frame for 7,513PPD
-smp 16:9m 38s per frame for 2,871PPD

This represents a fall of 40.7% and 80.6% respectively compared to the modified 2.6.28-10 kernel.

We advise you not to upgrade to 9.10 just yet, but will keep you informed of any future changes.

Optimization for a2 core v2.10

Wed, 21 Oct 2009 22:54:01 +0200

Introduction

The performance problems of the a2 core on dual cores are in the past! "tear", a user of the official folding forums, has discovered a variable linked to the MPI layer which seems to solve the performance problems of the a2 core on native and virtual dual-core machines.

Technically, this variable changes the behaviour of the core with respect to memory management for MPI processes. By default, the Folding@home core uses shared memory to exchange information between processes.This memory is a memory area not reserved for a given process, so everyone can read and write.

After activation of the variable, MPI uses a more conventional memory setup: shared memory is no longer used, with each process allocated a reserved memory space, and the exchanges are performed over TCP/IP (via the localhost in this scenario). MPI is used here in the way it was originally intended: the exchange of data between processes distributed over a network (i.e. a cluster).

Application of the variable

The application of this variable is very simple. Before you launch the client, you must type the following command, then run the client as normal:

Code :


export MPICH_NO_LOCAL=1

This command must be typed in the same window that will be used to launch the client. If you use an automated script, place it before the command to launch the client. If you run the client in a terminal, type and execute the command before running the client (in the same terminal window).

Real-world results

The tests were conducted on the following setup:

VMware workstation 6.5.2, 640MB RAM, 2 cores
Windows XP SP3 host OS
Ubuntu Server 9.04 64-bit guest OS, based on kernel 2.6.28-15-server
Linux core A2 2.10, Project: 2662 (Run 1, Clone 98, Gen 56)
Q6600 processor at 3GHz (333*9), with 800MHz DDR2 RAM (CAS4-4-4-10

Results before applying the variable

Quotation:

[17:14:20] Completed 7500 out of 250000 steps (3%)
[17:27:42] Completed 10000 out of 250000 steps (4%)
[17:41:06] Completed 12500 out of 250000 steps (5%)
[17:55:31] Completed 15000 out of 250000 steps (6%)

Time per frame : 13:22, 14:24 and 14:25.

Click to enlarge the image.

In this configuration, it would seem that the process is assigned to a core (column P is the core on which the process runs), and the other three processes share the second core.

Results after applying the variable

Quotation:

[19:41:23] Completed 35000 out of 250000 steps (14%)
[19:53:42] Completed 37500 out of 250000 steps (15%)
[20:05:59] Completed 40000 out of 250000 steps (16%)
[20:17:53] Completed 42500 out of 250000 steps (17%)

Time per frame : 12:19, 12:17 and 12:54

Click to enlarge the image.

Here, the processes are properly distributed: two per core.

Conclusion

Finally, the gain varies between 28 seconds (3.5%) and 2 minutes 8 seconds (14.8%) per frame. The gain is significant, and the first tests by users seem to imply that performance is back to the level of core v2.08. Considering the simplicity of application, why deprive yourself?

Warning!This optimization has a downside: it will make the core sensitive to network events (disconnected cable/wifi, DHCP renewal etc) which can cause the core to crash and resulting loss of the WU. To avoid this problem on a machine connected via wifi it is recommended to use a fixed IP for your machine.

General Information about Storage@home

Sun, 13 Sep 2009 17:44:37 +0200

The Folding@Home project has been taken to an unprecedented scale. This is the first - and largest - distributed computing project in the world, in terms of raw power. On our side, we the contributors/clients have a small portion of our drive on which to store data files for the project (no more than the current WU and any pending work items), and at Stanford there is the other. Each WU and result file is carefully preserved. The results for a given project are then combined to create the videos of proteins that have been released by the project.

All of this data is kept in storage servers at Stanford. The terabytes are countless; people speak of more than 400TB of valuable scientific data. However, such storage is very expensive, and the power of the projects equipment is increasing, and the PS3 and the GPU has only increased this need for storage space.

The principle of Storage@Home is simple; data derived from the WUs of Folding clients are sent to your PC. When a server needs to access data that you are mirroring, your computer is accessed and the data uploaded.

However, this system requires some forward planning. First, redundant data must be stored on multiple clients, as it would be disastrous to lose simulation data if John Smith had a hard drive crash. Redundancy also allows load balancing, which enables better data availability for servers. The use of encryption, signature data, and a digital fingerprint ensures that the content has not been modified or damaged, and that the sender is authorised by Stanford.