MXM-Upgrade

Preface

When one wants to buy a new graphics card or is considering the purchase of a sparkling new gaming system, be it desktop or notebook, the prudent buyer often resorts to checking benchmarks rather than putting one's trust in the enthusiasm of fanboys or marketing slogans. For desktop parts, this is a fairly straightforward operation. There are numerous sources of benchmark information, often from respected sources such as Tom's Hardware, Anandtech,...

Benchmarks for notebook components are also available in abundance, from sources such as Notebookreview, Notebookforums, Laptopvideo2go and many, many more. Unfortunately, they often all share the same, fundamental flaw. Because GPUs can often not be swapped or if they can no compatible GPU is available, two GPUs are often compared in entirely different systems. Some differences can be taken into account, albeit with a high degree of inaccuracy and uncertainty. One can add or subtract a bit for the system with the slower RAM, different CPU architecture, slower harddrive etc. The influence of these parts is often limited compared to the performance of the GPU, but still. Unfortunately, the problems don't end there. A driver can influence the end results. Perhaps one person's system is running a hungry process in the background, maybe his harddrive is running out of breath,... The amount of possible party poopers is endless and they may all result in a final judgment or choice that is off the mark.

However, as primary and only source of MXM cards on the net, I find myself in a unprecedented and privileged situation. I have access to a a wide range of MXM cards and I have a notebook that can handle them all. The plan to compile a set of benchmarks that is unparalleled and unprecedented as far as consistency, reliability and representability go formed some time ago. At the time I had a 6200 32MB/TC, a 6600 256MB and a 6800 256MB at my disposal, which I judged to be too limited to kick off the project. When MSI Europe offered me the chance to try a 7600 256MB Type II card, I knew it was time. The card will be offered in their upcoming L735 and L745 series of notebooks, aka the M1042 and M1039. I am happy to report that the card runs very stable and performs very well, even though the card I received appears to be an engineering sample as is evidenced by this picture..

While the article was waiting for it's launch on notebook review, a batch of 7900GS 512MB arrived on my doorstep. Needless to say I added benches for them as soon as I could.

UPDATE: Added (partial) x1800 results.

Test System

Fujitsu-Siemens AmiloM 3438
Samsung HDM060II (60GB)
PM 1,73GHZ (2MB L2)
2 x 256MB DDR (553MHz)
1440 x 900 screen resolution
Driver: 84.71

As can be seen, the memory is obviously a weak part in this system. As the Farcry benchmark required more than the available 512MB, I assume some performance was lost there. This is a previous generation system, and hardly the fastest in that category. I leave it up to the specialists to argue about it, but the advantage modern dual cores have over this aging PM is obvious. But in the end, I was looking for an apples-to-apples comparison and that's exactly what you'll get here.

I used the driver of a newer FSC model, the Pi1547, as I know the model is very close to mine. It is not the latest driver, but a stable and good one that supports all the GPUs benched in this article after the .inf file was replaced with the corresponding modded .inf from my good friends over at laptopvideo2go.

A more in depth view on my notebook can be found here.

Tests

• 3DMark01
• 3DMark03
• 3DMark05
• 3DMark06
• Aquamark3
• Farcry
• Metalgear
• X2

The entire range of available 3DMark test suites has been executed (01,03,05 and 06). Aquamark3 was the final synthetic benchmark. I added 3 "real life" tests: Farcry, X2 and Metalgear. Farcry is based on the 'real' game, the rest are free benchmark demos that can be found on the Guru3D website.
I would have loved to run some Oblivion or Prey benchmarks, but unfortunately they both require the retail version of the game so I decided against it.If someone is willing to shell 100€ I'll be happy to add them ;)

I ran all 3DMark suites twice to make sure no large discrepancies exist between runs. This notebook 'features' the dreaded lagging issue, so when the effects of this issue were suspected, the benchmark run was omitted from the results and run again. The Farcry benchmark was programmed to run 5 consecutive times. This has some importance as I noted the first run was always a bit slower. I assume this has to do with parts of the level being loaded during the first run and being available during the following runs. The picture below fueled that assumption.

Note the harddrive LED being lid. During the first run, a lot of disk access occurred. During the following runs it all but vanished. Furthermore, I noted that when running the benchmark as a whole twice, the second run was much faster (>10%) For this reason, I ran all the Farcry benchmarks twice and kept to the second run. The results leveled around 50fps. It was suggested that the engine enables additional eye candy around that point if the GPU can handle it. Perhaps this is also the reason why a second run was sometimes so much better? The first run perhaps allows the engine to asses what is under the hood.The X2 benchmark was run only once but consists of multiple levels. I ran the Metalgear only once because... I was getting bored with benchmarks. All benchmarks results are available for download.

I decided against adding thermal readings as the thermal setup of all cards has not been equal. I ran all cards with the original Fujitsu-Siemens heatpad, which isn't exactly a performance stud as it is and it degraded after being applied an reapplied several times. On top of that, I was forced to run most cards without heatpads between the heatsink and memory BGAs. While this hasn't affected performance as far as I know, it certainly has an influence on thermal performance. To make sure no thermal throttling occurred, I ran 3DMark06 with Rivatuner in the background and checked on the logs after the run.

To make up for that, I added some power readings for each card. The first one is the ampère reading on an idle desktop and the second one is while running the 3DMark06 benchmark. I removed the battery for this test because the battery charger takes some power as well. The displayed or mentioned reading is what I noted as 'average maximum reading'. Not exactly a scientific approach and it is suggested to take these measurements as an indicator, not so much as Final Judgment. To get an idea of power consumption, multiply the reading by 20. As voltage tends to fluctuate under different load conditions, this is once again not a scientific approach but an indication.

6200

Find anything missing?

Yeah, this card sports a whopping 16MB DDR. One single chip. One. The technology that 'enables' this is called 'TurboCache' or 'TC'. The idea is anything but new but regained some popularity after the introduction of PCI Express. While AGP feature less bandwidth from the GPU to the CPU as the other way around, PCI Express allows for the same transfer speed in both directions. This makes the idea of sharing system memory with the GPU slightly less insane, but the fact remains that you're not only looking at less performance for the GPU, but that you're crippling your CPU's performance while you're at it. While this may work reasonably well for expanding 128MB of DDR in the rare cases where more memory is required, it simply stinks for a card with only 16MB of dedicated RAM. As will be evidenced by the benchmarks.

3DMark01: First run 6023. Second run 6004.

Average: 6013

3DMark03: First run 1787. Second run 1791.

Average: 1789

3DMark05: First run 881. Second run 881.

Average: 881

3DMark06: First run 238. Second run 238.

Average: 238

Aquamark3: First run 17613. Second run 17595.

Average: 17604

Farcry: 10,29 fps

X2: 3.211fps

Gunmetal: Benchmark1 average: 4.35 fps Benchmark2 average: 4.49 fps

Benchmarking a slow card is never much fun. Benchmarking a very slow card even less so. That said, it offers some premium benefits not found with other cards. It allows to manually verify framecounts (slowly count the frames) and you can enjoy the craftsmanship the designers have put to use while adding detail and effect to the game. Needless to say that after running hours and hours of benching, these benefits start to pale. Imagine my relief when the final Gunmetal benchmark was run. That benchmark was a very long one and the 6200 logged another abysmal performance.

Imagine my horror when I noticed the Gunmetal benchmark was run at 852 x 480 and had to be done all over again. Anyways...

The Rivatuner report can be found here. Highlights:

• 300MHz Core
• 300MHz(600 effectively)memory
• 32-bit memory bus
• 16MB DDR
• 4pp,3vp

With the 6200, the Amilo consumes about 1,39 amps idling on the desktop. While running 3DMark06, the 'highest average' was 2,48 amps .

6600 - 256MB

The 6600 was a very successful midrange card from the previous generation of cards. It was used in a wide range of notebooks. Pretty much the launch point for the MXM technology, this card was found as both a Type I and Type II card. This one is a 256MB Type II incarnation.

3DMark01: First run 14591. Second run 14516.

Average: 14553

3DMark03: First run 5829. Second run 5867.

Average: 5977

3DMark05: First run 2617. Second run 2619.

Average: 2618

3DMark06: First run 1413. Second run 1431.

Average: 1422

Aquamark3: First run 43655 Second run 43186.

Average: 43420

Farcry: 35,08 fps

X2: 36.323fps

Gunmetal: Benchmark1 average: 23,49 fps Benchmark2 average: 28,31 fps

Not breaking any speed record, the 6600 can still be labeled 'decent'.

The Rivatuner report can be found here. Highlights:

• 373,5MHz Core
• 300MHz(600 effectively)memory
• 128-bit memory bus
• 256MB DDR
• 8pp,3vp

 

 

6800

I shipped my own Amilo 6800 to a costumer to make his life a bit easier, so I now have one of my own 'Amilo Special' 6800's fitted. Works like a charm. This is the 'original' configuration of this notebook, what it is designed for. One problem with the card I sell is that they lack the specific mounting provisions for heatsinks of various notebooks. The second picture below shows an easy albeit hardly beautiful solution. I used a XL screw for the picture. The screw goes through the heatsink and is bolted down. This procedure works just fine but dóes bring the risk of warping the card if too much torque is applied. To the left you'll also notice the two metallic blocks (coils) that protrude from the surface.It are these components that commonly require the heatsink to be dremmeled, a small yet annoying task.

3DMark01: First run 18978. Second run 18777.

Average:19377

3DMark03: First run 9447. Second run 9412 .

Average:9429

3DMark05: First run 4103. Second run 4094.

Average:4098

3DMark06: First run 2102. Second run 2101.

Average:2101

Aquamark3: First run 56866. Second run 56700.

Average: 56783

Farcry: 52,77 fps

X2: 52,866 fps

Gunmetal: Benchmark1 average: 40,45 fps Benchmark2 average: 45.49 fps

The Rivatuner report can be found here. Highlights:

• 375 MHz Core
• 300MHz(600 effectively)memory
• 256-bit memory bus
• 256 MB DDR
• 12pp,5vp

The 6800 logs 3.84 amps on 3DMark06 (3,78 pictured). Interesting to see is how the influence of the GPU on idle consumption is limited at best. Or rather, the difference drops below the level of significance these rudimentary test can give you. I guess this warrants a 'job well done' for the nVidia engineers.

7600

As said, we own a special "thank you" to the folks of MSI for sending us an advanced sample of their MXM lineup. As MXM reported earlier this year, MSI www has plans with MXM. So when I found the card in the mail, I was anxious to find out exactly how far they advanced. When I opened the package, I was a bit puzzled by the fan on the GPU. When I compared the picture some I took earlier this year at CEBIT, things cleared up a bit. While this particular fan can be found on some chipsets on MSI motherboards, it was also on the MXM cards that I saw on the Geminium technology demonstrator. For the time being, please don't ask if and when we'll be able to offer these cards. We currently simply don't know.

 

It can be assumed this particular card served some time or was scheduled to do so on the Geminium platform. While the actual or commercial purpose of the Geminium is not exactly clear, I think it would be a fair guess to assume it serves at least as a lab tool to debug MXM cards for notebooks. Indeed, it allows notebook designers to separate the designs of a notebook and their graphic subsystems. This allows the development of one to advance even if the other is bogged down. .

We removed the fan with some difficulty as it was attached with pretty strong double sided thermal tape. After clearing the drab, the GPU was revealed. A first sign of the fact that I had a good performer in my hands were the memory chips: ready to be clocked at 500MHz (1000Mhz effective). As with most 'bare' MXM cards, including mine, this card did not come with mounting provisions, so I added a simple screw and a bolt. It should be noted that this approach works perfectly well, but as the heatsink is not 'stopped' by any support, there's a risk of bolting it down too firmly and warping the MXM card. Common sense is a robust solution for this. And then... Benchmark time!

3DMark01: First run 19190. Second run 19016.

Average: 19103

3DMark03: First run 9281. Second run 9255.

Average: 9268

3DMark05: First run 4120. Second run 4019.

Average: 4070

3DMark06: First run 2235. Second run 2231.

Average: 2233

Aquamark3: First run 58533. Second run 58481.

Average: 58507

Farcry: 57.15 fps

Never mind the fact the bench says I had a 6800 installed. I can assure you it was the 7600.

X2: 49.413 fps

Gunmetal: Benchmark1 average: 40.55 fps Benchmark2 average: 45.52 fps.

The Rivatuner report can be found here.Highlights:

• 445.5MHz Core
• 500MHz(1000 effectively)memory
• 128-bit memory bus
• 256 MB DDR3
• 8pp,5vp

With the 7600, the Amilo consumes about 1,45 amps idling on the desktop. While running 3DMark06, the 'highest average' was 3,2 amps (3,15 pictured).

The palmrest area does indeed seem to heat up less.

7900GS 512MB

Needless to say we are very glad to be able to offer these cards. There's a waiting list and supply is erratic at best, but at least the show is running!

This one is once again based on the 'Arima Special' formfactor, meaning this will cause additional problems for some and it will offer a unique opportunity for others. Mounting provisions are again lacking and as often seen with other cards some of the components on the card (coils, to be precise) may force you to mill your heatsink with a Dremel or equivalent tool.

Please note the aluminium heatsinks+heatpads on the bottom memory banks.

A more thorough review will shortly be published on MXM-Upgrade.com. At this point, I would like to add however that as the fan did not respond to the rising temperature of the card, I forced the fan using the ACPI functions of Notebook Hardware Control. As such, NHC was always running in the background.

Well, there's little more to say. On to the benchmarks!

3DMark01: First run 21089. Second run 21023.

Average: 21056

3DMark03: First run12610.Second run 12708.

Average: 12659

3DMark05: First run 5807.Second run 5809.

Average: 5808

3DMark06: First run 3083. Second run 3061.

Average: 3073

Aquamark3: First run 65028. Second run 64904.

Average: 64966

Farcry: 58.55 fps

X2: 68.317 fps

Gunmetal: Benchmark1 average: 46,52 fps Benchmark2 average: 48,31 fps.

Average: 47,42 fps

The Rivatuner report can be found here. Highlights:

• 350MHz Core
• 500MHz(1000 effectively)memory
• 256-bit memory bus
• 512 MB DDR3
• 20pp,7vp

The fairly low core clock surprised me a bit, especially compared to the 450MHz the 7600 was running at. Another nice feature is the automatic throttling as featured by the 7900GS... The 350MHz core is supposed to be 25MHz slower as the nVidia default and the 7900GS is reportedly a good overclocker.

With the 7900GS, the Amilo consumes about 1,80 amps idling on the desktop. While running 3DMark06, the 'highest average' was 3,76 amps (3,71 pictured). The reason for throttling clocks while the GPU is idle becomes quite clear. It consumes enough as it is,which is not surprising. Power under load is about the same as a vanila 6800, which is nice for people considering the upgrade

 

x1800 256MB

Obviously, this card was not benched with nVidia drivers, I used Omega 3.8.291 This was probably the most troublesome upgrade I had. There's off coarse the smal Dremel job required to make the heatsink fit, but that's not really a problem. Problem is that the x1800 is not recognized with Rivatuner, does not yield results for Aquamark or Gunmetal and eventually screwed up the ACPI management in my notebook.

At the time of this article, I used a hardware hack to resolve the fan problem. At the beginning of the x1800 benchmarks, I used the NHC hack as described above. However, while benching some nice people over at the Amilo forum pointed out that a full hardware reset resulted in the fan being correctly managed. It looks like it works but I really can't confirm it as Rivatuner can't give temps or a report for that matter. Unfortunately, my system didn't like that reset all that much. I can no longer change power schemes in Notebook Hardware Control or by any other means. That means that the Idle power measurement are not available because they are not representative. With the x1800 out, all went back to normal.

3DMark01: First run 20344. Second run 20337.

Average: 20335.

3DMark03: First run 10505. Second run 10445.

Average: 10475.

3DMark05: First run 5477.Second run 5499.

Average: 5488

3DMark06: First run 3083. Second run 3061.

Average: 3073

Aquamark3: All the resul I got:

Several fresh installs would not work. The entire bench runs just fine,but right before reporting it got the above error.

Farcry: 54.55 fps

X2: 50.595 fps

Gunmetal: No go. Benchmarking begins, slows down, slows down a bit further and then freezes altogether. No results.

No Rivatuner report. The x1800 is not recognized in any version of Rivatuner. Gathering some information:

• 450MHz Core
• ?(? effectively)memory
• 256-bit memory bus
• 256MB DDR3
• 12pp,8vp

Because the x1800 screwed up the performance managing, I could not set the notebook to dynamic switching. As such, I ommited Idle power measurements because they are not representative with the CPU going at the full 1.73GHz. The full load measurement:

Peaking around 3.83, the x1800 fits about in the thermal envelope of the 6800.

Results

The 6800 and 7600 are quite on par. In fact, the numbers differ so little that it is to be considered non-significant. The 6200 is not capable of putting up a fight, a fact that is increasingly highlighted in the more recent synthetic benchmarks.

The 7900GS smokes them all, which is hardly surprising. A mild eyebrow raiser was the relative small difference in 3DMarks01. The same can be seen with the Aquamark benchmark. Both tests don't use the entire feature set of the newer DirectX versions. This most likely holds the new GPUs back a bit as they are optimized mainly for the new DX versions. The x1800 also logs a decent performance, but is clearly not measuring up to the newer 7900GS.

The 6600 takes it's natural place in the order of things.

The synthetic benchmarks are duplicated in the game benches with the 7600 and 6800 trading first places and the 6200 not even in the picture. The 7900GS is clearly king of the hill

The Farcry benchmark was originally a bit of a surprise, especially regarding the better GPU's. It would seem (or so I was told) that the Farcry engine throttles down on additional environmental elements for lower GPU's and adds a few more if the framerate allows this. Not an expert in this field I can't confirm this, but it sounds logical.

The result of the x1800 in the X2 benchmark came as another surprise. The engine would not seem to like the x1800 much. The 6800 trumps it and the 7600 is closer than it should be.

The 6600 again logs a very logical result.

It came as a bit of a surprise that the Idle measurements differed so little, except for the 7900GS. These numbers would seem to suggest that for normal webrowsing, text editing, watching movies etc will not affect battery life that much altogether. We are looking forward to log more measurements to check this out. The measurements under load only acknowledge what one can assume, with a difference of about 26Watts between the 6200 and 6800. The 7900GS nicely fits the thermal envelope set by the 6800 and so does the x1800