Right after the Centrino buzz, we reported in March of ATI’s and NVIDIA’s new mobility graphics processors. This got the mobile industry excited because for the first time, we were told that full high performance DX9 mobile parts were going to be released to the market. No small news either, since they were going to bring the relative performance of desktop systems to mobile systems.

It has been a bit hard to get together a Mobility Radeon 9600 and GeForce FX Go5600 head to head because of the odd product cycles that have been going on. Toshiba was the first to release a GeForce FX Go5600 based system, but this was on the Japanese product cycle, which is around spring/summer. Since the relationship between Japan and North America are intertwined via manufacturers, the US and Canada saw units trickle into their marketplaces. Meanwhile, Europe was really the earliest to see production systems that were based on the Mobility Radeon 9600, but due to marketplace relationships, North American didn’t see any of these products retailed. Mobility Radeon 9600 in North America wasn’t really seen until VoodooPC’s Envy M: 460 hit the market, which was just several weeks back.

Required or not these days, students back in North America are commonly buying notebooks for school use, and NVIDIA’s marketing has decided to go after the back-to-school cycle with their GeForce FX Go5600. Meanwhile, ATI’s Mobility Radeon 9600 will be aimed toward the US fall refresh cycle (sometime around late Q3 and early Q4), which means we will see more design wins in the near future. Though, at the moment, both graphic processors can be considered shipping components. Either way, the two cycles have lead to a shifted timeline between Mobility Radeon 9600 and the GeForce FX Go5600. With GeForce FX Go5600 arriving earlier than its competitor, it was sometimes unfairly compared to Mobility Radeon 9000 (code named M9). M9, though, was not a DX9 part nor supported AGP8X.

Today, we have the benchmark results to show for all of the countless hours. Not only do we have Half-Life 2 for viewing pleasure, but we will throw in the anticipated AquaMark 3. This should give you the full spectrum look into the latest and greatest from NVIDIA and ATI, with our look into full DX9 desktop and mobile graphic processors. You may have seen other media report benchmark scores [for these two mobile parts] that have been called into question, specifically involving odd margin results. In our time spent benchmarking the two mobile graphics processors, we have yet to be able to recreate similar scenarios.

ATI - Mobility Radeon 9600

As you know, the Mobility Radeon 9600 (code named M10) and Mobility Radeon 9600 Pro (also code named M10) are mobile versions of the desktop Radeon 9600 (RV350). Seeing as we have already gone over that graphic processor, we won’t bore you with the details (read more here). Recapping from our last mobile graphics preview, the Radeon 9600 (including both mobile version) are produced on a 0.13 micron process, incorporates some serious memory controller optimizations, a new version of Hyper-Z compression technology, and support for component output. Power specifications, according to ATI technical documents, notes that the Mobility Radeon 9600 runs at 1.0V, and consumes 0.5W in Windows idle. We have still not been able to track down the maximum power consumption specification.

Since our preview, not much has changed in the clocks of the Mobility Radeon 9600. The official clock speeds are still at 350MHz core clock and 300MHz DDR memory clock (600MHz effective). So far, the highest frequencies in a shipping system that we have been notified of are 350MHz core clock and 270MHz memory clock. Despite market issues, we have been able to take a look at a Mobility Radeon 9600 system clocked at 350MHz core clock and 250MHz core clock.

We have had reports of memory timing and AGP issues with the Mobility Radeon 9600 A13 revision, and lately, we were informed that this problem was related to non-Intel chipset based mobile systems. But even with a non-Intel chipset based mobile system, the problems weren’t guaranteed to appear. ATI’s A14 revision of this chip solves this problem, while we should note that A13 will not encounter any problems of this nature on an Intel chipset based mobile system.

We covered Mobility Radeon 9600 Pro in past mobile graphics preview, which features OVERDRIVE, inherently higher clock speeds, and GDDR2-M, but we have still not been able to see any designs that feature this product. We are expecting to see some desktop replacement and high-end gaming notebooks that feature Mobility Radeon 9600 Pro sometime in early Q4. For a more detailed look into Mobility Radeon 9600 and Mobility Radeon 9600 Pro, read our original preview.

NVIDIA – GeForce FX Go5600 and GeForce FX Go5650

For high-end mobile systems, NVIDIA brings the GeForce FX Go5600 (NV31M), which is appropriately based on the GeForce FX 5600 (NV31), the specs of which can be found here. The specification of the Go version compared to its desktop big brother is virtually the same. As we cited in our preview, it is produced on a 0.13 micron process, is a full DX9 part, shares a number of similarities with the NV30, and consumes 1.0V while running.

The original GeForce FX 5600 Ultra graphics processor, which includes the Go part, are supposed to be clocked at 350MHz core clock and 350MHz memory clock (700MHz effective). However, we have only been aware of original GeForce FX Go5600 shipping at the highest of 270MHz core clock and 300MHz memory clock. The reason why we use the term “original” is because the GeForce FX Go5600 has undergone a refresh. Like the GeForce FX 5600 Ultra, the Go part has two versions now. One is the GeForce FX Go5600 and the other, the GeForce FX Go5650. The first of the two is based on the wire bond design, and the second is the flip-chip. The flip-chip design allows the GeForce FX Go5650 to hit higher frequencies when compared to its predecessor. The flip-chip version of the desktop GeForce FX 5600 Ultra is clocked at 400MHz core and 400MHz memory (800MHz effective). As of late, we have only seen a shipping version of the GeForce FX Go5650 clocked at 325MHz core clock and 295MHz memory clock. So basically, there are two versions of the desktop NV31 and the NV31M. Though, the desktop version doesn’t have a new name incantation like the mobile line. NVIDIA has informed us that they have already started the transition, and we expect to see all future mobile systems (and those announced within a few weeks) to hit the shelves with the new flip-chip version of NV31M.

In the GeForce FX Go5200 and GeForce FX Go5600 preview, we mentioned the support for component output and a MPEG2 decode assist engine. However, we still have yet to test these two features, and will report back once we do, accordingly. For a more detailed look into GeForce FX Go5200 and GeForce FX Go5650, read our original preview.

The Test

For NDA reasons, as we have cited in previous reviews, we can’t go into details when describing the platform we tested. The only thing we can reveal is what is cited below, and that the system has a modular GPU design. This allows us to isolate the GPU in the same manner as a desktop system.

The term “qualified” has been tossed around, and we are sure that it will be used when discussing this review. Unlike desktop components, the graphic vendors can’t validate their own graphics processors for mobile use. The final approval goes to mobile system vendors like IBM, Dell, HP, etc., which is based on bidding price, heat/thermal emissions, and power consumption among other things. Even though validation comes into question when making this comparison, we do it because these two graphic processors, in and of themselves, are shipping in systems.

Windows XP Professional Notebook Test System
Motherboard(s)	855PM DDR
Memory	512MB PC2100
Video Card(s)	ATI Mobility Radeon 9600 128MB 128-bit DDR NVIDIA GeForce FX Go5650 128MB 128-bit DDR
Ethernet	Onboard Ethernet Adapter
Operating System	Windows XP Professional SP1
Video Drivers	ATI 7.93 NVIDIA 44.82
Benchmarking Applications	Blizzard Entertainment Warcraft III: The Frozen Throne – High Quality Settings Microsoft FlightSim 2004: Century of Flight – High Quality Settings Edios Tomb Raider: Angel of Darkness - High Quality Settings Ubi Soft Splinter Cell – Normal Settings Valve Half-Life 2 – Normal Settings Massive AquaMark 3 – High Quality Settings

We should, additionally, note that we used DirectX 9.0b in our system configuration.

Warcraft III: The Frozen Throne

Warcraft III isn’t very graphics intensive nor does it use DX9 in any form, but since it is one of the most popular games on the market, we use it for our benchmark purposes. There isn’t a built in benchmark for this title, so we used a frame counter on eight minutes of game playback.



The scores show Mobility Radeon 9600 with an 86% lead on the GeForce FX Go5650. This clearly isn’t small peas, but we should note that one can play Warcraft III with a GeForce FX Go5650 quite comfortably. However, as we increased the AA and AF settings, the differential between the Mobility Radeon 9600 and the GeForce FX Go5650 increased, with Mobility Radeon 9600 leading the way most of the time. This is something we commonly saw throughout many of our benchmarks.

Microsoft FlightSim 2004

Microsoft FlightSim 2004 is another popular game on the market and, basically, uses DX8 pixel shaders on a DX9 API. This is another title that didn’t have a built-in benchmark utility, so we had to resort to a frame counter on a demo run of the game. The scenario is one in which the flight is carried over a city, and city buildings appear on the horizon as you get closer.

In Microsoft FlightSim 2004, the GeForce FX Go5650 was able to close the performance gap to 26% when we turned on AA and AF. Without high settings, the Mobility Radeon 9600 was able to pull about a stark 105% lead over its competition. It seems in this title, the GeForce FX Go5650 was able to handle AA and AF better than its counterpart.

Tomb Raider: Angel of Darkness

Tomb Raider: Angel of Darkness is one of the first playable titles to use DX9’s pixel shader 2.0. The title has a built-in benchmark, but it auto detects hardware settings and selects the optimal quality settings for the best game play. In this case, the GeForce FX Go5650 was auto-detected, and the game selected lower quality settings than it did on the Mobility Radeon 9600. We ran the benchmark in four different settings to give an idea of the different code paths, and the respective ability of each graphics processors to run through each scenario. As the character [Lara Croft] ran through the pipes and waded through water, the image quality of each scenario reflected the settings we set.

As we bump up to higher and higher code paths, we see the differential between the two mobile graphic processors increase, as well. The Mobility Radeon 9600 takes a 51%, 88%, 96%, and a 181% lead, respectfully. When we hit the DX9 code path, the scores on both ends get to be extremely low and actual game play becomes unreasonable. While Tomb Raider: Angel of Darkness is DX9, it is still nothing like the use of DX9 in Half-Life 2. Read on to see those benchmarks.

Splinter Cell

We used Splinter Cell in the past briefly, but AA and AF aren’t recommended by the creator because of the way it was designed. Even so, the game is still very detailed and takes its toll on even the newest of hardware.

We ran the cards through all benchmark scenarios, but to avoid repetition, we will only include benchmark scores from two of the scenarios. As we push up the resolutions, the Mobility Radeon 9600 increases its 10% lead at 1024x768 by another 6%. At 1600x1200, this scenario hits sub 20fps for both graphics processor, and game play is, for the most part, unreasonable.

The largest difference between the two cards ended up being 28% at 1600x1200, but neither card ended up with sub 10fps scores in any of the scenarios.

We should note that the GeForce FX Go5650 was able to decrease the differential between the Mobility Radeon 9600 in one of the benchmark scenarios. However, the difference between the two only started at 6% at 1024x768, and ended up dropping to 3% at 1600x1200.

One glitch that we stumbled upon in this title was the interference of SpeedStep with the internal benchmark counter. With SpeedStep enabled, the game derives the wrong processor frequency, and therefore, reports the fps wrong. The developer has since been informed of this.

Half-Life 2

We have all been waiting for this title to hit the market, as it employs DX9 to almost the brim. The pixel shaders employed in this title alone can make one awe. For cost reasons, this engine is likely to be more attractive to gaming developers, as the DOOM III engine costs nearly 1 million dollars US. Valve has taken a different approach by slashing their prices down, and has asked for higher royalties for their license to make up for the lower upfront cost.

From the code standpoint, NVIDIA’s NV3x has had to take a lower code path in HL2, which Valve had to incorporate specially into their design, so that reasonable game play would be achieved on NV3x based cards. You can read more details relating to Half-Life 2 in Anand’s coverage.

In our mobile coverage, we forced on the DX9 code path, 32-bit depth, tri-linear filtering, and other high settings for which NVIDIA hardware would not automatically allow. For this review, we ended up deciding to keep AA and AF turned off because Half-Life 2 is a very intensive game with the use of pixel shaders to match. The scores that we are reporting, however they may raise a brow, are reflective of game play. The version of Half-Life 2 we used was source v0.4.

Shockingly, only 3 out of 8 times was the GeForce FX Go5650 able to surpass the 10 fps barrier. Even at its best, the GeForce FX Go5650 was only able to close the gap between the Mobility Radeon 9600 to 234%. “Slow as a pregnant yak” was a phase that we often heard in reference to these scores. While we wouldn’t put it in this exact context, the Mobility Radeon 9600 beats the GeForce FX Go5650 “no questions asked” in all of these scenarios, with the highest difference of 415% (36.6fps vs. 7.1fps).

What more can be said? The Mobility Radeon 9600 comfortably passes through all benchmark scenarios easily and never hits sub 30 fps, but this is putting everything lightly. For a mobile system that uses a Mobility Radeon 9600, Half-Life 2 won’t be a software title that is intimidated by this. Like its desktop counterpart, the Mobility Radeon 9600 has the wits to match even the best offering of NVIDA in this benchmark. Meanwhile, the GeForce FX Go5650, as per Valve’s recommendation, will need to run in a lower codepath (DX8 with lower quality settings) to attain reasonable game play.

AquaMark 3

From the developers of Massive, we have AquaMark 3 to show and tell. This is another highly anticipated DX9 benchmark title that is based on the krass-engine, which was used in their previous AquaNox titles.

For the purpose of our review, we turned the highest settings on (including 4xAA and 8xAF) and ran the resolution at 1280x1024 throughout. There are nine different chapters to AquaMark 3, which test everything from the particle system (particle upon particle texture build) to the terrain system (aquatic plants and the like as you pass over the bottom of the ocean floor). According to the faq, "AquaMark 3 approximately generates 30% of it's screen pixels (without overdraw) through ps2.0." The detail is quite impressive and can be visually seen in the bubbles of the water that are shown, as well as the light reflectivity in the bubbles of water. However, the detail is still shy of what we saw in Half-Life 2. There are explosions related to tanks and underwater craft firing projectiles, and the explosions are smoother than what we have seen many in DX8 titles. Time permitting later, we will explain this in more detail, so stayed tuned.

The scores that we achieved in AquaMark 3 are similarly reminiscent of our scores in Half-Life 2 but without such large margins. In AquaMark 3, the GeForce FX Go5650 achieves sub 10 fps scores in all but one of the scenarios. Meanwhile, the Mobility Radeon 9600 on the average is situated in the mid teens. Minimally, though, the Mobility Radeon 9600 shows its clear lead over the GeForce FX Go5650 with a 58% lead. At its best, the Mobility Radeon 9600 doubles the margin between its counterpart, and this just reinforces the GeForce FX Go5650’s trouble in true DX9 benchmarks.

Conclusion

With what turned out to be not one, not two, but almost a six-month wait, we finally got the head-to-head we were looking for. And with the scores in mind, we are extremely pleased with the way Mobility Radeon 9600 turned out. It seems definitely ready for the next generation games and benchmarks. In our various benchmark runs, we were even able to roughly gauge the heat emission between the Mobility Radeon 9600 and the GeForce FX Go5650. While we can’t release full results, we can state that in our Half-Life 2 benchmark runs, the Mobility Radeon 9600 was able to noticeably generate less heat. We are still waiting for all battery consumption benchmarks to finish, and we will report back as soon as that is completed.

Results aside, it was a bit frustrating to see NVIDIA and ATI take so long to get the chips to market. After all, we reported back in March on these two solutions, and it took us quite some time (albeit almost 6 months) before we started to see real tangible retail systems. Granted, they were in other overseas markets, but the main technology market is still North America.

ATI isn’t completely without fault, as their product announcement comes after their tradition of the Mobility Radeon 9000, which was touted as the first mobile graphics chip to be announced and shipped within a week. Hopefully, we will see the next generation of mobile graphics processors (M11 and NV36M) with an announcement much closer to their full market release. (Of the two, we have only been able to see M11, which is definitely something to keep your eyes peeled for as we near official announcement.) Ideally, each company’s marketing should hold off until the date nears, and not jump the gun to respond to the other.

With the GeForce4 4200 Go ultimately replaced by the Go56xx, NVIDIA is starting to head in the right direction. Power consumption and heat emissions for the GeForce FX Go based notebooks have succeeded in many things for which the GeForce4 4200 Go did not. However, NVIDIA has fair way to go to take their mobility graphics processors up to the same speed as Mobility Radeon 9600 in many of the next-generation games on the horizon.

The developer of Half-Life 2, Valve, is the first developer to voice their displeasure for the NV3x architecture with such intensity, because it has forced them to write additional codepaths particularly for NVIDIA hardware; thus, costing them time, money, and extra resources. This was something not needed to run on ATI hardware, which is why they entered into an agreement with ATI. The order of the agreement was based on already existing hardware benchmark scores to a marketing agreement, not the other way around as some have speculated.

Now, the only way for NVIDIA hardware to run reasonably well in full DX9 games such as Half-Life 2, AquaMark 3, among others, is to lower several image quality related settings: no fog, 32-bit dropped to 16-bit, low dynamic range, etc. The current selection of older DX8 games may suit the GeForce FX based systems (desktop and notebook) just fine, but we are on the heels of a software change to DX9, which is why we are in the process of revising our graphics benchmark suite. The result of GeForce FX benchmarking in DX8 is that consumers are getting use to the higher fps rates in UT2003 and Jedi Knight 2. If Valve didn’t program a special codepath for NVIDIA hardware, customers would be calling up their technical support, and ultimately sending back the software title (RMA issues), which would result in Valve's loss of money. This ends up leaving both the programmer and the NVIDIA consumer dissatisfied because neither side gets to see the full DX9 experience appreciated. Don’t forget that programmers are also artists, and on a separate level, it is frustrating for them to see their hard work go to waste, as those high level settings get turned off. We can’t even begin to hypothesize or speculate the performance results for Go5200, which is a full DX9 part, had we sought to include it in this review.

Update 9/17: We are finished with the battery consumption runs, and we can report back that there is no noticible difference between the two mobile graphic parts, in this respect. We ran both under the highest battery conservation settings (PowerPlay and PowerMizer) and the standard MobileMark settings. Due to NDA reasons, we cannot release the numbers, but the margin between the two result were negligible.