Tech-Hounds.com

Scaling: GeForce 7900GS vs Radeon X1950 Pro

With a new platform, we think the time has come to establish a new base of performance. Something to compare to for future reviews. The move to the new platform brought about several changes - moving from AMD to Intel, the use of dual core processors, a new motherboard and chipset, DDR2-SDRAM and onboard sound. Though the advantages are many and significant, we're still in the process of getting used to this new system. Our last article delves on our experiences with the Core 2 Duo E6300 and the Intel P965 chipset. Now it's time to look at other components of the system.

Of course, the next important component crucial to gaming performance is the graphics card. After all, what's the point of having a new fast system - processor, chipset and memory - if you do not have a powerful enough graphics card to keep up with the processor or vice versa. Now, every gamer knows that choosing a graphics card is a little harder than choosing a processor. Unlike a processor, graphics card have some 'influence' on output quality. With a processor, the main concern is usually just how fast the processor can run games or benchmarks. Output quality between one processor and the other are identical down to the last bit. With graphics cards, we have a lot more to be concerned about: raw performance, performance at higher resolutions, performance with and without anti aliasing (AA) and anisotropic filtering (AF), image quality, graphical features, shader support etc. That's outside the many hardware oriented nature like heat dissipation, temperature, noise levels and power consumption that also applies when we're choosing a processor.

The introduction of a new operating system - Windows Vista - also brought on a new graphics API - DirectX 10. At the moment, the only add-on graphics cards supporting DirectX 10 are NVIDIA's GeForce 8 series, all of which cost above USD 350, including the newly launched GeForce 8800GTS 320 MB. ATI have yet to release their DirectX 10 lineup. However, it will be a while before DirectX 10 (and Vista) games start to emerge and possibly even longer for DirectX 10 compliance hardware to become the requisite.

Vista only complicates matters further because the new operating system needs a new set of drivers and writing good, stable and high performing drivers is never an easy task. It's only just recently NVIDIA released WHQL drivers for Vista. From early previews, there are some notable features missing from these drivers - the most glaring is the lack of SLI support for all GeForce 6, 7 and 8 series of cards. In addition to that, performance with Vista can vary widely between faster, similar to lower than Windows XP. So, for the time being we've decided  not to use Vista, prefering to give it and the drivers more time to mature.

Despite all of that, the arrival of a new API and products supporting it always complicates the decision of which graphics card we should buy (even if we're still using Windows XP). Do you buy the latest, greatest hardware and become an early adopter, paying a premium for that part or do you stick to the tried and true older hardware, costing less but may not be up to the task of running newer games? These are the question we have to ask ourselves when we were choosing the graphics card to use for our new platforms. More likely than not, these are also the questions facing gamers everywhere right now. NVIDIA past mistake - the GeForceFX - still haunts gamers everywhere and the manufacturer themselves to this day. This is why some gamers choose to adopt a 'wait-and-see' position on DirectX 10 and the new GeForce 8 series.

Of course, there's also those who just like to keep their options open. Rumors of ATI's first DirectX 10 line of products, from the high-end R600 to the entry level RV610 are swarming the web as we speak. The latest news as of the time of this article reveal that AMD have changed its plans and will launch their products on the second quarter rather than Q1 as originally planned. Perhaps waiting for Vista demand to pickup (contrary to what Bill Gates are saying, Vista have not prove to be a sales driver for new PCs and components).

We think that for the moment, the toughest competitors to the GeForce 8 series is not ATI's next generation of cards, but their current X1900 series. Contrary to what some people believe, most users do not buy high end graphics cards such as the GeForce 8800. Mainstream cards costing about USD 200 - 250 with just the right mix of price and performance are the mainstay of both manufacturers sales. NVIDIA still retains the GeForce 7900 / 7950 series for this market, just like ATI still has the Radeon X1950 Pro / GT. Rumors are already circulating that NVIDIA plans to release mainstream DirectX 10 parts very soon. Performance reviews of the Radeon X1950 Pro puts in par with the GeForce 7900 / 7950 series of cards, of which the GeForce 7900GS shares pretty much the same price.

So today, we will be running test on both the GeForce 7900GS and the Radeon X1950 Pro to see which card provide the more 'omph' for your hard earned, gaming dollars. However, since these cards are basically derivatives of the GeForce 7900GTX (some of us like to think 7800GTX is more appropriate) and Radeon X1900XT, we won't be doing any iamge quality tests today. Image quality output from these cards should be the same as their high end siblings. Today's article will focus on scalability or the scaling capabilities of these two cards (in addition to seeing their performance in games).

Driver Problems

Before we begin, we have several issues we have to get out of the way first. During the process of testing for this article, we've encounter several problems and annoyances with graphics card drivers from both ATI / AMD and NVIDIA. We briefly mentioned the Catalyst 7.1 bug with Homeworld 2 in our last article. This bug causes huge performance drops when we enable Shadows with Homeworld 2. NVIDIA's current drivers - Forceware 93.71 do not have this problem, but we think to make comparison possible we should be using the same settings for both cards / drivers. We also mentioned another performance limiting bug on both drivers with F1 Career Challenge that makes it impossible to use this benchmark for today's test.

There are some other performance 'aspects' we're still trying to figure out. It could be the drivers or related to the new platform we're using. While running F.E.A.R's Performance Test, we notice minimum frame rates with the new platform are actually lower than waht we got with our old platform - the single core Athlon 64 3500+ at default clocks with the NForce 4 SLI chipset. Our first thought was this is a driver related matter, but this behavior is present on both cards. Our second suspect is the onboard audio drivers, which we're still looking into. However, the load should be negligible with dual core processor such as the Core 2 Duo E6300, particularly if it's running at 7 x 400 MHz. This was not the case - both cards scored lower minimum fps on F.E.A.R Performance test although we're running the E6300 at 2.8 GHz. We're pretty sure it's not due to graphical limitations of the card since the same card - GeForce 7900GS - did offer higher minimum frame rates with our old platform.

Another performance 'aspects' we saw was some very annoying stuttering in some games (F.E.A.R, Serious Sam II) with the Radeon X1950 Pro. We also saw no stuttering before when we paired this card with our old platform. With 2 GBs of RAM, there should not be any memory or storage related stuttering, so we were very puzzled by this. Thankfully, we found what was causing the stuttering, which we will elaborate further later on. Although newer drivers are already available from both manufacturers (Catalyst 7.2 and Forceware 9x.xx WHQL), the bulk of testing we're already finished by the time these drivers are out, so we will be sticking to the older drivers for this article.

Catalyst AI Peculiarities

After some additional testing, we found that the culprit for the second problem we descibred above (stuttering on the Radeon X1950 Pro) was actually Catalyst AI. As you may know, Catalyst AI enables some performance enhancing application specific optimizations. From past experiences, not all application profiles offer a performance increase in our benchmarks. If you look back at our Crossfire article (which you can read here), we found the embedded profile for Star Wars: Knights of the Old Republic was actually producing slightly lower performance. This again seems to be the case with the stuttering / choppiness problems we're having today with F.E.A.R and Serious Sam II.

So, we reran the test both with Catalyst AI set to standard (the default setting) and disabled. You can see the results below. In case you're wondering, the settings and resolution are chosen because they prove to be the crossover point between graphics and system limitations. This means with we can still see some gain with a higher clocked processor without hitting the limits of the graphics cards.

	Catalyst AI Disabled	Catalyst AI Standard	Catalyst AI Advanced
Call of Duty - 1600 x 1200
Min	126	137	141
Avg	253.448	275.286	279.026
Max	525	642	552
Homeworld 2 - 1600 x 1200
Min	164	165	159
Avg	382.762	387.768	387.181
Max	675	679	679
SW: KOTOR - 1024 x 768
Min	49	36	44
Avg	98.284	85.646	89.560
Max	121	110	110
Richard Burns Rally - 1024 x 768
Min	183	184	184
Avg	277.039	276.854	277.085
Max	387	393	393
F.E.A.R - 1024 x 768
Min	46	48	41
Avg	103.032	117.813	116.495
Max	243	289	289
Quake 4 - 1024 x 768
Min	83	97	96
Avg	144.825	164.524	165.405
Max	227	234	238
Serious Sam II - 1024 x 768
Min	74	72	60
Avg	89.924	90.523	87.291
Max	114	117	114
Oblivion - Tree Test - 1024 x 768
Min	39	39	39
Avg	44.852	44.292	44.305
Max	56	54	54
GTR 2 - 1024 x 768
Min	61	61	60
Avg	154.420	154.644	153.821
Max	260	261	261
NFS: Carbon - 1024 x 768
Min	43	43	40
Avg	56.253	56.141	55.081
Max	70	75	74

We can all see that the embedded profile for SW: KOTOR still offer slightly lower performance. Other notable differences between the settings, particularly between disabling and enabling Catalyst AI are Call of Duty, F.E.A.R and Quake 4. You can also see the differences for these games between 'Standard" and "Advanced" are minimal and not significant. What's significant is that running F.E.A.R and Serious Sam II with Catalyst AI set to "Standard" fixed all or most of the stuttering problems. For the rest of the article, we will be setting Catalyst AI to "Advanced", mostly to keep consistency if we were to benchmark any Crossfire configurations later on. Since performance differences are minimal, the results taken should not be any different if we were to set Catalyst AI to "Standard".

CPU and GPU Scalability

What do we mean when we talk about scalability, be it on the CPU and GPU? Well, simply put, scalability is the rate at which performance scales with clock increases. For example, Core 2 Duo scales well with clock increases.However, when we talk about CPU and GPU scalability, we're not talking about increasing their clocks - we're talking about will the graphics card be able to keep up - offer a performance increase - if we were to use a higher clocked processor. The overclocking potential of Core 2 Duo platforms allows us to test GPU scalability. If a card performs better - offer higher frame rates - with a higher clocked processor, it means the benchmark is still system limited. Another, more interesting way of looking at this is that we have not hit the graphical limits of that particular graphics card.

Why is that important? Well, that's related to the next question - why do we care about scalability at all. Well, the answer is again very simple. Having a graphics card that scales nicely means we will still be able to enjoy higher frame rates should we opt for a processor upgrade down the road. So, there's still a valid reason to hang on to that graphics card. With multirendering solutions like Crossfire and SLI, we can opt to get a second card similar to what we have rather than buying a new, faster card.

Generally, articles about scalability only uses average fps to measure performance and increase in performance. We don't think this approach offer a complete picture of scalability. Games and benchmarks vary in loads, sometimes we are more system bound, other times we are more graphics bound. There will be cases where the average frame rate will fail to indicate any increase from the use of a higher clocked processor. However, when we also look at the minimum and maximum fps, chances are we can see a difference. If minimum fps increase with the use of a higher clocked processor, then that benchmark still scales well - assuming the point where we experience minimum fps are the same. After all, having higher minimum fps usually means higher low frame rates, which should help the gaming experience considerably.

Performance

We ran tests with three different configurations - at Core 2 Duo E6300's default clock (1.86 GHz), then overclocking it to 2.4 GHz (6 x 400 MHz) and 2.8 GHz (7 x 400 MHz). Memory clocks were set to DDR2-800 and timings used were SPD timings. To keep relevance to actual gaming settings, we benchmark the games using our usual graphical test settings, not lowered settings to 'induce' system limited situations. Older games like Call of Duty, Homeworld 2 and Richard Burns Rally are very system limited with fast graphics cards, even with cards such as NVIDIA GeForce 7900GS and ATI / AMD Radeon X1950 Pro. This represent the 'optimal' scaling test. Newer games such as F.E.A.R, Quake 4, Serious Sam II may behave differently as we will see later on. We also included two new games in addition to The Elder Scrolls IV: Oblivion - the arcade favorite Need for Speed: Carbon and the heavily simulation oriented GTR 2.

Graphical settings for these games are set to their highest possible values. Notable differences are F.E.A.R where we disable "Soft Shadows", Quake 4 where we use High Quality - not Ultra Quality - but anisotropic filtering was set to "1" for default tests and "16" for AA / AF tests. We used Serious Sam II built in "Maximum" quality preset then change resolution and AA / AF settings accordingly. All of these settings can be seen here. Our TES IV: Oblivion settings are as close as we can get to Ultra High Quality - you can see our test settings here. For Need for Speed: Carbon, we used the game's "Maximum" video quality preset, but change the filtering to "Trilinear" for default tests - AA AF tests are run with "4x" and "Anistropic" (talk about your typos, come on EA!!). NFS: Carbon test were done with the Sprint Race on Lincoln Boulevard. Below are the settings we used for GTR 2 Clear weather was used for our run of Monza GP tests.

Image quality settings in the Forceware driver for the NVIDIA GeForce 7900GS was set to "High Quality". We did leave the "Threaded Optimization" on, but all other optimizations are turned off (trilinear, anisotropic sample and mip filter optimizations). Based on past experience, at these settings image quality is comparable to ATI's default image quality settings. V-sync was turned off on both cards via the driver control panel and also through in game settings whenever possible. For some games (Call of Duty, Homeworld 2, Richard Burns Rally,GTR2), we force AA and AF through the driver's control panel. 



Our test setup
Intel Core 2 Duo E6300 socket LGA-775
4 x 512 MB A-DATA Vitesta 5-5-5-18 PC6400 DDR2-SDRAM
Gigabyte Radeon X1950 Pro 256 MB graphics card
Leadtek PX7900GS GeForce 7900GS 256 MB graphics card.
Gigabyte P965-DS3P Intel P965 motherboard
Maxtor DiamondMaxPlus9 80 GBs Serial ATA 8 MB buffer
LiteOn 1673S DVD-RW
Tagan TG530-U15 530 watts ATX/BTX power supply

Windows XP Professional with Service Pack 2 installed
ATI Catalyst 7.1 reference driver
NVIDIA Forceware 93.71 reference driver
Intel Chipset Software Installation Utility 8.1.0.1006
DirectX 9.0c
all respected games used for benchmarks have been updated to their latest, final builds.

We'd like to thank Gigabyte for providing the Gigabyte P965-DS3P motherboard and the Gigabyte Radeon X1950 Pro graphics card for this test. We'd also like to thank Leadtek for providing the PX7900GS TDH and Tagan for the Tagan TG530-U15 power supply.

The results:

X1950 Pro 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	40.20%	41.41%	42.42%		38.61%	12.00%	8.05%
Average	32.16%	27.09%	19.40%		18.11%	10.81	6.59%
Maximum	19.74%	25.97%	11.87%		13.20%	0.88%	0.56%

7900GS 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	39.22%	41.58%	40.59%		40.20%	-0.99%	0.00%
Average	30.94%	21.50%	12.34%		6.57%	2.44%	0.16%
Maximum	15.85%	14.93%	0.45%		0.00%	0.47%	0.69%

We chose to use a table instead of graph to show the scalability of these two graphics cards. If we were to use a graph, we would be forced to make separate graphs for minimum, average and maximum fps - a table is much easier to do and can be just as easy to read. The first row is for minimum fps, the second is for average fps and the third is maximum fps. By providing all three, we can see at what points these benchmarks scale. For the Call of Duty - Dawnville demo, we can see that for the most part it is on the lower frame rates we see the biggest boost - so those parts in the benchmark are heavily system limited. We're seeing something in the area of 40 % increase on both graphics cards. Notice how the Radeon X1950 still manage a 12 percent boost of minimun fps at 1280 x 960 with AA and AF enabled while the GeForce 7900GS frame rates hardly change at all. Another proof of more pixel processing headroom is the 13 percent boost in maximum fps at 1024 x 768 with AA and AF.


Call of Duty - Dawnville 1600
(click here for a more detailed graph)

	1.8 X1950 Pro	2.8 X1950 Pro	1.8 7900GS	2.8 7900GS
<30 fps	0	0	0	0
30-45 fps	0	0	0	0
45-60 fps	0	0	0	0
60-90 fps	0	0	0	0
90-120 fps	8	0	8	0
>120 fps	68	76	68	76
Total	76	76	76	76

results are in seconds

The table above don't really show anything, but it sure is breath taking. We spent most of our time at above 90 fps, running at 1600 x 1200. At this resolution, the GeForce 7900GS experience a performance penalty so the Radeon X1950 Pro was noticeably faster for most of this benchmark. This is true most the time at 1.8 GHz, but only about half the time at 2.8 GHz. The second half of the benchmark are the most system bound - we saw a 50 to 75 fps increase at one point by overclocking the processor to 2.8 GHz. However, we don't see any difference between the two cards. Graphical wise, the first parts of the benchmark are the part that's particularly interesting. It would seem that the GeForce 7900GS already hit its graphical limits while the Radeon X1950 Pro prove it can still offer higher frame rates with a higher clocked processor.


X1950 Pro 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	87.21%	107.14%	111.25%		106.98%	104.65%	55.42%
Average	31.13%	32.07%	26.39%		30.04%	25.76%	5.89%
Maximum	24.20%	29.52%	32.34%		25.37%	27.44%	4.41%

7900GS 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	150.00%	105.05%	74.74%		57.29%	54.69%	57.14%
Average	31.60%	26.69%	6.09%		0.30%	-1.45%	-1.30%
Maximum	27.70%	26.12%	8.20%		0.85%	0.00%	-0.54%

First, let's look at default test results where AA and AF is not a factor. Frame rate wise, the GeForce 7900GS is the better card here. We experienced notably higher frame rates with it than with the Radeon X1950 Pro. The GeForce 7900GS is faster all the way up to 1600 x 1200, though the difference decreases with each increase in resolution and higher processor clocks. The situation is reversed when we enable AA and AF. The Radeon X1950 Pro simply leaves the GeForce 7900GS way, way behind. The frame rate differences at 1600 x 1200 AA AF shows this very clearly - we're seeing practically a 100 percent higher minimum fps and 125 percent higher average fps with the Radeon X1950 Pro and an 6300 running at 2.8 GHz.

From the table we can see the Radeon X1950 Pro scalability persist even when AA and AF is enabled. It does hit its graphical limits at 1600 x 1200 with AA and AF, but that's still much better than the GeForce 7900GS which already hit its limits at 1024 x 768 AA AF. We can also see how heavily system limited this benchmark really is - minimum fps increase with the use of a 2.8 GHz E6300 are phenomenal. Average frame rate wise, the Radeon X1950 Pro only began to slow scaling at 1600 x 1200, while the GeForce 7900GS don't have any extra headroom anymore.


Homeworld 2 - Vaygr Bomber Strike 1600
(click here for a more detailed graph)

	1.8 X1950 Pro	2.8 X1950 Pro	1.8 7900GS	2.8 7900GS
<30 fps	0	0	0	0
30-45 fps	0	0	0	0
45-60 fps	0	0	0	0
60-90 fps	0	0	0	0
90-120 fps	0	0	0	0
>120 fps	70	70	70	70
Total	70	70	70	70

results are in seconds

Looking at the graphs, it became clear that the Radeon X1950 Pro offer the better performance with a higher clocked processor - it scales better. On the first parts of the benchmark, we see the use of a higher clocked processor allows the Radeon X1950 Pro to offer a similar level or significantly higher frame rates than the GeForce 7900GS. The GeForce 7900GS do have the tendency to lose steam at 1600 x 1200, so we have to take that into account. We do think the GeForce 7900GS performed better on the middle parts of the benchmark, offering consistently higher frame rates. The last parts of the benchmark is not very useful - the Radeon X1950 Pro is slightly faster because of higher maximum fps.


X1950 Pro 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	10.00%	-3.23%	4.55%		2.94%	0.00%	50.00%
Average	29.06%	18.33%	16.74%		20.57%	13.24%	25.49%
Maximum	31.76%	24.71%	24.68%		27.50%	18.99%	35.90%

7900GS 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	-3.85%	-3.03%	3.85%		3.13%	-16.00%	-5.56%
Average	26.90%	27.91%	10.46%		14.33%	-5.77%	-1.77%
Maximum	25.90%	42.37%	19.66%		16.22%	-5.10%	-4.29%

Our KOTOR benchmark is a very different beast - the most graphically intense parts are the ones with the lowest frame rates. So, system bound situations occur on the high frame rate parts of the benchmark. The GeForce 7900GS excel in this benchmark, offering consistently higher average and maximum fps with higher processor clocks than the Radeon X1950 Pro. The average frame rate differences are significant - 25 to 50 percent higher depending on the resolution and processor clock (at default settings without AA AF). We don't really recommend enabling AA and AF above 1024 x 768 on both cards, seeing frame rate drops below 30 fps occur more frequently at 1280 x 1024 and 1600 x 1200. Even with AA and AF enabled, the GeForce 7900GS is definitely faster than the Radeon X1950 Pro.

Scalability wise, the GeForce 790GS is also the better card here. Remember, here we need to pay more attention to average fps than minimum fps. Without AA and AF, the GeForce 7900GS scored more than 25 percent increase in performance at 1024 x 768 and 1280 x 1024 when paired with a higher clocked processor. We could argue the Radeon X1950 Pro offer slightly higher increase in performance at 1600 x 1200 and when AA AF is enabled, but the nominal frame rates are still not close enough to GeForce 7900GS frame rates. No ball.


SW: KOTOR - Endar Spire 1024
(click here for a more detailed graph)

	1.8 X1950 Pro	2.8 X1950 Pro	1.8 7900GS	2.8 7900GS
<30 fps	0	0	0	0
30-45 fps	1	1	0	0
45-60 fps	2	2	1	0
60-90 fps	22	7	3	3
90-120 fps	0	16	25	9
>120 fps	0	0	0	11
Total	25	25	29	23

results are in seconds

Do remember that our choice to use "Advanced" for Catalyst AI caused lower frame rates in this benchmark. However, disabling Catalyst AI will not help the Radeon X1950 Pro reach GeForce 7900GS frame rates. Pairing the card with an E6300 running at 2.8 GHz does allow higher frame rates and a better gaming experience - 7 seconds spent on 60 - 90 fps compared to 22 seconds with an E6300 running at default clock speed. However, at the same settings, we also see a much substantial increase with the GeForce 7900GS - 9 seconds spent on 90 - 120 fps compared to 25 seconds. The Radeon X1950 still dropped to 30 to 45 fps with a 2.8 GHz E6300 while the GeForce 7900GS practically never dropped to below 60 fps .


X1950 Pro 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	33.33%	3.62%	0.00%		0.83%	0.00%	0.00%
Average	43.10%	31.31%	10.55%		5.64%	0.30%	0.06%
Maximum	43.70%	29.10%	25.49%		31.76%	3.36%	0.00%

7900GS 1.8 GHz to 2.8 GHz

	1024	1280	1600		1024 AA AF	1280 AA AF	1600 AA AF
Minimum	17.32%	2.54%	-0.99%		0.00%	0.00%	0.00%
Average	38.83%	21.94%	5.73%		1.80%	0.08%	0.01%
Maximum	41.01%	35.16%	25.90%		23.57%	0.00%	-0.65%

Both cards seems to trade places back and forth. In this benchmark, the Radeon X1950 Pro tries to prove its case that it is the better card. Let's look at default test results first. Minimum and average fps are definitely higher on the Radeon X1950 Pro and the rate of increase - scalability - is higher. Moving on to higher resolutions, the situations don't change all that much. Even at 1600 x 1200, the Radeon X1950 Pro still manage to offer higher frame rates when paired with an E6300 running at 2.8 GHz. AA and AF results are also interesting. At 1024 x 768, the GeForce 7900GS results are pretty much constant while the Radeon X1950 Pro results are higher when the E6300 processor is running at 2.4 and 2.8 GHz. The table confirms what we already seen in the graph - in addition to offering higher frame rates, the Radeon X1950 Pro scales better in this benchmark.