In NVIDIA's usual style, there will be two flavors of the GeForce FX available in February - the GeForce FX 5800 Ultra and the regular 5800.

The GeForce FX 5800 Ultra will retail for $399, just like ATI's Radeon 9700 Pro; we'd expect street prices to be a bit higher than the 9700 Pro, at least initially. As NVIDIA originally claimed, the FX 5800 Ultra will be shipping with a 500MHz core clock at 1GHz memory clock (effectively 1GHz that is). The card we received is what will be shipping in February and features the unique 2-slot cooling solution that NVIDIA shocked everyone with at Comdex.

We'll talk about cooling in a bit but just by looking at the card you can tell that the GeForce FX 5800 isn't cheap to manufacture, not by a long shot. All of the materials that go into the cooling system and mounting it on the card definitely increase manufacturing costs well beyond what ATI's board partners have to deal with, so it will be interesting to see if street prices for the GeForce FX will be able to rival those of the 9700 Pro.

The 0.13-micron NV30 GPU and its 1GHz memory are power hogs and thus the 4-pin standard HDD power connector are present on the PCB. If you fail to plug in the power connector the card will still function and it will let you into your OS without any problems. The card will underclock itself from the stock 500/1GHz settings to 250/500MHz; assuming that the GeForce FX's drivers are installed a window will pop up reminding you to plug in the power connector.

The regular GeForce FX 5800 will be priced around $299 and ships with a much lower 400MHz core clock and 800MHz memory. By reducing the core and memory clocks, NVIDIA's board partners should be able to produce regular 5800 solutions in a single-slot design although none were ready at the time of this review for us to include pictures of. We did include benchmarks at 400/800 to give you an idea of where the performance of the regular 5800 will lie.

The most stunning part of the GeForce FX 5800 Ultra is its utterly insane cooling system and the first thought that popped into our minds when we saw it was: noise. With the GeForce FX, NVIDIA borrowed some of their mobile technology to help minimize the amount of noise their cooling system would produce.

For starters, when the 3D pipeline is not in use the card will operate at lower clock speeds (300 core / 600 mem to be specific). The lower clock speeds enables NVIDIA to run the very loud fan at a much slower speed and still provide adequate cooling to the GPU and its memory. As soon as there is activity in the 3D pipeline, the clock speeds increase to their 3D setting (500/1GHz) and the fan spins quicker to deal with the increased temperatures.

To give you an idea of how loud the GeForce FX 5800 Ultra can get we've recorded two videos several inches away from a GeForce FX 5800 Ultra and a Radeon 9700 Pro for comparison purposes. The FX 5800 Ultra video starts out in 2D mode and then we start up a game, triggering an increase in clock and fan speed which you'll definitely be able to hear. Remember that these videos were made very close to the actual source of the sound, so it's much like putting your ear up to either one of these cards.

ATI Radeon 9700 Pro in action (AVI)
NVIDIA GeForce FX 5800 Ultra in action (AVI)

To further extend the comparison, the GeForce FX 5800 weighed in at 77 dBA on our sound meter with the fan spun all the way up vs. 64 dBA on the Radeon 9700 Pro. With the fan in its "slow" mode the card was still louder than the Radeon 9700 Pro but not nearly as bad. The rear heatsink on the GeForce FX 5800 also got quite warm, the maximum temperature of the surface reached 131F as it is still not being passively cooled. Remember that the fan setup on the GeForce FX only cools the GPU and the memories mounted on the front of the PCB, the back is cooled passively.

An interesting phenomenon we discovered while attempting to overclock the GeForce FX was that when the card got too hot (courtesy of our overclock), it automatically throttled itself down to its 2D speed (300/600MHz) and reduced the fan speed accordingly in the middle of a game. We're not exactly big fans of this method of protection, since it would make more sense to reduce the clock speed to its 3D default setting and not the significantly slower 2D clocks.

If you are curious, the card underclocked itself after running at a 540MHz core clock for a matter of minutes, and we couldn't get the memory up above 1100MHz too reliably either; as expected, there's not much headroom in the GeForce FX 5800 although that may change over the life of the chip as the manufacturing process is tuned.

NVIDIA supplied beta drivers to us that properly recognized the GeForce FX 5800 Ultra - v42.63. Other than an issue with enabling NVIDIA's 4XS AA mode (we'll discuss that later), we had no issues with the drivers on either our AMD or Intel test systems. After extensive testing however, it's clear that NVIDIA does have a good amount of optimization to work on before the GeForce FX is truly ready to compete with the likes of the Radeon 9700 Pro; you'll understand why as we take you through the performance tests.

The drivers have very similar control panels to what we're all familiar with, with a few new additions that are GeForce FX specific:

You can set 2D/3D clocks independently, we found it curious that NVIDIA set the 2D clocks so high

AA and Anisotropic Filtering settings are controlled in OpenGL and Direct3D by this single set of sliders:

The Performance - Balanced/Aggressive slider at the top controls the anisotropic filtering quality/performance settings. The Balanced setting offers higher image quality while the Aggressive setting offers greater performance.

Since the focus of this review is the performance of the two fastest desktop GPUs there are a few assumptions we can make, the biggest being that you don't buy a card like the GeForce FX 5800 Ultra to run without Anti-Aliasing or Anisotropic Filtering enabled. For this reason the vast majority of our benchmarks will be focused on the performance with AA/AF enabled, but we will include an abridged set of numbers without those features turned on to be as complete as possible.

Before we get into the truly strenuous benchmarks let's have a look at some scores without AA & Anisotropic Filtering enabled. First things first, synthetic benchmarks using 3DMark 2001:

3DMark 2001 SE helps confirm that NVIDIA is able to deliver as expected in terms of fill rate; with a much higher core clock (500MHz vs. 325MHz) it's no surprise that the GeForce FX 5800 Ultra comes away with a 37% higher fill rate in 3DMark than the Radeon 9700 Pro. What you have to take into account however is the fact that the Radeon 9700 Pro actually has more raw memory bandwidth than the GeForce FX (when you ignore the always-on color compression of the FX), which can significantly reduce any real-world fill rate advantage NVIDIA may have.

The rest of the 3DMark numbers are generally in favor of the GeForce FX but there are a few cases when the 5-month old Radeon 9700 Pro comes out on top. Let's move on to some real-world benchmarks before we conclude this quick performance intro and start focusing on why you'd really want to purchase one of these cards - for Anisotropic/AA performance.

Unreal Tournament 2003 AA/Aniso Disabled 1024x768x32 Radeon 9700 Pro GeForce FX 5800 Ultra GeForce FX 5800 Radeon 9700 205.4 204 198 189.2 | 0 | 41 | 82 | 123 | 164 | 205 | 246

What once used to be only true for video cards at 640x480 is now true for these power houses at 1024x768, being mostly CPU/driver limited that is. At 1024x768 we're not really stressing anything but what's very interesting and impressive on ATI's part is that the Radeon 9700 Pro is slightly faster than the GeForce FX. If you remember back to all previous ATI cards, in CPU/driver bound situations ATI was almost never on top. It always took turning on AA or cranking up the resolution in order to let ATI's bandwidth advantages (if they existed) to kick in and give them the performance advantage, but here we have ATI on top from the start. It just goes to show you how far along ATI's driver team has come and how mature the Radeon 9700 Pro is, considering that it has been shipping for close to 6 months now.

Unreal Tournament 2003 AA/Aniso Disabled Minimum FPS - 1024x768x32 Radeon 9700 Pro Radeon 9700 GeForce FX 5800 Ultra GeForce FX 5800 105 84.1 73.7 71.6 | 0 | 21 | 42 | 63 | 84 | 105 | 126

Due to popular demand we've included graphs of minimum frame rates as well as average frame rates in situations where the benchmark reports the data to us. Here we see that the GeForce FX, for whatever reason, has much lower minimum frame rates than the Radeon 9700 Pro. We have no explanation for this other than we hope it is an issue with the current GeForce FX drivers.

Unreal Tournament 2003 AA/Aniso Disabled 1280x960x32 GeForce FX 5800 Ultra Radeon 9700 Pro GeForce FX 5800 Radeon 9700 181.2 160.2 160.2 138.6 | 0 | 36 | 72 | 109 | 145 | 181 | 217

As we increase the resolution the GeForce FX 5800 Ultra actually begins to pull away from the Radeon 9700 Pro, here establishing a 13% performance advantage. But now let's have a look at the minimum frame rates:

Unreal Tournament 2003 AA/Aniso Disabled Minimum FPS - 1280x960x32 Radeon 9700 Pro Radeon 9700 GeForce FX 5800 GeForce FX 5800 Ultra 96.3 83.8 69.9 65.2 | 0 | 19 | 39 | 58 | 77 | 96 | 116

Once again we see that while the average frame rates are quite high, the GeForce FX 5800 falls far behind even the regular Radeon 9700 in the lowest frame rates attained during the course of the benchmark.

Unreal Tournament 2003 AA/Aniso Disabled 1600x1200x32 GeForce FX 5800 Ultra GeForce FX 5800 Radeon 9700 Pro Radeon 9700 132.5 110.3 108.6 93.3 | 0 | 27 | 53 | 80 | 106 | 133 | 159

At 1600x1200 even the regular GeForce FX 5800 is able to outperform the Radeon 9700 Pro in average frame rate numbers, but once the attention shifts to the minimum frame rates...

Unreal Tournament 2003 AA/Aniso Disabled Minimum FPS - 1600x1200x32 Radeon 9700 Pro GeForce FX 5800 Ultra GeForce FX 5800 Radeon 9700 66.1 60.2 59.5 56.6 | 0 | 13 | 26 | 40 | 53 | 66 | 79

The GeForce FX doesn't do nearly as bad in comparison to the Radeon 9700 Pro here, as the minimum frame rates are relatively similar across the board at this high of a resolution. It's nice to note that all of these cards can provide minimum frame rates at around the 60 fps mark at 1600 x 1200.

The only other benchmark we ran with AA/AF disabled was Serious Sam 2: The Second Encounter (Demo) using the Little Trouble benchmark. As usual we used the Extreme Quality defaults but disabled Anisotropic Filtering (we'll be saving those benchmarks for a later section).

Serious Sam 2 AA/Aniso Disabled 1024x768x32 GeForce FX 5800 Ultra GeForce FX 5800 Radeon 9700 Pro Radeon 9700 160 148.2 146.6 139.2 | 0 | 32 | 64 | 96 | 128 | 160 | 192

Once again we see that the GeForce FX 5800 Ultra holds a small but reasonable (9%) performance advantage over the Radeon 9700 Pro.

Serious Sam 2 AA/Aniso Disabled Minimum FPS - 1024x768x32 GeForce FX 5800 Ultra Radeon 9700 Pro GeForce FX 5800 Radeon 9700 94.5 93.5 90.2 87.4 | 0 | 19 | 38 | 57 | 76 | 95 | 113

The minimum frame rate numbers are a bit more reliable under Serious Sam because we can report the sustained minimum frame rates, rather than the instantaneous minimums; thus these scores are much less prone to being thrown off by minor issues that would cause significant but short drops in frame rate.

Serious Sam 2 AA/Aniso Disabled 1280x960x32 GeForce FX 5800 Ultra Radeon 9700 Pro GeForce FX 5800 Radeon 9700 125.8 119.7 109.1 106.4 | 0 | 25 | 50 | 75 | 101 | 126 | 151

At higher resolutions the performance gap between the GeForce FX 5800 Ultra and the Radeon 9700 Pro is reduced instead of increased, in this case it is only 5%. This is a perfect case of ATI's raw memory bandwidth advantages outweighing any fill rate advantages that the GeForce FX has.

Serious Sam 2 AA/Aniso Disabled Minimum FPS - 1280x960x32 GeForce FX 5800 Ultra Radeon 9700 Pro Radeon 9700 GeForce FX 5800 83.5 81.3 76.1 70.1 | 0 | 17 | 33 | 50 | 67 | 84 | 100

The minimum frame rates continue to be very close...

Serious Sam 2 AA/Aniso Disabled 1600x1200x32 GeForce FX 5800 Ultra Radeon 9700 Pro Radeon 9700 GeForce FX 5800 88.4 87.4 76.5 76.3 | 0 | 18 | 35 | 53 | 71 | 88 | 11

Finally at 1600x1200 we see that virtually all performance advantage has been erased, as the GeForce FX 5800 Ultra is barely 1% faster than the Radeon 9700 Pro in this case.

Serious Sam 2 AA/Aniso Disabled Minimum FPS - 1600x1200x32 Radeon 9700 Pro GeForce FX 5800 Ultra Radeon 9700 GeForce FX 5800 61 60.9 56 54.5 | 0 | 12 | 24 | 37 | 49 | 61 | 73

Now that we've gotten that out of the way, let's really stress these cards by turning on some eye-candy...

With the Radeon 8500 and 9700 Pro, ATI brought anisotropic filtering to the mainstream by introducing a filtering engine that minimized the performance hit seen with anisotropic enabled. Because of the pressure ATI put on NVIDIA, the GeForce FX driver lets you choose between two different anisotropic filtering modes that adjust the engine between optimizing for performance and image quality, the two options being Balanced and Aggressive.

ATI has two similar options, at least in name, in their drivers for the Radeon 9700 Pro; ATI calls the two options performance and quality.

In order to compare the various anisotropic filtering modes we chose this scene from the Inferno Flyby in Unreal Tournament 2003. The white box is the area we zoomed in on (400%) to investigate the quality of ATI's anisotropic filtering vs. NVIDIA's engine in the GeForce FX.

First let's look at the highest image quality settings offered by NVIDIA, which in the case of the GeForce FX is an 8-sample algorithm. To view the different screenshots simply hold your mouse over the appropriate link and the corresponding screenshot will appear. By default, anisotropic filtering disabled is the screenshot that will be shown if you don't mouse-over one of the other links:

[ Aniso Disabled | 8X Aggressive | 8X Balanced ]

Here you can see there noticeable differences between NVIDIA's aggressive and balanced settings, although both are a definite improvement over no-anisotropic filtering as you would expect.

Next we have ATI's AF options:

[ Aniso Disabled | 8X Performance | 8X Quality | 16X Performance| 16X Quality]

What makes ATI's solution so great is that there is virtually no discernible image quality difference when you compare the Performance and Quality settings. We included shots of ATI's 16X setting as well just to show that there is a slight increase in image quality, but for a fair comparison we will put NVIDIA's 8X up against ATI's 8X.

"NVIDIA Performance-Balanced is closest to ATI Quality mode. NVIDIA Performance – Aggressive matches closest to ATI Performance mode."

Now that we've concluded that NVIDIA's 8X Balanced AF mode is comparable to either of ATI's 8X modes, let's figure out how all of the options perform:

Impact of Anisotropic Filtering on Performance NVIDIA GeForce FX 5800 Ultra Aniso Off 2X Aggressive 2X Balanced 4X Aggressive 4X Balanced 8X Aggressive 8X Balanced 100% 99% 92% 99% 81% 98% 75% | 0 | 20 | 40 | 60 | 80 | 100 | 120

NVIDIA's Performance - Aggressive AF setting results in virtually no performance hit, which isn't anything to complain about however once you take into account the fact that it doesn't compare visually to ATI's Performance setting it is not as impressive.

NVIDIA's Performance - Balanced setting incurs a much larger performance hit, but it is the only option that is visually comparable to ATI's AF.

Let's see how the decreased clock speed and lesser memory bandwidth of the regular GeForce FX 5800 changes the picture:

Impact of Anisotropic Filtering on Performance NVIDIA GeForce FX 5800 Aniso Off 2X Aggressive 2X Balanced 4X Aggressive 4X Balanced 8X Aggressive 8X Balanced 100% 100% 84% 97% 69% 93% 63% | 0 | 20 | 40 | 60 | 80 | 100 | 120

With less memory bandwidth to spare, the regular FX 5800 does take a noticeably larger performance hit when AF is enabled.

Finally we have the Radeon 9700 Pro:

Impact of Anisotropic Filtering on Performance ATI Radeon 9700 Pro Aniso Off 2X Performance 2X Quality 4X Performance 4X Quality 8X Performance 8X Quality 16X Performance 16X Quality 100% 100% 77% 94% 62% 90% 57% 89% 56% | 0 | 20 | 40 | 60 | 80 | 100 | 120

ATI's Performance AF setting incur a much larger performance drop than NVIDIA's Performance - Aggressive mode, but that's understandable considering that ATI's setting offers noticeably improved image quality. On the flip side of things, ATI's Quality setting takes a larger performance hit than NVIDIA's Performance - Balanced mode.

So what do we compare? Based on our image quality assessments from the previous page, it is clear that we can't compare NVIDIA's Performance - Aggressive mode to either of ATI's AF modes. And although there is only a 1% performance drop between ATI's 8X and 16X modes, in order to compare apples-to-apples we stick to 8X AF modes with both cards. The choice comes down to whether we use ATI's Performance or Quality modes for the Radeon 9700 Pro, since there's virtually no perceivable difference in image quality we focus the majority of our benchmarks on NVIDIA's 8X Performance - Balanced AF vs. ATI's 8X Performance AF.

Now that we've investigated the GeForce FX's Anisotropic Filtering modes, the next step is to see how well its Anti-Aliasing engine works. We followed the same procedure as we did with our AF investigation, and in doing so we managed to uncover a bug in the 42.63 drivers NVIDIA sent us for testing; it turns out that a bug in this build of the drivers causes visual artifacts if NVIDIA's 4XS Anti-Aliasing is enabled, and thus we had to exclude that setting from our tests.

An example of the visual artifacts present with 4XS AA enabled on the GeForce FX

NVIDIA plans to fix this bug in a later revision of the driver, but given that it was the weekend they could not fix it in time for this review.

Below you'll find the scene we used from the same Inferno Flyby in Unreal Tournament 2003 for our AA quality tests:

The white box is the area we zoomed in on (400%) to investigate the actual quality of the AA algorithm each of the cards employed. Once again, we'll start with the newcomer - NVIDIA's GeForce FX.

To view the different screenshots simply hold your mouse over the appropriate link and the corresponding screenshot will appear. By default, AA disabled is the screenshot that will be shown if you don't mouse-over the other links:

[ AA Disabled | 2X | QC | 4X | 6XS | 8XS ]

The first thing you should notice is that neither of the 2-sample AA algorithms (2X or Quincunx) seem to be doing much of anything in the way of actually anti-aliasing. We brought this to the attention of NVIDIA but heard nothing in return and can only conclude that there are still unfinished issues with the current GeForce FX drivers. There is a performance drop (as you'll soon see) with 2X/QC AA enabled, and there is a slight difference in image quality but to say that the 2X mode is doing a good job is stretching it.

Luckily the 4X and higher settings seem to be working just fine; remember that all of the 'S' modes (e.g. 6XS, 8XS) are only available under Direct3D and not OpenGL; this isn't too big of a problem considering that even the GeForce FX doesn't have the power in it to run most of today's games at high resolutions with anything above 4X enabled at reasonable frame rates.

Just so you're familiar with what we're comparing to let's take a look at the Radeon 9700 Pro:

[ AA Disabled | 2X | 4X | 6X ]

With the Radeon 9700 Pro ATI got rid of the Performance/Quality AA settings and just went to one single algorithm, which makes our job much simpler. By looking at ATI's 2X AA mode you can already tell that it's doing more than NVIDIA's 2X setting, but to make things easier let's look at the two compared directly.

This comparison works the same way our ATI vs. NVIDIA AF comparisons worked, by default NVIDIA's 2X is showing in the image below but all you have to do is hold your mouse over the image and you'll see ATI's 2X AA:

Hold your mouse over the image to view ATI's 2X AA

With this comparison you can get a good idea of exactly how little NVIDIA's 2X AA is doing. Looking at the ATI screenshot you can see that there are now intermediate steps between the jaggies, whereas that's completely absent from NVIDIA's 2X screenshot.

So what about NVIDIA's 4X mode? We've already established that it's doing considerably more than their 2X mode, but how does it compare to ATI's 4X AA? You know the drill:

Hold your mouse over the image to view ATI's 4X AA

This is a much closer competition now, but you can still see a slight advantage in ATI's 4X mode. The AA samples in the ATI screenshot seem to be blended together better than what we see in the screenshot produced by the GeForce FX; it's not a huge deal, and no where near as dramatic as the 2X comparison, but it's still worth noting.

So what do we compare? Since we've established that NVIDIA's 2X setting needs some work, it looks like we'll be pitting ATI's 4X up against NVIDIA's 4X

We benchmarked the GeForce FX 5800 Ultra and the ATI Radeon 9700 Pro on every single 266MHz FSB Athlon XP to produce an interesting CPU scaling chart. The tests were conducted at 1024x768 with no anti-aliasing or anisotropic filtering enabled in order to increase the dependency on the CPU and driver subsystems. The results were interesting to say the least:

The interesting thing is that the GeForce FX never overtakes the Radeon 9700 Pro, it consistently falls behind by 4 - 6%. With a higher core clock, the only explanation for the GeForce FX falling behind here would be that the Radeon 9700's raw memory bandwidth advantage helps it out. We debunked this theory by increasing the FX's memory bandwidth by 10% (read: overclocked it to 1100MHz memory) and noted no more than a 0.2% increase in performance, so this test was clearly not saturating the memory bus of the GeForce FX. In theory, the GeForce FX should be outperforming the Radeon 9700 Pro here, but in reality it was not at all.

All evidence points to the GeForce FX's drivers holding it back, but this seems rather atypical of NVIDIA, or does it? Historically, NVIDIA has always provided a driver release 3 - 6 months after the launch of a new GPU that has improved performance anywhere from 10 - 20% on average. With the GeForce FX being a relatively new core, we'd expect to see something very similar from them down the road. The difference this time around is that NVIDIA has a much more powerful competitor in ATI, the Radeon 9700 Pro is no sitting duck. With very mature drivers and solid performance across the board, the GeForce FX vs. ATI's Radeon 9700 is almost reminiscent of the Radeon 8500 vs. NVIDIA's GeForce3 Ti 500, but with the roles reversed.

So there you have it, NVIDIA's response to ATI's Radeon 9700 Pro - but does anyone else feel unfulfilled by the GeForce FX? A card that is several months late, that is able to outperform the Radeon 9700 Pro by 10% at best but in most cases manages to fall behind by a factor much greater than that. Granted the problems that plagued the launch of the FX weren't all up to NVIDIA's control, after all the decision to go 0.13-micron was made 1 - 2 years ago based on data that was available at the time. ATI took a gamble on producing a 0.15-micron part and NVIDIA did the same on their 0.13-micron NV30, and it looks like ATI guessed right.

While we were reviewing the FX, looking at its performance and investigating its image quality, we found ourselves reminiscing of ATI's launch of the Radeon 8500. A card that was long overdue, but in the end unable to outshine the top performer at the time. Although the current state of the GeForce FX is much better than what we had with the first Radeon 8500, the word impressive isn't what we'd use to describe it. The performance is an improvement over the Ti 4600, without a doubt, but it does not place NVIDIA back in a position of dominance, which is what everyone was expecting from NV30. This isn't the end of NVIDIA, the company is quite healthy and they've got a number of products in the pipeline with great potential (GeForce FX included) but it does mean that the road to regaining dominance in the market will be an even more difficult one to traverse.

ATI has not been sitting idle all this time, and progress on the R350 core has been coming along quite well. We proved early on that the 0.15-micron R300 core could reach speeds of up to 400MHz, and with the GeForce FX NVIDIA has established that shipping cards with 800MHz - 1GHz memory is feasible, if ATI can put together a R350 with specs close to what we're implying then even a driver-tweaked FX will not stand a chance. NVIDIA has told us that the GeForce FX will be in stores next month, and we'd expect R350 in about a month following that. It will be a close race, but what ATI has going for them right now is a much more mature driver set than NVIDIA for their flagship GPU. The 3+ month advantage ATI had in bringing the R300 into production and to market gave ATI a much bigger advantage than just being the king of the hill for a while, it gave them quite a bit of time to fine-tune and optimize their drivers for this very occasion; this is a luxury that ATI has not had previously but they have made excellent use of it today.

NVIDIA's focus at this point is NV31 and NV34, after all, that's where the money is. The small percentage of the market that will go after the NV30 will not make or break NVIDIA, but should ATI compete like this in other market segments then there will be cause for worry. As we mentioned at the start of our GeForce FX Preview - "Kudos to ATI."