Summary: Are you ready for NV35? NVIDIA's highly anticipated successor to NV30 is here and it has a new name: GeForce FX 5900 Ultra. Boasting a 256-bit memory interfance and 256MB of DDR memory, this card packs a powerful punch. NVIDIA has also added its UltraShadow technology specifically for next generation titles that use shadows extensively like Doom III. But is this enough to overtake ATI's RADEON 9800 PRO 256MB? Find out in this article!
NVIDIA’s fall from grace in the high-end segment has definitely been one of the most closely followed stories of 2003 to this point. When ATI stunned the hardware world with RADEON 9700 PRO last summer, everyone expected NVIDIA to counter with something stronger. After all, this was the company that had used the six-month product cycle to beat the competition into obsolescence with great success. GeForce FX 5800, and more specifically, GeForce FX 5800 Ultra were meant to restore NVIDIA’s presence in this space, but both products have been AWOL since they were announced.
This came as quite a surprise to industry watchers, after years of near flawless execution, not only had NVIDIA missed fall, but they also missed the critical holiday shopping season. As a result, ATI’s RADEON 9700 and RADEON 9500 cards enjoyed the distinction of being the only DirectX 9 solutions on the market and all the fanfare that comes with being on top in performance and features.
Just what went wrong with NVIDIA and the GeForce FX 5800 family? In all honesty, the answer depends on who you talk to, but the decision to go with bleeding-edge technologies like DDR2 memory, and TSMC’s relatively untested 0.13-micron manufacturing process are largely considered to be the leading roots of the problem. Essentially NVIDIA took a gamble by incorporating these components in GeForce FX 5800’s design and lost, while ATI played it safe, sticking with TSMC’s more mature 0.15-micron process (which incidentally, is used on a wide range of NVIDIA products) and DDR memory.
These differences in approach resonate throughout the architecture of both products in much the same way AMD’s Athlon XP differs from Intel Pentium 4. Like the Pentium 4, GeForce FX needs its high clock speed to attain optimal performance; RADEON 9700 on the other hand relies on a more balanced approach of additional functional units with a more palpable clock frequency.
NVIDIA must have felt that with GeForce FX 5800’s 125 million transistors, 0.13-micron was more feasible than 0.15-micron: the smaller process keeps GeForce FX 5800’s die size in check, making the chip cheaper to produce. Hitting those astounding clock speeds NVIDIA was shooting for was also more attainable at 0.13-micron. At least, that’s probably what NVIDIA engineers were thinking early on in the chip’s design, in practice we now see that NVIDIA and TSMC have had a difficult time getting sufficient quantities of chips at the clock speeds NVIDIA was shooting for, resulting in a lot of parts that weren’t up to snuff and thus have to be thrown away.
NV35: The graphics inferno
As this story was progressing in the public limelight, NVIDIA had another design team that was quietly plugging away on its follow-up to GeForce FX 5800, internally codenamed “NV35”. Just as GeForce4 was NVIDIA’s “refresh” to GeForce3, NV35 was designed as an update to GeForce FX 5800. NVIDIA has followed this basic strategy since its inception. If you recall NVIDIA’s previous refresh products, they’ve all offered some pretty remarkable performance improvements over their predecessors. Does NV35 live up to this legacy?
GeForce FX 5900: Building on NV30’s roots
From the outset, NVIDIA has stated that the NV30 core found in GeForce FX 5800/5800 Ultra would be the foundation for NVIDIA’s entire range of DirectX 9 products. From the $79 GeForce FX 5200 all the way up to the $499 GeForce FX 5900 Ultra (NV35), NVIDIA’s entire GeForce FX family supports the same 2.0+ pixel and vertex shaders that will be found in the next generation of games and 3D applications. Lets take a look at the rest of the bells and whistles that make up the GeForce FX 5900:
256-bit memory interface supports up to 256MB memory
NVIDIA CineFX 2.0 engine
Intellisample HCT performance technology
Advanced display pipeline with full NVIDIA nView capabilities
Like GeForce FX 5800, GeForce FX 5900 is built on TSMC’s 0.13-micron manufacturing process with copper interconnect technology. NVIDIA recently announced a foundry deal with IBM that will be used for its next generation part, due this summer. This next generation core will take advantage of IBM’s latest manufacturing technologies such as copper interconnects, silicon-on-insulator transistors (SOI) and low-k dielectric insulation on larger, 300mm wafers.
GeForce FX 5900 contains approximately 5 million more transistors than GeForce FX 5800, bringing the total to 130 million transistors. NVIDIA has also reduced the core clock frequency of GeForce FX 5900 Ultra to 450MHz (from 500MHz), which should help the 5900 Ultra achieve higher yields than the 5800 Ultra. To correspond with this, NVIDIA has also dropped the FX Flow heat pipe cooler in favor of a more traditional aluminum heatsink. The end result is that GeForce FX 5900 Ultra operates considerably quieter than GeForce FX 5800 Ultra – it’s essentially a night and day difference! We’ll discuss the new cooling unit in more detail a bit later.
The other aspect that really stands out is GeForce FX 5900’s 256-bit memory interface, twice the width of any previous NVIDIA product. This was implemented to address one of the chief complaints of GeForce FX 5800 Ultra: its lack of memory bandwidth.
SIDEBAR: IBM’s $2.5 billion facility that will be utilized by NVIDIA began operation last year and is located in East Fishkill, NY.
256-bit memory interface
If you recall the original GeForce FX 5800 Ultra, the card often outperformed ATI’s RADEON 9700 PRO when running without antialiasing and anisotropic filtering enabled. Once these features were turned on however, GeForce FX 5800 Ultra was often left eating ATI’s dust. This is partially because GeForce FX 5800 Ultra was giving up nearly 4GB/sec of memory bandwidth to the RADEON 9700 PRO. Once you turn on these visual quality features, the demands on the memory subsystem of the graphics card increases, especially as you crank up the screen resolution. The memory subsystem in GeForce FX 5800 Ultra literally became a bottleneck.
Since its announcement, id Software’s Doom III has been one of the most hotly anticipated titles on the market. In addition to the game itself, its underlying game engine will be used in a number of upcoming titles for years to come. Quite frankly, the jaw-dropping graphics in this game are going to move lots of video cards as gamers will upgrade their current graphics card to experience all of the visuals of John Carmack’s latest creation.
With UltraShadow, shadows can be calculated more quickly by defining the areas that need to be calculated, allowing the graphics core to only render the area of the shadow that is necessary, increasing efficiency. (Much like the z-buffer is used to cull unnecessary pixels.) This is accomplished by the programmer, who must define the bounded portions of the scene, depicted as “depth bounds” in the image above. The graphics core limits all calculations of lighting effects to this area, allowing the shadow to be rendered more quickly and with greater efficiency -- the more complex the scene (for example a room with multiple light sources and objects), the more significant the performance improvement. Unfortunately, UltraShadow is only currently exposed through OpenGL, not Direct3D. NVIDIA plans to address this shortly however.
SIDEBAR: Besides Doom III, another title NVIDIA is showing that will use shadow volumes extensively is a game known as Abducted from Contraband Entertainment.
Besides the addition of UltraShadow technology, NVIDIA has improved upon its CineFX architecture first introduced with GeForce FX 5800. NVIDIA has optimized all stages of the pixel pipeline in CineFX 2.0. According to NVIDIA, CineFX 2.0 doubles the floating-point pixel shading power of GeForce FX 5800. This should result in a dramatic performance boost, as pixel shader programs can be executed more quickly.
Intellisample HCT is NVIDIA second generation of compression and caching technology, and has been redesigned for optimal performance with GeForce FX 5900’s new 256-bit memory interface. NVIDIA has introduced new algorithms that are more successful at compressing texture, color and Z data than Intellisample 1.0, resulting in improved antialiasing performance at high resolutions.
The 3D engine
The remainder of GeForce FX 5900’s engine remains unchanged from GeForce FX 5800. This means that the same 4-pixel pipeline architecture with two texture units per pipe (4x2) that caused so much controversy earlier this year is still used. ATI on the other hand has implemented eight pixel pipelines with one texture unit per pixel pipeline (8x1) in its RADEON 9700, 9800 series. ATI is also proud to proclaim that it has one more vertex engine than NVIDIA.
GeForce FX 5900 variants
NVIDIA plans to produce three GeForce FX 5900 variants: the GeForce FX 5900 Ultra 256MB (the board we’re testing today), a 128MB GeForce FX 5900, and a 128MB value solution. The 256MB 5900 Ultra board will ship in June with an estimated selling price of $499. The middle GeForce FX 5900 part will also ship in June but with reduced clock frequencies and a $399 price tag. To avoid confusion, hopefully NVIDIA will use the “Pro” designation to denote these cards.
NVIDIA’s GeForce FX 5900 Ultra is one long graphics card! At nearly 10 inches in length, the GeForce FX 5900 Ultra reference board is the longest graphics card we’ve ever tested. Fortunately, NVIDIA has stated that retail boards will be one inch shorter than our reference card so final cards should fit in a wide variety of system cases with greater ease.
Like the original GeForce FX 5800 Ultra, the 5900 Ultra is a two-slot graphics solution, meaning that you’ll need to keep the PCI slot adjacent to your AGP slot empty for the graphics card to fit properly. Power circuitry is located on the back of the board and is more extensive than the implementation in GeForce FX 5800 Ultra. A 4-pin Molex power connector is used to keep the graphics card fed with juice, as the AGP interface isn’t capable of meeting the power requirements of GeForce FX 5900 Ultra.
If the end user forgets to connect the external power connector to the system power supply, the GeForce FX 5900 Ultra card will still operate, but at a reduced clock frequency. The card will alert you immediately once the operating system is loaded that it’s operating at reduced frequencies, although we still wish it were a little more aggressive at informing the end user of the source of the problem. The current error message “your NVIDIA based graphics adapter is running on a severely limited power source” is a little vague, an inexperienced user may think the system power supply itself is the source of the problem.
New cooling system
NVIDIA’s GeForce FX 5800 Ultra enjoyed a notorious reputation for being one of the loudest graphics cards in recent memory. Its FX Flow cooling system could literally be heard in nearby rooms. As a result all kinds of comical sketches and names were attached to the GeForce FX 5800 Ultra, “leaf blower”, “dustbuster”, and “hair dryer” were the most common synonyms used to describe it.
NVIDIA’s drivers have taken a pretty rough beating lately. We’ve noticed all kinds of visual bugs and other quirks in some of their earlier drivers that we used to test the GeForce FX 5800 Ultra with, the current 43.45 driver also offers reduced image quality in comparison to NVIDIA’s previous driver release.
Fortunately, NVIDIA has attempted to address these issues with its latest 44.03 driver, and they’ve changed a few features in the driver itself. For starters, the “application” setting in the “performance and quality settings” menu is gone. According to NVIDIA, this feature was meant as a debugging tool for developers and shouldn’t have been present in the first place, or at least in the way it was implemented. The application setting was never optimized for performance.
NVIDIA now has three options available in its current driver: performance, high performance and quality. By default, the quality setting is used, just like ATI’s current CATALYST release. On the next page we’ll examine the visual quality of this mode in more depth.
Like the other GeForce FX boards, the GeForce FX 5900 Ultra runs at a slower clock frequency in 2D mode: 350MHz on the core versus 450MHz in 3D mode (the memory frequency remains the same). This allows the card to run quieter, although as we mentioned earlier, even in 3D mode the card remains fairly quiet.
One aspect that has concerned us in the past however is during overclocking. In NVIDIA’s Detonator 42.69 release we found that the card would often underclock itself under some extreme overclocked conditions, resulting in performance that was below the default clock speeds. Fortunately this bug has been squashed in 44.03.
While we’re discussing overclocking, one really handy feature NVIDIA has implemented in 44.03 is the “auto detect” setting. This option tests your graphics card to automatically determine the highest speeds possible. It isn’t 100% accurate (in our use it was a little too aggressive) but it certainly saves a lot of the time and guesswork that goes into overclocking.
SIDEBAR: Did you know that NVIDIA still produces TNT2 Vanta cards?
To test the image quality of the GeForce FX 5900’s anisotropic filtering engine, we first cranked up Quake 3’s DM7 map. We wanted to see exactly what the GeForce FX 5900 Ultra card was doing, so we used Q3’s “r_colormiplevels” console command to highlight the mip-maps being used and the transitions between each of them.
Ideally the colors should blend together, resulting in smooth transitions from each mip-map level (denoted as red, green, or blue in the images). First up, we have a few GeForce FX 5900 Ultra images with the performance setting enabled:
And now the RADEON 9800 PRO in performance mode:
As you can see, the gradients between each mip-map level are harsh with both cards. Lets look at the quality setting for GeForce FX 5900 Ultra:
RADEON 9800 PRO:
The levels blend much smoother in quality mode, both cards look incredibly good to our eyes, but if we have to determine a winner, we’d give the nod ever so slightly to the RADEON 9800 PRO.
Next we conducted visual quality tests with Unreal Tournament 2003 in both cards quality mode:
Textures on both cards are pretty sharp. You really have to take a real close look to see any differences.
SIDEBAR: Matrix Reloaded later this week, can’t wait!
To test anti-aliasing image quality, we used 3DMark 03’s image quality test. If you’ve followed some of our other preview articles, you’ve probably noticed that we prefer to use 3DMark for all image quality anti-aliasing tests. The reason we prefer to use 3DMark is simple, jaggies can creep up at a moment’s notice. One slight twitch of the mouse can make a difference. The only way to ensure that we’re getting the exact same image is to use 3DMark’s image quality test, which grabs a screen capture of a static display.
We chose game test 1 because it’s a bright map (and thus jaggies are easier to see), and also because flight simulations are notorious for jagged edges. The long, straight lines of a planes’wing are a great place to spot them, while the edge of the horizon is a perfect place to spot shimmering. In this case all you have to do is look along the leading edge of the B-17’s wing and tail to see the jagged edges. We used the screen resolution of 800x600 for testing.
First, lets take a look with anti-aliasing turned off (taken with GeForce FX 5900 Ultra):
It’s pretty easy to see all the jagged edges, they’re easily visible on the leading edge of each plane’s wings.
Now lets look at 4xAA:
Jaggies are more apparent on the GeForce FX 5900 Ultra, even in 4xAA mode. The easiest place to spot them is on the lead B-17 in the picture, right between the two engines. While this area is relatively free of jagged edges on the RADEON 9800 PRO card, GeForce FX 5900 Ultra is filled with them.
SIDEBAR: Microsoft recently released DirectX 9.0a.
CodeCreatures Benchmark Pro
Serious Sam 2 - OpenGL
Quake III - High Quality
Comanche 4 demo
Unreal Tournament 2003 - flyby
Unreal Tournament 2003 - botmatch
Unreal Tournament 2003 - flyby
Unreal Tournament 2003 - botmatch
Unreal Tournament 2003 Flyby
Unreal Tournament 2003 Botmatch
Quake III – High Quality
After an absence of over eight months, NVIDIA has finally returned to the high-end segment with the GeForce FX 5900 Ultra. In fact, they’re not only back in the game, NVIDIA’s 256MB GeForce FX 5900 Ultra is the fastest graphics accelerator on the market, albeit with slightly lesser image quality than ATI’s RADEON 9800 PRO.
Lets look at the results, in real world game testing the GeForce FX 5900 Ultra finished on top in every benchmark except Tom Clancy’s Splinter Cell. This title in particular is clearly highly dependant on the graphics card’s fill-rate as evidenced by the older GeForce FX 5800 Ultra. With its 500MHz core clock, the GeForce FX 5800 Ultra finished ahead of all other graphics cards in our testing with Splinter Cell.
The addition of the 256-bit memory interface in particular really allows the GeForce FX 5900 Ultra to shine in high-resolution environments with antialiasing and anisotropic filtering, one of GeForce FX 5800 Ultra’s weaknesses in comparison to the offerings from ATI.
NVIDIA also addressed the noise issue with GeForce FX 5800 Ultra. Rather than incorporating heat pipe technology similar to that found on the original FX Flow cooling system, NVIDIA opted to go instead with a traditional heatsink/fan combination. Describing the GeForce FX 5900 Ultra’s heatsink as “traditional” however is probably a bit conservative, as the heatsink is so large it physically consumes the PCI slot adjacent to your AGP slot! Fortunately the cooling fan NVIDIA has utilized is much quieter than its predecessor, so you won’t have to crank up the volume while you’re gaming.
The question now facing NVIDIA is availability. After their track record with GeForce FX 5800 Ultra, many will be skeptical of the June availability date. In talking with NVIDIA however, they sound very confident that they’ll be able to hit their June target. In a remarkable achievement, NVIDIA went to full production from NV35’s first silicon. As a result, NV35 has been in production since February and they claim to have shipped tens of thousands of NV35 cores to their launch partners by the time you’ve read this. For now though, we’re going to take a wait and see approach. After all, GeForce FX 5600 Ultra and 5200 Ultra’s are still hard to come by at retail, with the latter being built on TSMC’s proven 0.15-micron manufacturing process.
If everything we’re seeing today holds true, the 256MB GeForce FX 5900 Ultra will be a worthy competitor to ATI’s RADEON 9800 PRO, but at $500, its asking price will be too steep for many to afford. That’s why we really want to see what NVIDIA’s 128MB variants of the GeForce FX 5900 will be capable of. Unfortunately, we’re going to have to wait a little longer to get that answer, but in the meantime start saving your pennies because all of these cards come with pretty hefty price tags.
SIDEBAR: Are you drooling over the 256MB GeForce FX 5900 Ultra, or were you expecting more? Perhaps you’re really interested in seeing 128MB numbers? Speak!
|© Copyright 2003 FS Media, Inc.|