It's the most hyped product Nvidia has ever released, and a major milestone for the company. But can the Geforce FX become the new 3D card of choice? Jogn Gillooly screams the answer over the howl of its cooler.
It is the most hyped product NVIDIA has ever produced, and a major milestone for the company. But can the GeForce FX become the new 3D card of choice? John Gillooly screams the answer over the howl of the FX.
Over the past few years we have been able to use NVIDIA's six month product cycle to calibrate our calendars. Without fail the graphics giant stuck to this Moore's Law on speed development cycle, riding it all the way to number one. But this relentless pace has hit its first huge roadblock in the form of the GeForce FX.
Months after this next generation GPU was announced at Comdex, actual cards are virtually nonexistent. Rumours emerged that NVIDIA was not only controlling all the production of the high end GeForce FX 5800 Ultra cards, but only 100,000 of the beasts would ever leave TSMC's door. While NVIDIA strongly denies these rumours, word coming from many sources in Taiwan is that they are 100% correct.
The reality of this means that we will be unlikely to see retail cards on the market until April, and then it will still take a while for volume to appear and prices to drop to sane levels.
There will be two models of GeForce FX 5800 on the market. Most of the hype so far has been directed at the 500MHz core, 1GHz DDR-II GeForce FX 5800 Ultra model, but reality and product availability means that 99% of cards that we will see will be based on the 400MHz core, 800MHz DDR GeForce FX 5800 model. It is an unfortunate reality but most of the limited production run of Ultra cards will end up on sale in the US marketplace rather than down here.
Rather than twiddle our thumbs, patiently waiting for retail cards, we called upon the Atomic Spy Ring, who called on 3 months spent playing Splinter Cell to formulate a cunning plan to get an MSI GeForce FX 5800 engineering sample into the labs for some intensive benchmarking. Now this engineering sample is a curious beast. Based on A02 revision GeForce FX silicon, it uses the gigantic FX-Flow cooler that is required for the GF FX 5800 Ultra cards.
It is unlikely that any of the normal GeForce FX 5800 cards will use these coolers. The slower models will simply not need such an extreme solution, and using such a meaty contraption would only inflate the costs and make the card needlessly loud. The card will also be made by each company, rather than doled out by NVIDIA, which will make for a much more diverse range of cards and more reasonably priced solutions.
It is fair to say then, that while the performance of the card is indicative of how the final GeForce FX 5800 will go; the looks will be completely different. Going by our previous experience will early revisions of NVIDIA silicon, such as the GeForce3, there will undoubtedly be some performance improvements before final silicon is available, at which point we will have a second look at the retail product.
Our strong interest in the FX 5800 is based around one significant point:- Memory bandwidth. Competition between 3D chip makers has revolved around what features are supported, and there has usually been a trade-off between different architectures based on this. NVIDIA and 3dfx were historically known for duking it out over the relative benefits of 16 and 32 bit colour, or high resolution vs. antialiasing. ATI and NVIDIA have for the past couple of years leapfrogged each other with more advanced shading units. But inherent to all of the feature set gains has been memory bandwidth. New features only ever work at speed when they aren't being choked by the memory bus.
With the RADEON 9700 Pro ATI smashed this barrier for a generation or two by biting the bullet and introducing a complex and expensive to implement 256-bit memory bus. It initially pumped the costs of the card up, but as any quick check of pricing will show, this initially high price has dropped down to a very reasonable level.
NVIDIA took a different tack with the memory bandwidth issue, choosing to use the expensive and rare DDR-II memory running a 1GHz for its GeForce FX 5800 Ultra card and 800MHz for the GeForce FX 5800. In the case of the 5800 Ultra this translates to memory bandwidth of 16GB/sec, for the 5800 it works out at 12.8GB/sec. The actual numbers quoted by NVIDIA are greater, 48GB/sec for the 5800 Ultra and 25.6GB/sec for the 5800. This is because of the implementation of lossless colour compression through the pipeline. Thanks to the 256-bit memory bus, ATI manages to get a cool 19.4GB/sec out of the RADEON 9700 Pro's 620MHz DDR RAM.
Memory bandwidth is of utmost performance because it is the main influence on video performance under antialiasing and anisotropic filtering. With video hardware at the point where running games at normal setting barely impacts performance, anyone who is looking at these high end beasts will want to be taking available of these image quality enhancers.
Our other big concern is how the GeForce FX's shading units perform. On paper NVIDIA has a much more advanced shading unit than ATI, and the relative performance under DirectX 8 and 9 is a major selling point for these cards.
Beat that bandwidth
To sort out just what these disparate feature sets and clock speeds do for performance we lined the MSI GeForce FX 5800 up against a HIS RADEON 9700 Pro. While the GF FX 5800 will end up pitched at the non-Pro RADEON 9700, at launch we can expect the GF FX 5800 and the RADEON 9700 Pro to be priced in the same arena. And as you will see, performance between the cards is interesting, but it is the relative performance hits that really tell the interesting tale.
Testing was done using an i845PE motherboard, 3.06GHz Pentium 4 and 1GB DDR RAM. The software environment was Windows XP Pro with Service Pack 1, DirectX 9, Catalyst 3.1 drivers for the RADEON and Detonator 42.68 drivers for the GeForce FX.
Our first test is the DirectX 9 benchmark, 3DMark03. We have only included a small number of 3DMark03 tests, not only because we are still evaluating the benchmark, but also because the Detonator 42.68 drivers were supplied to use for benchmarking NVIDIA cards with 3DMark03, and hence may contain optimisations. We tested at normal settings, and then with post processing, which uses the shaders in the DX8 and DX9 tests to add depth of field and overbrightening effects. Perhaps the most telling results are the Mother Nature ones. While the GF FX sits behind the RADEON in the DX8 tests, in this pixel shader overload test the GF FX maintains parity with the RADEON, and on a purely subjective note, the test looks crisper and cleaner on the GeForce FX than the RADEON.
With an initial close victory to the GF FX's shading unit, we pulled out the meatiest of the DX 8 benchmarks, Codecreatures, and ran the cards. Again, the GeForce FX is slightly slower than the RADEON 9700 Pro, but it maintains an even gap as resolution increases. This picture changes dramatically when performance shifts from relying on the shader unit to relying on memory bandwidth. We set up three configurations for testing antialiasing and anisotropic filtering performance: - none; 2x AA (Quincunx for the GF FX) and 2x Anisotropic filtering; and the heavy duty maximum settings of 6x AA and 8x Anisotropic filtering for the RADEON 9700 and 8xS AA and 8x Anisotropic filtering for the GF FX.
Our first test of these settings was 3DMark2001SE Pro. Again, while the GF FX is slower overall than the RADEON 9700 Pro, this margin is not by much. The really interesting part of these results is the hit seen when running the maximum settings. In this case the RADEON 9700 Pro's 256-bit memory bus shows tha
Issue: 133 | February, 2012