During ATI's "Shader Days" event, Valve's Gabe Newell conducted an unexpected presentation, sharing a set of benchmarks obtained from Half-Life 2, the game Microsoft themselves recommended yesterday, as the DirectX 9.0 standard. Mr Newell's presentation included some exciting and interesting information about the use of DirectX 9.0 features in H-L 2 but was far more interesting as far as Valve's conclusions about graphic solutions were concerned. Gabe Newell also clearly demonstrated his frustration at a certain company over the recent attempts to influence the outcome of benchmarks using driver-based optimizations.
The conclusion of the presentation was clear The Radeon 9800 Pro is twice the speed of the GeForce FX 5900 Ultra.
Half-Life 2, DX9 and Vice-Versa
According to the Newell presentation Half-Life 2 is a proper DX9 application and :
-is properly written to take advantage of DX9
-Uses a wide range of DX9 functionality
-Half-Life 2 is an accurate predictor of future DX9 application performance
-We see H-L 2 as a new benchmark for the type of amazing experiences that can be delivered on the Windows platform
DX9 Features
-High Dynamic Range Lighting (HDR)
-Bump Mapped Characters
-Soft Shadows
-Improved full scene Anti-Aliasing
DX9 Performance
-More efficiently exposes hardware
-9800 PRO, DX9 has same framerate as DX8 with more features
DX9 Ease of Implementation
-New source functionality will be DX9 specific
It was interesting to get Valve's view on recent attempts by hardware vendors to produce unrealistically high benchmark results. Newell presented the reasons why Valve are so concerned about such attempts. According to the presentation those are:
-Valve customers have traditionally made purchase decisions based heavily on benchmark data
-Because of engineered benchmark methods benchmarkk data has decorrelated from real-world performance.
-Valve customers will be pissed.
Although Newell heavily hinted at nVidia he did not restrict his comments to any one hardware manufacturer. He also cited cases during H-L 2 development where fog was completely removed from a level of the game by the driver in order to improve performance. The main methods used to produce such unrealistic results are:
-Camera path-specific occlusion culling
-Visual quality trade offs e.g. lower filtering quality, disabling fog.
-Screen-grab specific image rendering
-Lower rendering precision
-Algorithmic detection and replacement
-Scene-specific handling of Z writes
-Benchmark-specific drivers that never ship
-App-specific and version specific optimizations that are very fragile
As far as the performance of the top-of-the-range cards is concerned, Newell was categorical; The low-end GeForce FX 5200 Ultra and mid-range GeForce FX 5600 Ultra are wholly unplayable at DX9 level. The top-of-the-line GeForce FX 5900 Ultra manages just over 30 fps and is well behind ATI's mid-range Radeon 9600 Pro. This information leads to the conclusion that the Radeon 9800 Pro is twice as fast as nVidia's GeForce FX 5900 Ultra.
According to Newell, Valve spared no expense or effort in order to optimize for the NV3x cards. According to Valve, five times as much time was spent optimizing the NV3x path as was spent optimizing the generic DX9 path. Newell and Valve also claim that they were surprised by the discrepancy and that ATI hardware did not require the optimization.
Although the optimization did produce better results for the GeForce 5900 Ultra, Newell claimed that such optimized performance gains would disappear in the future as new DX9 functionality will be able to use fewer and fewer partial precision functions. He even went as far as to suggest that the best possible optimization would be to :
Treat NV3x as DX8 hardware. He went on to add that most developers would not have the budget to work on nVidia specific optimizations and that 5200/5600 card owners should not even attempt to play H-L 2 with DX9. In that light however, it did seem odd that under DX8 the aging GeForce Ti 4600 outperformed both of the brand new 5200 and 5600 boards. Are you confused yet?
It seems that Newell's presentation attempted to give a clear view of under what conditions H-L 2 will look best. According to Valve then the full H-L 2, DX9 experience can only be enjoyed using either a RADEON 9600 PRO OR 9800 PRO.
In order to answer possible future criticism of his presentation Gabe Newell concluded the presentation by stressing that the following point is a possible rumour:
ATI's OEM deal with Valve is the real rerson for these benchmark results He added that these results were the reason Valve had an OEM deal with ATI and referred any future critics to Microsoft who, apparently, can confirm Valve's honest results and intentions. He also stressed the company's commitment to working with all companies in order to best serve their customers.