Windows 7 WARP brings DX10 rendering to the CPU

Hoping to avoid a repeat of the Windows Vista-capable debacle, where Intel integrated graphics offerings weren't capable of handling many Vista eye candy effects, Microsoft has developed a DX10/DX10.1 software rasterizer for Windows 7, allowing all 3D rendering to be handled by the CPU. Potentially faster CPUs would be better suited to handle these tasks than Intel's IGP. 

The Microsoft technology, dubbed WARP (Windows Advanced Rasterization Platform) 10 supports DirectX 10 and DX10.1 as well as 8x MSAA and anisotropic filtering. According to this Microsoft document, WARP10 has been optimized for use on multi-core CPUs:

As WARP10 makes extensive use of multiple CPU cores, the best performance of the rasterizer will be found on modern quad core CPU’s. WARP10 also runs significantly faster on machines with SSE4.1 extensions and we have done significant testing and performance tuning on machines with eight or more cores and SSE4.1 as we believe these high end machines will be more and more common during the lifetime of Windows 7.

When WARP10 is running on the CPU we are limited compared to a graphics card in a number of ways. The front side bus speed of a CPU is typically around or under 10GB/s where as a graphics card often has dedicated memory that is able to take advantage of 20-100GB/s or more of graphics bandwidth. Graphics hardware also has fixed function units that can perform complex and expensive tasks like texture filtering, format decompression or conversions asynchronously with very little overhead or power cost. Performing these operations on a typical CPU is expensive in terms of both power consumption and performance cost in cycles.

...

The typical performance numbers we are seeing on an Intel Penryn based 3.0GHz Quad Core machine show that WARP10 can in some cases even outperform low end integrated Direct3D 10 graphics GPU’s on a number of benchmarks!

Low end discrete graphics hardware is typically 4-5x faster than WARP10 at running these benchmarks and obviously, these GPU’s have minimal use of CPU resources as well. Mid-range or high-end graphics cards are significantly faster than WARP10 for many applications particularly when an application can take advantage of the massive parallelism and memory bandwidth these graphics cards provide.

We don’t see WARP10 as a replacement for graphics hardware, particularly as reasonably performing low end Direct3D 10 discrete hardware is now available for under $25. The goal of WARP10 was to allow applications to target Direct3D 10 level hardware without having significantly different code paths or testing requirements when running on hardware or when running in software.

The Microsoft article goes on to state Crysis benchmarks with a variety of CPUs running the game in DX10 low quality mode at 800x600. Core i7 965 ran the game at 7.36 fps, while the Q9650 topped out at 5.69 fps and 2.4GHz Core 2 Duo 2.62 fps. AMD's Phenom 9550 ran the game at 3.01 fps.