Tessellation
In today's complex graphics, tessellation offers the means to store massive amounts of coarse geometry, with expand-on-demand functionality. In the NVIDIA GF100 GPU, tessellation also enables more complex animations. In terms of model scalability, dynamic Level of Detail (LOD) allows for quality and performance trade-offs whenever it can deliver better picture quality over performance without penalty. Comprised of three layers (original geometry, tessellation geometry, and displacement map), the final product is far more detailed in shade and data-expansion than if it were constructed with bump-map technology. In plain terms, tessellation gives the peaks and valleys with shadow detail in-between, while previous-generation technology (bump-mapping) would give the illusion of detail.
Using GPU-based tessellation, a game developer can send a compact geometric representation of an object or character and the tessellation unit can produce the correct geometric complexity for the specific scene. Consider the "Imp" character illustrated above. On the far left we see the intial quad mesh used to model the general outline of the figure; this representation is quite compact even when compared to typical game assets. The two middle images of the character are created by finely tessellating the description at the left. The result is a very smooth appearance, free of any of the faceting that resulted from limited geometry. Unfortunately this character, while smooth, is no more detailed than the coarse mesh. The final image on the right was created by applying a displacement map to the smoothly tessellated third character to the left.
Tessellation in DirectX-11
Control hull shaders run DX11 pre-expansion routines, and operates explicitly in parallel across all points. Domain shaders run post-expansion operations on maps (u/v or x/y/z/w) and is also implicitly parallel. Fixed function tessellation is configured by Level of Detail (LOD) based on output from the control hull shader, and can also produce triangles and lines if requested. Tessellation is something that is new to NVIDIA GPUs, and was not part of GT200 because of geometry bandwidth bottlenecks from sequential rendering/execution semantics.
In regard to the GF100 graphics processor, NVIDIA has added a new PolyMorph and Raster engines to handle world-space processing (PolyMorph) and screen-space processing (Raster). There are sixteen PolyMorph engines and four Raster engines on the GF100, which depend on an improved L2 cache to keep buffered geometric data produced by the pipeline on-die. The end result is an 8x increase in geometric performance over the GT200 GPU in the GeForce GTX285.
Not surprisingly, NVIDIA's investment into increased geometric power has resulted in a giant leap over similar output from ATI's Radeon HD 5870. Using the Unigine DX11 benchmark, the NVIDIA GF100 performed at 43 FPS on average compared to 27 FPS for the Radeon HD 5870. Using the Microsoft DirectX-11 SDK, the NVIDIA GF100 renders six cubemap faces in one pass, and compared to the HD5870 the geometry was produced 4x faster.
Four-Offset Gather4
The texture unit on previous processor architectures operated at the core clock of the GPU. On GF100, the texture units run at a higher clock, leading to improved texturing performance for the same number of units. GF100's texture units now add support for DirectX-11's BC6H and BC7 texture compression formats, reducing the memory footprint of HDR textures and render targets.
The texture units also support jittered sampling through DirectX-11's four-offset Gather4 feature, allowing four texels to be fetched from a 128×128 pixel grid with a single texture instruction. NVIDIA's GF100 implements DirectX-11 four-offset Gather4 in hardware, greatly accelerating shadow mapping, ambient occlusion, and post processing algorithms. With jittered sampling, games can implement smoother soft shadows or custom texture filters efficiently. For example, NVIDIA's own testing indicates that the accelerated jitter sampling on the GF100 is 3.3x faster than Radeon HD5870.
The previous GT200 GPU did not offer coverage samples, while the GF100 can deliver 32x CSAA. The jump from 8x AA to 32x CSAA appears to only reduce GF100 performance by approximately 7% in NVIDIA's lab tests.
GF100 Compute for Gaming
As developers continue to search for novel ways to improve their graphics engines, the GPU will need to excel at a diverse and growing set of graphics algorithms. Since these algorithms are executed via general compute APIs, a robust compute architecture is fundamental to a GPU's graphical capabilities. In essence, one can think of compute as the new programmable shader. GF100's compute architecture is designed to address a wider range of algorithms and to facilitate more pervasive use of the GPU for solving parallel problems. Many algorithms, such as ray tracing, physics, and AI, cannot exploit shared memory-program memory locality is only revealed at runtime. GF100's cache architecture was designed with these problems in mind. With up to 48 KB of L1 cache per Streaming Multiprocessor (SM) and a global L2 cache, threads that access the same memory locations at runtime automatically run faster, irrespective of the choice of algorithm.
NVIDIA Codename NEXUS brings CPU and GPU code development together in Microsoft Visual Studio 2008 for a shared process timeline. NEXUS also introduces the first hardware-based shader debugger. NVIDIA's GF100 is the first GPU to ever offer full C++ support, the programming language of choice among game developers. To ease the transition to GPU programming, NVIDIA developed Nexus, a Microsoft Visual Studio programming environment for the GPU. Together with new hardware features that provide better debugging support, developers will be able enjoy CPU-class application development on the GPU. The end results is C++ and Visual Studio integration that brings HPC users into the same platform of development. NVIDIA offers several paths to deliver compute functionality on the GF100 GPU, such as CUDA C++ for video games.
Image processing, simulation, and hybrid rendering are three primary functions of GPU compute for gaming. Using NVIDIA's GF100 GPU, interactive ray tracing becomes possible for the first time on a standard PC. Ray tracing performance on the NVIDIA GF100 is roughly 4x faster than it was on the GT200 GPU, according to NVIDIA tests. AI/path finding is a compute intensive process well suited for GPUs. The NVIDIA GF100 can handle AI obstacles approximately 3x better than on the GT200. Benefits from this improvement are faster collision avoidance and shortest path searches for higher-performance path finding.
In a PhysX fluid simulation example, there was a 2x improvement in the GF100 over the previous GPU architecture (67 in GT200 to 141 FPS on the GF100). Concurrent kernel calculations receive a 20-40% improvement using Phys X 3.0. In fact, NVIDIA suggests that the new architectural enhancements offer a 2x-3x improvement in compute performance in most cases. Other improvements, such as geometric realism, occur at 8x over GT200 while image quality at 32x CSAA produces 3x faster shadow maps for 2x over GT200.
continued
