NVIDIA GF100 Architecture Details

[youtube j9F3W-v6PNI]

GT100 tesselation demo

After the first global overview in September 2009, NVIDIA has released new details on its new Fermi GT100 architecture. Here is a summary of NVIDIA’s GT100 architecture features in equations:

• The CUDA core is the primary working unit of the GF100 (Each CUDA core is fully IEEE 754-2008 compliant) – GF100 = 512 CUDA cores
• Streaming Multiprocessor (SM) = 32 CUDA cores – GF100 = 16 SM
• 4 SFU per SM (SFU – Special Function Unit – executes transcendental instructions such as sin, cosine, …) – GF100 = 64 SFUs
• Graphics Processing Cluster (GPC) = 4 SM – GF100 = 4 GPC
• 1 Raster Engine per GPC (raster engine = rasterization, z-culling). A raster engine processes 8 pixels per clock – GF100 = 32 pixels per clock
• 1 PolyMorph Engine per SM (PolyMorph Engine: execution unit that handles geometry for GF100: vertex fetch, tessellation, viewport transform, attribute setup, and stream output) – GF100 = 16 PolyMorph Engines
• 4 Texture Units per SM – GF100 = 48 Texture Units
• 6 partitions of 8 ROPs (ROPs perform blending or AA) – GF100 = 48 ROPs

GT100 die

SM (Streaming Multiprocessor) detail

CUDA core detail

The tessellator of the PolyMorph Engine is THE BIG FEATURE of GT100:

[youtube K3m9rPltA_s]


References:

• NVIDIA’s GF100 official page
• NVIDIA’s GF100: Architected for Gaming @ AnandTech
• NVIDIA GF100 Architecture Preview – Fermi brings DX11 to the desktop @ PC Perspective
• Nvidia GeForce GF100 : la révolution géométrique ? @ Hardware.fr