NVidia has published a whitepaper of their upcoming Fermi (GT300) architecture detailing the new features and various case-study results. The cliff-notes version:
- 32 CUDA Cores per Streaming Multiprocessor
- 8x double precision performance over GT200
- Unified Address Space with full C++ Support
- Full IEEE 754-2008 32-bit and 64-bit Precision
- Predication
- ECC Memory SUpport
- 10x faster context switching
- Out-of-Order thread block execution
- 3Billion Transistors
- 768K L2 Cache (didn’t exist in GT200)
- Up to 16 concurrent kernels (Didn’t exist in GT200)
Looks like Fermi is going to be an awesome processor to work with. Read the full report on their site.
NVIDIA_Fermi_Compute_Architecture_Whitepaper.pdf (application/pdf Object).