Our 16-GPU Monster-Class Supercomputer
the world’s most compact (18"x18"x18") and inexpensive ($3000) supercomputer
new MIT CUDA Course
Built as part of a collaboration between the Cox Lab (Rowland Institute at Harvard) and the DiCarlo Lab (McGovern Institute for Brain Research at MIT).
A longer description is available at the Cox Lab web page of this project.
Performance
- more than 8 teraflops of theoretical performance
- around 4 teraflops of observed performance (mainly 3D convolutions)
Mission
- Primary purpose: brain-inspired artificial vision
- Secondary purpose: seizure prediction, collaborative filtering (e.g. Netflix), etc.
Environment
- GNU/Linux
- Python, C, C++, Cython
- CUDA, PyCuda
more pictures on flickr
Specs
- Graphics cards: 8x NVIDIA 9800GX2
- Motherboards: 2x MSI K9A2 Platinum 790FX
- CPUs: 2x AMD Phenom 9850 2.5 GHz Quad Core
- Memory: 8GB OCZ
- HDD: 2x Western Digital Velociraptor 300GB
- Cooling: 2x Rosewill CPU Coolers and 14x 120mm Fans
- Case: Mountain Mods Duality Custom Case
- PSUs: 2x 850W Zalman ZM850-HP and 2x 750W Silverstone
Cost
- Computer parts: less than $3000
- Graphic cards: generously donated by NVIDIA
keywords: MIT 16 GPU CUDA NVIDIA Monster Supercomputer