RecessionCone's Supercomputer!

How is the configuration? Processor? RAM? Is it one GPU/PC or multi-GPU/PC? I assume you're using rack server and not regular PC casing? So far, are all the hardware ran reliably or you have a defective part somewhere? I ask this because last year I bought 4x8GB RAM (corsair) and one of the stick is defective. Since you obviously bought a lot, do you experience any?
 
With so many GPUs going there will be noticeable attrition. High-end GPUs have historically NOT been very reliable, compared to computing devices in general. :p
 
Why don't you buy K20s or similar computing optimized GPUs like computing centers do?
!/$ ?
If your application needs lots of RAM, lots of SP FLOPS, but doesn't need ECC or whatever other feature a K20 provides, then the Titan X is an obvious choice.
 
Driver management/support ... does this come into play? Is there ever a need to upgrade drivers, or perhaps in this instance just add 80 more gpu's and use existing driver?
 
These are all good reasons to not buy a dedicated GPGPU card. But I am wondering about the specific application then...as it is quite unusual (I mean, there is a good reason that those Tesla's exist imo).
 
Neural networks fit that bill quite nicely. They are happy with fp32 (there are even uses for fp16 hence Pascal), they burn lots of RAM, and Maxwell is also more efficient with its flops then Kepler.
 
The object of the thread hasn't returned to give even any hints about what his supercomputer is doing... Can we get at least a teaser, Mr. @RecessionCone Sir, please? :)
 
Back
Top