next »

[isthisthingon]

"The Fermi chip has 3 billion transistors and 512 processor cores, and includes a new technology called GigaThread 3.0 that can manage thousands of threads in parallel, Huang said."



http://www.pcworld.com/article/172934/nvidia_says_its_new_fermi_cpu_will_run_supercomputers.html?tk=rss_news

[perkiset]

"...thousands of threads in parallel..."

[kurdt]

Oh my god... I want that!

[vsloathe]

4 of them in quad SLI imo

[isthisthingon]

4 of them in quad SLI imo

With enough internal parallel processing I wonder if the SLI synchronization would have a point of diminishing returns.  They are still bound to the PCIe bus after all.  I think the dual SLI even with current GPUs is close to the sweet-spot no?  But talk about liquid cooling and insane wattage requirements.

[kurdt]

4 of them in quad SLI imo

With enough internal parallel processing I wonder if the SLI synchronization would have a point of diminishing returns.  They are still bound to the PCIe bus after all.  I think the dual SLI even with current GPUs is close to the sweet-spot no?  But talk about liquid cooling and insane wattage requirements.

I always thought that SLI doesn't do anything good for parallel data crunching but instead it's more for graphical applications.

[isthisthingon]

It's totally GPU.  SLI makes it so you can have say 2 PCIe cards in one box that share the graphics processing load.  However, unless I'm mistaken, their jobs require synchronization between them since GPU processing is not discreet.  So the PCIe bus has to pass sync data between them.  Therefore having 2 PCIe cards doesn't net you exactly twice the performance.  It's faster but not twice as fast. 

But I'm not the expert in this arena for sure.  I think vs is the Cache heavyweight when it comes to gaming hardware.  He described one of his systems and I began muttering like Beaker from Sesame Street.

[kurdt]

It's totally GPU.  SLI makes it so you can have say 2 PCIe cards in one box that share the graphics processing load.  However, unless I'm mistaken, their jobs require synchronization between them since GPU processing is not discreet.  So the PCIe bus has to pass sync data between them.  Therefore having 2 PCIe cards doesn't net you exactly twice the performance.  It's faster but not twice as fast. 

But I'm not the expert in this arena for sure.  I think vs is the Cache heavyweight when it comes to gaming hardware.  He described one of his systems and I began muttering like Beaker from Sesame Street.

Yeah but like you guessed, vsloathe is talking about gaming but I'm talking about CUDA. Last time I checked SLI wasn't supported very well and everybody who I asked recommended to write your own task dispatch system because all cores are viewed as individuals. I don't know enough to say if PCIe bus slows it down but at least for now I don't see any reason why it would affect a lot if you are sending 1+1 to core 1 and 2+2 to core 2 and so on. But with very large processing pieces syncing might become a little issue. However I think at least with my CUDA coding skills the bottleneck is my code and not the hardware

But this new GigaÜberThread system is most likely added to CUDA as well so it might make CUDA even more powerful by default.

[vsloathe]

Read about something that is coming out soon that is a chip which will allow mobo manufacturers to create basically SLI that is completely agnostic with regard to the type of cards it's using. For instance you can use a 512MB Radeon with a 1GB Geforce 260GT and they'll play together just fine.