University of Antwerp builds desktop supercomputer with 13 NVIDIA GPUs

Thomas De Maesschalck Posted on by
Last year I wrote about the 4000EUR FASTRA desktop supercomputer with eight NVIDIA GPUs, and now the University of Antwerp's ASTRA group presents FASTRA II. This new system features a total of 13 NVIDIA GPUs and delivers up to 12 teraFLOPS. You can find more details in my FASTRA II article, the performance of this GPGPU system is really impressive.

Thomas De Maesschalck

About the Author

Thomas De Maesschalck

Thomas has been messing with computer since early childhood and firmly believes the Internet is the best thing since sliced bread. Enjoys playing with new tech, is fascinated by science, and passionate about financial markets. When not behind a computer, he can be found with running shoes on or lifting heavy weights in the weight room.
Thanks. Your comment is waiting for moderation.
Your comment was published.

Comments (8)

  1. Anonymous

    Holy cow! 12 teraflops from one box! Nice. That's a cluster-worthy system that carries some punch. And for just €6000!

    Reply
  2. Anonymous

    That's so cool Are you guys considering the move to OpenCL? I wonder how this system would perform with a bunch of highend Radeon 5970 video cards if so :p http://techreport.com/articles.x/17986/2 Peak pixel fill rate (Gpixels/s) GeForce GTX 295 - 32.3 Radeon HD 5970 - 46.4 Peak bilinear texel filtering rate (Gtexels/s) GeForce GTX 295 - 92.2 Radeon HD 5970 - 116.0 Peak bilinear FP16 texel filtering rate (Gtexels/s) GeForce GTX 295 - 46.1 Radeon HD 5970 - 58.0 Peak memory bandwidth (GB/s) GeForce GTX 295 - 223.9 Radeon HD 5970 - 256.0 Peak shader arithmetic (GFLOPS) Single-issue Dual-issue GeForce GTX 295 - 1192 1788 Radeon HD 5970 - 4640 N/A Good luck guys!

    Reply
    1. Anonymous

      i cant actually see the 5970's having an improvement over the 295s, as the 295s would be bottlenecked already on that motherboard. the 295's use a little more than 8 PCIe lanes and that motherboard can only support 4 PCIe slots with 16lanes adding any more cards reduces all of the slots to 8 lanes. adding the 5970 will only bottleneck it further and still give about the same performance.

      Reply
  3. Anonymous

    so cool... can both fastra I and II work together with cluster?? and a tutorial of building and combine supercomputer would help many people including me i think :)

    Reply
    1. Anonymous

      I dont understand the choice of 32-bit! centOS operating system. Why not use Windows Server 64-bit ! ? wouldnt that solve the memory addressing issues. ?

      Reply
      1. Thomas De Maesschalck

        The memory issues were caused by the BIOS, the GPUs required a larger block of physical address space than the BIOS was able to provide. Without the BIOS adjustments they made the system wouldn't even boot.

  4. Anonymous

    good motherboard suggestion? Hello Thomas, Very nice system indeed. I have a mixed experience with GTX 295 GPUs on different motherboards, all advertised with PCI16 2.0 slots. A setup with a X58 chipset - made by Supermicro and similar to the Asus P6T7 - giving a CPU to GPU transfer speed of 4.4 GB/s and another one with a P43 chipset - an Asus P5QL - giving only a speed of 1.1 GB/s. Of course quite a different price tag but I am somewhat deceived about the second one throughput performance. Anyone with a clever suggestion for a decent motherboard to host my GTX 295 GPUs preferably accepting Core 2 processors? Kind regards, Alain UCL - Belgium

    Reply
  5. PeterW

    Hello, Have you considered using a TYAN FT72B7015 4U barebone, (http://www.tyan.com/product_SKU_spec.aspx?ProductType=BB&pid=412&SKU=600000150)? It provides 8 double width PCIe x16 v2.0 slots. So you could be using 8 Teslas C1060 (8 GPUs) or 8 GTX295 (16GPUs) without make changes to the hardware. Peter

    Reply

Reply to Anonymous

Replying to Anonymous