Build Your Own


Guide to Building Your Own Tesla Personal Supercomputer System    

This guide is to help you build a Tesla Personal Supercomputer. If you have experience in building systems/workstations, you may want to build your own system. Otherwise, the easiest thing is to buy an off-the-shelf Tesla Personal Supercomputer from one of these resellers.

As with building any system, it is at your own risk and responsibility. There are many possible components to choose from when you build such a system. NVIDIA provides general guidance, but cannot test every configuration and combination of components.

We also get a lot of questions about why Tesla products are better than consumer products for building personal supercomputers. Click here for more details.

Minimum Specifications of Main Components

These minimum specifications are for folks who want to label their system a “Tesla Personal Supercomputer”. You may choose to build a workstation with fewer Tesla GPUs in it.

  • 3x Tesla C2050
  • Quad-core CPU: 2.33 GHz (Intel or AMD)
  • 12 GB of system memory (4GB of system memory per Tesla C2050)
  • Linux 64-bit or Windows XP or Win7 64-bit
  • System acoustics < 45 dBA
  • 1200 W power supply

Example of Complete 4 Tesla C2050 System Configuration

This is a list of suggested components to build a 4x Tesla C2050 PSC. Several of these components such as the memory, CPU, power supply, case, can be substituted with an appropriate equivalent and suitable component. We do not certify any components for the PSC; this is left to the system builders.

4 Tesla C2050 Configuration
Motherboard Tyan S7025
PCI-e bandwidth 4x PCI-e x16 Gen2 slots
Tesla GPUs 4x Tesla C2050
CPU Dual-socket Intel Xeon Nehalem
Memory 24 GB (6x 4GB) DDR3 DIMMs (motherboard takes up to 64 GB)
Power Supply Coolmax CUQ-1350B 1350W
Case Lian Li PC-P80
Hard drive 640 GB
DVD drive DVD burner
CPU fan, heat sink For the Intel Xeon

Pre-built SuperMicro SuperServer 7046GT-TRF
4 GPU system, dual-socket Nehalem, up to 96 GB of memory

3 Tesla C2050 Configurations
  3 Tesla Configuration + 1 Quadro 6000
PCI-e bandwidth 4 PCI-e x16 Gen2 Slots 4 PCI-e x8 Gen2 Slots
Motherboards ASUS P6T7 WS SuperComputer AsRock X58 SuperComputera
Tesla GPUs 3x Tesla C2050 3x Tesla C2050
Display GPU 1x Quadro FX or NVS 1x Quadro FX or NVS
CPU Intel Xeon E55xx / Core i7 Intel Core i7
Min Memory 24 GB DDR3 24 GB DDR3
Power Supply 1350W 1350W

The Tesla C2050 computing processor is a dual-wide PCI-e x16 Gen2 board. It can be used in a Gen 1 PCI-e x16 slot as well, but this leads to lower system bandwidth between the CPU and GPU, which may impact application performance (depends on the application).

So, you need to use motherboards that have 3 or 4 PCI-e x16 slots that are dual slot apart from each other.We also recommend more recent chipsets, such as Intel Tylersburg.

Choice of CPU is determined by the motherboard you use. We recommend that you use at least a 2.33 GHz quad-core CPU such as:
  • Intel Xeon or Core i7 quad-core
  • AMD Phenom or Opteron quad-core
System Memory
Since each Tesla C2050 has 3GB of GPU memory, we recommend at least 3GB, preferably 4GB, of system memory per Tesla C2050. So, for a 3x Tesla C2050 system, include at least 12 GB of system memory and for 4x Tesla C2050 system, configure with at least 16GB of system memory.

Power Supplies
There are several choices for power supplies. Choose a good power supply that is rated at least 1350 Watts. An example is the Coolmax CUQ-1350B 1350W power supply. While we recommend each Tesla C2050 to have one 6-pin and one 8-pin, some power supplies such as Coolmax CUQ-1350B can provide enough load with two 6-pin power connectors. It is extremely important to understand specifications when selecting power supplies.

Hard Drive, DVD, other components
Hard and DVD drive choices are left to you. In general, it is best to have at least 160GB of hard drive.

Computer Cases / Chassis
The case / chassis choice is important, since when you plug in four Tesla C2050 boards in, it needs a case that has 8 slots (which is larger than the usual ATX case).

Commercially available off-the-shelf cases with 8+ slots are
  • Lian-Li PC-P80
  • ThermalTake ArmorPlus

It may be possible to take a 7-slot chassis and cut an extra 8th slot to accommodate the four dual-wide GPU boards. Try at your own risk!

System Cooling
Some system cases (like the Lian-Li) come with system fans of their own. In general, we recommend at least one case fan placed so that it is blowing air on the side of the Tesla boards (i.e., blowing air directly at the motherboard). The key is to keep the ambient temperature inside the case below 45C.

Operating Systems
We recommend either Linux 64-bit or Window XP 64-bit for best operation of this high-memory, high-performance system. If you prefer Win7, we recommend our TCC driver for highest CUDA performance. TCC driver is exclusive to Tesla product family. The latest TCC driver can be found in the normal driver download page on nvidia.com; for an older TCC driver, go to: //www.nvidia.com/object/tesla-win7-64bit-259.03-whql-driver.html. Please make sure you read the Release Highlights on what is not supported with TCC driver.

Verifying your System
Once you assemble your system and install an operating system on it, please download the CUDA driver, CUDA toolkit and optionally the CUDA SDK examples from CUDA Zone.

After you download the CUDA driver, toolkit and SDK examples, run the following commands from the CUDA SDK.
  • deviceQuery
    • This will report the number of Tesla GPUs in the system
  • bandwidthTest --memory=pinned --device=N
    • Run for each C2050, where N = 0, 1, 2 for 3 C2050s in the system
    • This will report the PCI-E bandwidth to and from the CPU and each of the GPUs
    • Peak PCI-E x16 Gen2 bandwidth is between 5 and 6 GBytes/sec. PCI-E x16 Gen1 and PCI-E x8 Gen2 are about half that.
  • nbody --benchmark --n=131072 --device=N
    • Run simultaneously for as many instances as Tesla GPUs in your system.
    • That is, for 3 Tesla C2050s, run 3 instances with N = 0,1,2 simultaneously
    • This will run the nbody program on all the Tesla GPUs

Reporting problems/questions
NVIDIA offers no direct support to individuals building their own Tesla Personal Supercomputers.  However, we recommend using the CUDA forums to ask other CUDA developers about their experiences in building these systems.

Learn more about Tesla C2050