2. Solved
  3. LocalAI
  4. Accelerating LocalAI with an Nvidia Tesla M40: right cables and 3D printing

Running AI locally is hot, but decent hardware is expensive. As a Digital Problem Solver, I prefer to look at what can be done with things that are already there. My goal? Setting up a cheap, powerful local AI environment (LocalAI)

Nvidia Tesla m40 from Server Room to Desktop

Who says enterprise hardware only belongs in data centers? As a Digital Problem Solver, I like to delve into technical puzzles to find practical solutions. My latest project? Install an Nvidia Tesla M40 server graphics card into a standard desktop PC to run AI models locally. No hassle, just good technology.  

Of course, that didn't happen without a fight. To get this server hardware running safely in a normal enclosure, I had to tackle two major challenges: power and cooling.

1. The Power Puzzle: Prevent a sea of ​​flames

The biggest pitfall with enterprise cards is the power supply. Although the 8-pin connection on the Tesla M40 looks exactly like a standard PCIe connector, the plus and minus poles are exactly reversed. If you plug in a standard PCIe cable from a modular power supply (such as my Corsair CX750M) here, you immediately send the card to the scrap heap.  

Because I think in terms of requirements and processes, I have prepared this tightly:

  • The correct pin-out: The Tesla M40 expects an EPS/CPU power signal instead of PCIe. 
  • The power supply solution: an extra PCIe cable to power the Tesla (CP-8920143 )
  • The solution: A specific Tesla GPU adapter cable ( type 030-0571-000 ) which internally converts the poles exactly right. This way the card safely receives the power it needs.

2. Cooling: From passive heatsink to active airflow

Another unique feature of the Tesla M40 is that it is passively cooled. In a server room, the case fans are blowing so loudly that a gale of wind is constantly blowing through the heatsink. In a normal desktop case, the air becomes stagnant and the card overheats within a few minutes.  

As a developer, I like pragmatic solutions. So I fired up the 3D printer:

  • 3D-printed Blower Duct: I had a special duct from Thingiverse ( https://www.thingiverse.com/thing:6038375 ) printed that fits perfectly on the end of the M40.
  • Active cooling: I can mount a powerful blower fan on this. With this we imitate the targeted airflow of a server, but in a jacket that fits in my desktop.

The result?

With the right cables and 3D printing in hand, the foundation has been laid. The platform is ready to be built iteratively, so that I can soon run a stable, local Docker environment with AI models. A wonderful hardware puzzle with a functional result!