Nvidia's Tesla P40 is a great graphics card for my need. But before I begin with my evaluation let me quickly summarize what the P40 is about:
- It's from Nvidia's Pascal generation GP102 high end chip which means it is roughly comparable with the GTX 10XX series of consumer graphics cards
- It has no connectors so you need to get the video signal out some other way (integrated graphics or a cloud gaming like setup)
- It has a whopping 24GB of RAM (for the time) which is good for modern games but great for Machine Learning, Hardware Acceleration and running Large Language Model inference
- It runs a relatively low clock speed only slightly better than the 1050/1050 TI (which accidentally is the GPU I am replacing)
- It's passively cooled so you need to get creative. The most effective way is a radial fan and a 3d printed adapter both you can grab on ebay. But it is said to be very loud and annoying.
- It is cheap. More than one order of magnitude to a more modern GPU with similar amounts of RAM.
I bought one for less than 300€ (including customs and a fan). The fan has yet to arrive though but I could not wait and wanted to test it. My first idea which was inspired by some videos on the internet was to remove the outer casing and strap a pair of large coolers to the heat sink. So I tried to open it but I couldn't get the screws on the top or bottom open. They are too small for any tool I own and the screwdriver I bought that should work just doesn't. I'm not even quite certain that they are hex screws since I can hardly see the shape under the lacquer.
Luckily I had a brilliant idea which you can see below. I folded a piece of paper such that on one side it matches a 80mm fan an on the other side matches the opening. Then I taped everything into the right shape. After struggling because it was slightly to large and the power supply cable of the GPU did not quite fit I got it working. And I am shocked by the efficiency of it. In my research I found a lot of people who did not manage enough airflow using such fans and keeping noise levels bearable. I also have a theory (using an 8th graders intuition of physics which is surely very wrong): My adapter is way longer and compresses the air evenly into the opening. This might be advantageous in comparison to very short or even bent adapters where the fan mostly blows against walls. And for the record: My 80mm fan is 12V and draws up to 0.6A and I used a A4 sheet (long side taped to the fan/jammed into the gpu).
The next thing was to get it running under arch linux but that's another story. In the end I got it working. My setup is a laptop with a 12th gen intel CPU with thunderbolt 4 and an eGPU dock with the P40. I consider this a nearly optimal setup for the connector-less GPU especially as my primary use case is as an accelerator. Hotplugging the eGPU still does not work so I need to reboot every time but other than that everything is fine.
The most annoying thing by far is however that unlike some of it's predecessors it seems that noone managed to overclock a Tesla P40. Because of the low clock speeds the performance in games is similar to the GTX 1070 and way below the GTX 1080 TI. And during the whole test it draws only slightly above 100W on average (the highest value I saw was ~125W) with the temperature maxing out at 61C. Now on the one hand I'm happy because the P40 still seems to be a cheap and pretty good solution but on the other hand I'm very much frustrated because I know this card could do so much more if only the clock speeds could be increased.
PS: I could never get the GPU utilization above 85% using benchmarks. Only stable diffusion succeeded in that. That's something I want to look into. An obvious idea is that the eGPU setup is the bottleneck here and steaming the rendering results back to the iGPU takes too much of thunderbolt 4's 40Gbps bandwidth. Another idea I had was whether something could be done with a Quadro P6000's (also based on the GP102 chip) VBIOS but clocked 200MHz faster...