Improvised Hybrid Cooling for Nvidia RTX A6000

Dumi Loghin
5 min readNov 30, 2021

In this article, I will show you how I adapted a CPU All-In-One (AIO) cooler to the state-of-the-art Nvidia RTX A6000 GPU card. This video card costs more than $5,000, so be careful when you try the steps shown in this story. If you do not have experience building a PC, I advise you not to try these steps.

Fig. 1 — Top view of the RTX A6000

Motivation

I run some intensive machine learning experiments, and I observed the GPU was hitting 85°C, which is quite high. At that time, there was no custom water block for RTX A6000, so I decided to improvise and use a CPU AIO to cool the GPU. I am using my older Cooler Master MasterLiquid Lite 120 AIO, as shown in Fig. 2.

Fig. 2 — Cooler Master MasterLiquid Lite 120 AIO

I will also use a Kraken G12 mounting kit as shown in Fig. 3. Note that this is not compatible with either RTX A6000 (in fact, there is no mounting kit for this GPU as of November 2021), or the AIO. I had to drill new holes for mounting the AIO on the kit, and also for mounting the kit on top of the PCB.

Fig. 3 — Kraken G12 GPU mounting kit

Disassembly

There is a very good guide on how to disassemble the RTX A6000 card here. I won’t repeat the steps. I will just show you some photos of the disassembled GPU. Again, be careful when you handle the PCB, when you clean the die, and when you remove the thermal pads.

Fig. 4 — RTX A6000 PCB with and without the radiator

Preparing the Mounting Kit

The Kraken G12 mounting kit does not fit the RTX A6000 GPU. I had to do the following modifications: drill 4 holes for securing the AIO on the kit (highlighted with violet in Fig. 5), drill 4 holes to match those on the PCB to mount the kit (highlighted with red in Fig. 5, matching the PCB holes highlighted with red in Fig. 4), remove the extra metal pins (highlighted with yellow in Fig. 5).

Fig. 5 — Preparing the mounting kit

Next, I covered the fan hole of the mounting kit with a thin wood plank, and I mounted the AIO on the kit.

Preparing the GPU

To mount the kit on the GPU, I had to add 4 hexagonal standoffs, 1cm high, on the places highlighted with red in Fig. 4. On the back, I re-used the “N” brackets from the Kraken G12 kit, but I oriented them inwards, as shown in Fig. 6.

Next, I added heat sinks to the memory and voltage regulator modules (VRM), as shown in Fig. 6. Note that the space between the AIO pump and the PCB is very narrow, so I had to use thin (5mm) copper heat sinks in that area. I had to cut those heat sinks myself to fit the space.

Fig. 6 — The GPU with custom heat sinks for the memory and the voltage regulator modules

Adding a Blower

The memories and VRMs can get quite hot, so I needed a way to cool them down. My idea was to use a blower fan (see Fig. 7) and to orient the heat sinks such that the air flows through them. The problem is with the memories and VRMs that are obstructed by the AIO. As shown in Fig. 10, their temperature is significantly higher compared to the other modules.

Fig. 7 — The mounting kit with the AIO and the blower

Another disadvantage of the blower is its noise.

Fig. 8— Cooling solution assembled (top view)
Fig. 9 — Cooling solution assembled (back view)

Results

Let us see the results now. I used a WH5000A multimeter with a temperature probe to get the temperature of different hot regions of the GPU. For the GPU die, I used nvidia-smi tool. Indeed, the GPU die’s temperature is around 60°C with this AIO. There is a 25°C (or 30%) drop in temperature. The memories and VRMs are also relatively cooler, except for those where the airflow is not so good, as shown in Fig. 10. The minimum temperature is 42°C where the air enters, and the maximum is 77°C on the VRM that is not well exposed to the airflow.

Fig. 10— Measured temperatures

In summary, there are significant improvements in cooling with this improvised solution. The issues are (1) some components are not cooled well due to insufficient airflow and (2) the blower is very noisy. In a future article, we will explore water cooling on this impressive GPU.

--

--

Dumi Loghin

I am a Research Fellow in Computer Science with experience in parallel and distributed systems, blockchain, and performance evaluation.