First off this data is preliminary and does not yet include multiple runs on each block. We are also waiting on XSPC, Bitspower, Watercool, Swiftech and Alphacool to finish/deliver their blocks. There are three facets to our performance roundup. The first and most important is cooling of the GPU core itself. The main GPU core chip is what kicks out the most heat on the board and is where the cooling engine design work is concentrated. These blocks are tested at various pump speeds to show performance vs flow rate.
Testing is performed with a “normal” overclock. +100mV Core, +100mV Aux are run on a reference R9-290, while the core is clocked at 1175MHz and the memory at 1465MHz. This provides a sizeable heatload as will be evidenced by core temperatures being significantly higher at high flow than was the case with the Titan/780 waterblock testing. GPU temperatures are logged by GPU-Z, Dallas one wire temperature probes are used to measure coolant and ambient temperatures and are logged by wintest through a CrystalFontz CF-635. Loading is provided by furmark and a 40 minute warm up is allowed before 20 minutes of data logging is initiated. The warm up occurs before each data point is data logged. MX2 TIM is used and is allowed to burn in over night in order to remove any time-related “curing” effects. The motherboard is a Maximux VI Impact with a 4770K clocked at 4.4GHz and cooled by an EK Supremacy water block. Cooling is provided by an XSPC EX 560 radiator with 8 low speed yate loon fans. Flow rates are measured with a King Instruments rotameter. Multiple high flow QD4/VL4N quick disconnect fittings are used to quickly replace blocks. Here is how the testbench looks without the GPU:
Core Cooling Performance
As expected cooling performance increases as flow increases. There may be a point at which high flow may hurt performance but this is not shown to be within the normal expected flow rates that users normally run.
While the shape of the data is therefor expected there are some points to note. The data is extremely close and there is little difference in performance between the three blocks. In addition in the Titan/780 waterblock testing we saw the EK block to have significantly worse performance at lower flow rates. This has been improved. In addition AquaComputer’s block was significantly higher in restriction and therefore operated at lower flow rates for a given pump speed. This also has been fixed as evidenced by the AquaComputer block having similar flow rates to the EK and Koolance. Flow restriction numbers will be available soon.
VRM Cooling Performance
In our Titan/780 water block testing we measured VRM temperatures by measuring the temperature on the rear of the circuit board under the VRMs. Although this gave an indication of VRM temperature, it really measured the PCB temperature which was greatly affected by how well the neighbouring chokes were cooled. However AMD included VRM temperature probes that can be monitored by GPU-Z. There are two VRM probes. One is located close to the memory on the DVI/HDMI/DP end of the card, while the other is located in the main VRM section near where the 12V power cables will be connected. The first never gets particularly hot. In fact in one case the thermal pads were forgotten that would enable the VRMs to be cooled. The VRM only hit 20C over coolant which is very good. The other VRM temperature sensor on the other hand gets significantly hotter as can be seen in the following plot:
Taking the first three lines that are grouped together we see that VRM performance is spread a little more than core temperature was. Koolance uses the thickest thermal pads (1mm thick) and unsurprisingly has the worst performance. EK decided to use thicker pads for the VRMs than on the 780/Titan block and therefore loses ground to AquaComputer who has the thinnest pads so far. In addition the AquaComputer backplate provides a significant boost in VRM cooling. This was surprising and initially attributed to the “active” heatpipe. My expectation was that the active backplate was somewhat gimicky – a cool feature but one that might not prove to actually affect VRM temperatures much at all. Removing the heatpipe and running with just the “passive” part of the backplate shows this to match expectations. However the large effect of the backplate is mysterious right now. There are two theories – one that it’s providing extra pressure to the block to cool the VRMs, and the second is that there are thermal pads on the backplate that cover more area than the EK for example.
Flow Rate & Restriction
Like the CPU Block roundup Koolance again proves it’s supremacy in core cooling, but drops the ball with it’s thick thermal pads. However it’s lead is not really enough. 1C is a small gap, however AquaComputer shows that it’s backplate can really make a surprising difference in VRM cooling. As the roundup is not complete then awards can not yet be issued but it would seem that AquaComputer is making a very strong case with it’s all round performance! XSPC’s block is on the bench right now!