As promised I have now made up couple of extra renders showing the internals of my GPU waterblock concept, plus a little explanation of the physics behind it for those that are interested in such things. I’m working on the other less informative but prettier renders showing the full SLI assembly, which will be posted for your amusement as soon as I find time to finish them.
This is what the core of the GPU waterblock concept looks like when you strip away the facings, fixing/framing plate, the Plexiglass internal structure, the mounting screws and the graphics card itself - leaving just the waterblock
, rubber gasket
, Bitspower fittings
, and the Aqua Computer MPS flow sensor
The original idea was created with the Titan Black
in mind, whilst this particular CAD model was made (or at least started to be made) for the Titan X
. The basic rational behind the design was that since these are dual-width cards (and the DVI connectors prevent them being otherwise) there is no benefit to having a conventional thin waterblock - so I might as well make use of the space to try and build something more effective (and nice to look at). As such this is a deep-channel
design, in a similar vein to some of the motherboard-waterblock concepts shown earlier in the build log. It is machined from a big slab of copper (20mm thick), sealed with a full-area rubber gasket and capped by a Perspex window. This makes use of a costly amount of materials (far more than any retail product could afford to) but should have several advantages:
- The larger volume of metal means there is a higher thermal capacitance in direct conductive contact with the chip – i.e. a bigger mass of material requires more energy to raise its temperature by the same amount. In theory this should help to smooth out the effects of short-duration spikes in GPU loading (from say the ‘update model’ command in CAD) and make the temperature changes more gradual.
- The thicker block is able to accommodate standard G1/4 threaded ports directly into the edge of the copper slab. This means that, unlike normal designs, the fluid does not have to execute an instant 90-degree turn (or worse, in a parallel setup, an instant 90-degree split) in order to enter the block. With most retail waterblocks the flow geometry must also change at virtually the same time; from the circular port to a longer, narrower slit. Both of these things disrupt the flow and contribute to the formation of resistance effects in a way which, alllied to the smooth 45-degree flow-split of the manifold design, is mitigated by this approach.
- In a similar vein, the normal internal dimeter of G1/4 fittings is about 10mm. The deep-channel design allows for a flow area of 10x10mm to be maintained throughout the block, save at the critical point where it passes over the GPU. Thus, whilst the flow vector alters (as smoothly as possible), changes to the prevailing cross-sectional flow-area are minimised as much as possible. Additionally the thick block design means that heat can be effectively transferred into the coolant water from the sides of the channel as well as the base; yielding three square cm of thermal contact area for every cm the flow progresses.
This image illustrates the passage of coolant water through the block. The channel has been shaped to pass directly over all the main heat sources as far as is practically possible. Whilst this produces a more complex path than normal, changes in direction are kept as smooth as possible and should not be as problematic using a large-volume channel than with a conventional shallower one.
The incoming flow is split between the four cards by the manifold and then enters the block as shown via a mini D-plug fitted into the inlet port. It then passes only briefly through a short length of channel before exiting through another G1/4 port. From there it passes through an assembly of fittings, through the integrated MPS flow sensor, and back into the block.Here there are a number of issues with the design which I never managed to resolve before discontinuing work on it. Firstly the MPS is sited on the incoming flow, which means its integrated temperature sensor is reading the inlet water temperature which is of no use at all. There is nothing in the block design which would stop me running the flow the other way, but using the 900D case causes some limitations with the rest of the Ironbeast loop layout which prevent me doing so. Secondly, my experience with the radiator testing shows that the MPS sensor, being differential-pressure based, will misread if the flow comes directly of a turn in this way. Finally, but most problematically, that particular assembly of fittings would prove extremely tricky to install.
Upon returning to the block the flow channel turns upwards over the VRMs before splitting into two separate paths (this geometry would probably be smoothed out a bit in a more refined design). One path goes directly to the GPU area, whilst the other slightly longer path curves down to cover some of the memory and then enters the same area from another direction.
The main GPU heat transfer area, rather than using the normal linear channel setup, is modelled on the pin design of the Aqua Computer Cuplex Kryos
CPU waterblock (though an order of magnitude cruder in scale). This is basically two sets of channels cut across each other, producing a very high amount of surface area - and with the incoming flow streams intersecting at right angles, a region of turbulent flow is produced directly over the chip. Furthermore, even with the presence of the pins, the flow area available here is larger than that of the main channels. This increased area should mean that the local flow velocity
is temporarily reduced (because in an incompressible fluid the overall flow rate must remain a constant), meaning that each molecule of water spends a proportionally longer time directly over the GPU than it otherwise would. In combination these three factors should, in theory, produce very efficient rates of heat transfer.
The two exit paths from the GPU area mirror the entry paths, again with a longer branch to pass over the memory directly. In both cases the difference in path length and geometry will produce a slight bias in favour of the shorter route, but at this scale and pressure it shouldn’t be significant enough to matter. The two exit paths reintegrate into a single channel, which passes back out of the block through another mini D-plug. At this point the second manifold smoothly recombines the flows from the four graphics cards back into a single stream and connects to the next part of the loop.Edited by OCDesign - 10/28/15 at 5:52pm