For many years, stacking a computer full of 2, 3, or even 4 NVIDIA graphics cards was a viable option, and ideal for many GPU-compute applications like rendering, color correction, and machine learning. Prior to the RTX series of cards, the "reference" cooler designs from NVIDIA had one fan located near the front of the video card which pulled cool air in, pushed it across the heatsink, and exhausted that now-hot air out the back. Using multiple cards like that only required plenty of fresh air intake into the system, a little space between each card for the fan to breath, and a big enough power supply to keep them all running.
With the launch of the GeForce RTX cards, though, everything changed. This series debuted with dual-fan cooling solutions from NVIDIA, and while multi-fan cards had been available from OEMs in past generations they became the norm with RTX.
There are some advantages to this: a single video card can be kept cooler and quieter with two fans, for example… but there are also serious downsides. These coolers do not exhaust heat out the back of the card, instead of recirculating it within the computer's chassis.
With a single card, and a decent airflow through the case, that isn't usually a big deal – though it can be a challenge in small form factor systems, which tend to have restricted airflow. Once you move to two video cards, however, you've doubled the amount of heat being pumped back into the system – requiring even more is movement from chassis fans – and also introduced the opportunity for one GPU's fans to pull in already-hot air that just came off the other GPU. Hot air cannot absorb as much additional heat from the GPU as cooler air would, dramatically affecting the temperatures of the video card and leading to increased fan speeds, added noise, and potentially overheating. When video cards get too hot they throttle down their clock speeds to prevent damage, reducing performance.
With enough case fans and ventilation, as well as some space between the video cards, you might be able to get away with two multi-fan cards in a single system… but certainly not more than that. We've tried a full set of four RTX 2080 Founders Edition cards, and they overheat and throttle within minutes of being under load – with the impact felt before even completing a single run through OctaneBench:
How widespread is this problem? All of the GeForce RTX 20-series Founders Edition cards from NVIDIA share this dual-fan cooler design, and the recently-released Titan RTX does as well. None of those cards are suitable for multi-GPU use, which is a particular shame in the case of the Titan RTX since its 24GB of video memory make it very appealing (on paper) for GPU-based rendering. Only the Quadro line of "professional" video cards appears to be unaffected, with all announced Quadro RTX models sporting a single fan.
Why would NVIDIA do this? It depends on how you want to look at things. From a positive perspective, the dual fan cooler is going to be a better solution for almost all computers equipped with a single video card. It is both quieter and cooler, so long as there is decent airflow to help exhaust the extra heat inside the chassis. I would wager that a large majority of NVIDIA's GeForce customers use just one video card, so from that standpoint it is certainly not a bad way to go. The smaller number of users, often enthusiast gamers, who want two GeForce cards in SLI should be able to build (or buy) computers with additional cooling to handle the heat of two RTX cards.
On the flip side, this change dramatically hurts users who want three or four video cards for use cases like machine learning, GPU-based rendering, and color correction. They are prevented from using any of the Founders Edition cards in this generation, and most painfully cannot use the Titan RTX either. That seems odd since the Titans are traditionally marketed toward more demanding graphics applications. Why would NVIDIA want to cut them out? Well, for years NVIDIA has tried to encourage more "professional" users to purchase Quadro video cards rather than GeForce, or even Titan. Why? Quadro cards cost a lot more, meaning NVIDIA makes a lot more per card. They do also have some nice features: usually more VRAM at a given level of performance, for example, along with official certification for certain programs and ECC memory on higher-end models. There are definitely places where a Quadro card is a far better choice than a GeForce or Titan, but it seems like this cooling layout change may be a way for NVIDIA to push even more users toward cards with bigger profit margins – even if they don't need a Quadro from a compatibility or feature standpoint.
What software needs more than one GPU? Here is a chart showing some of the programs and use cases we have tested here at Puget Systems, and whether they do well with a single video card or are better off with two or more:
|Single GPU Applications / Use Cases||Multi GPU Applications / Use Cases|
|Adobe Creative Cloud (PS, LR, PP, AE)||Agisoft PhotoScan|
|Autodesk 3ds Max, Maya, Revit||DaVinci Resolve|
|Dassault Solidworks||OTOY OctaneRender|
|Maxon Cinema 4D||Redshift 3D|
|Pix4D Mapper||Chaos Group V-Ray (GPU version)|
|IrisVR Prospect||Machine Learning|
|Gaming / Game Dev / Virtual Reality||GPGPU Computing|
Are there ways to work around these cooling issues? Yes and no. In the case of GeForce RTX 20-series cards, OEMs like Asus, Gigabyte, and PNY have put out single-fan, blower-style cards. They may not look as fancy as the Founders Edition models, and you still need to be careful to ensure that there will be some space between each card for the intake fan to breathe, but these are an excellent option for those who want high performance with a relatively low price tag – and who aren't too limited by the 6, 8 or 11GB of video memory found on these cards. They do also tend to be louder under load, but at least they can maintain full performance.
Four Asus Turbo RTX 2080 Video Cards Under Full Load Without Throttling
Sadly, the Titan RTX is exclusively made by NVIDIA – so no single fan version is available. For users who want multiple GPUs with more than 11GB of VRAM using this generation of technology, the Quadro RTX cards are the only air-cooled option. The Quadro RTX 6000 should have similar performance to a Titan RTX, while equipped with the same 24GB of memory and other added features.
Another option that some might pursue is liquid-cooling, which comes in two varieties. Closed-loop or all-in-one liquid coolers are a single unit with the water block, pump, tubing, and radiator combined – but that is not viable here. Each video card needs its own cooler, and virtually no chassis has space to accommodate four separate radiators in a layout which would provide proper airflow to all of them.
Open-loop or custom liquid cooling is created with individual pumps, water blocks, reservoirs, tubing, and radiators – put together in whatever combination is required for the hardware that the builder wants to cool. This could be set up in such a way that four video cards were cooled, with one or two radiators handling all of them. For someone building a system on their own, with the right tools and expertise, this is a viable option. However, we moved away from offering this type of liquid cooling here at Puget Systems because of complications that arise in shipping and support. There are many potential points of failure in this type of system, and a leak can cause damage to components that aren't covered by manufacturer warranties. Shipping a fully built system across the country with this sort of cooling in place increases the likelihood of a failure, because of the stresses placed on hardware within a system during transit. Moreover, if any parts need to be swapped it is much more difficult to do so with liquid cooling in place. It may be right for some, but for a majority of users, the added cost and complexity of liquid cooling is just not worth it… especially when alternatives are available.
So what is the take away from all this? Graphics cards have different types of cooling available, and you need to be careful to select cards with the right layout for your situation. That means certain video cards are simply not going to be an option for some use cases, like the Titan RTX being unsuitable for multi-GPU configurations. Within today's RTX (Turing architecture) generation of cards, the best performance value is available with GeForce cards using dual fans for an individual card or single fan models for multi-GPU, while high RAM capacities are available with the Titan RTX if you need just a single card or the Quadro RTX series if you want multiple GPUs.
If you are looking for a new workstation and would like advice on this or any other hardware choices, check out our numerous articles and feel free to call or email our consultants for personalized guidance.