Harvesting thermoelectric energy from computer graphics cards

Question

During CES2015 a well know Graphic Processor Unit (GPU) manufacturer introduced a new GPU. These GPUs need complex thermal management to keep the processor cool. Most GPU accelerator card manufacturers develop new graphics accelerator cards that use primitive thermal management technology compared to today's advanced technologies. Most of us know this technology as fans that manage this unwanted thermal energy, as seen in this image of a graphics accelerator card:

What barriers do engineers need to overcome to convert this wasteful thermal energy to useful electrical energy?

Below is a temperature profile of a GPU card.

---

References:

• A Thermoelectric Generator That Runs on Exhaust Fumes
• Small Thermoelectric Generators
• Thermoelectric Technology Overview
• Extend Your Video Card's Life With Custom Auto Fan Speed Settings

Answer 1

There is heat that can be recaptured, but you won't get much of it away. As one of the commenters has mentioned, your absolute maximum is the Carnot efficiency.

$$\eta_{Carnot}=1-\frac{T_c}{T_h}$$

This is an idealized condition, you'll never reach this efficiency. But to find our limit, let's figure it out anyway. $T_c$ will just be room temperature, it might be slightly warmer inside the tower, but we'll give ourselves the benefit of the doubt and pick a nice round number at 20C (293K). $T_h$ will vary as the GPU works harder (this is one of the issues with this design in general; the power you get from the cooling system won't be consistent because GPU temperature varies depending on how much you're stressing the chip.) We don't want to run it too hot and damage the card, that defeats the purpose of a cooling system.

After some quick searching (Google "GPU operating temperatures," you'll see a bunch of forum posts that give a lot of different numbers, none of which I think are strong enough to cite, but I'm pooling their data to make my own assumption) it looks like most cards have a strong upper limit of ~100C before you start to do serious damage. However, running that hot will still reduce the life of your card, and judging by the picture in the question, this is a nice card for which we've paid a pretty penny, and we want to keep it around as long as we can. 70C is a good place to shoot for, but 80C (353K) is still probably pretty safe, and we want our best possible case. With those numbers, we get

$$\eta_{Carnot}=1-\frac{293K}{353K}=0.17$$

This means that, at the very maximum, the best we can do is to get 17% of the heat we're generating on the card back as electricity to power something in the tower. We can vary the card temperature, and as it goes between 60C and 100C, efficiency goes between 12% and 21%. Regardless, we're not getting a lot back.

That's the max efficiency though. This site, which sells thermoelectric generators, says that the top of the line TEGs will run at 8% efficiency. While this is better than the nothing that we'd have been getting before, the real issue here is cost and implementation. TEGs are not cheap, and cooling fans are. A basic cooling system is also much easier to install. Even if we can hook up a TEG to cool the card, we have to find something we can do with that electricity, and we don't want the variable power to be used for an critical components. Tower lights and extra fans are probably the extent of our usage.

So to answer your actual question in there, I'm sure we can find all kinds of creative ways to get that heat converted into electrical or mechanical work. Making it "useful" is an entirely different story.

Answer 2

Trevor Archibald has given you a really good answer, but I see from your comments a diffferent answer might be useful, as you still think this could be viable with the right economics.

It wouldn't be. The issue is engineering, not economics. It's a bad idea from the economic perspective, certainly; but changing prices won't make it a good idea. It would still be a bad idea. Let me explain.

low-grade heat

Low-grade heat is heat that is a few Kelvin or tens of Kelvin above room temperature.

getting rid of the heat quickly is the name of the game

George Herold points you in a comment to one reason why energy harvesting on the card would be a bad idea: the thermal conductivity of the card is designed to be high.

Getting rid of heat quickly is particularly important in IT equipment, where the electrical efficiency of the equipment is really really staggeringly poor. And that means that of the electricity you put in, almost all of it is going to get turned straight into heat. There is a theoretical minimum amount of energy needed to flip a bit, regardless of the medium on which the bit is stored. All the rest of the energy put in above that minimum, is going to turn into heat straight away. In order to protect the equipment, you need to get rid of that heat as fast as possible.

So the card is designed to get rid of heat as fast as possible. Anything you put in the way, such as your proposed energy harvesting device, will slow the rate at which heat leaves the card. That will raise the card's equilibrium temperature. And that will radically shorten the life of the card. That will happen regardless of the price of electricity.

it's not about the price of electricity

And this idea that if the price of electricity were high enough, it would make harvesting low-grade heat worthwhile, is simply wrong. If electricity is that valuable, then it's worth making the card more efficient in the first place, so that there was less waste heat: first, reduce the consumption of high-value energy, before trying to recycle low-value energy. And that brings me to ...

energy versus exergy

Heat is, in a large proportion of cases, a waste product. It's almost always the least useful form of energy. That's really what the Carnot efficiency limit is telling you: that to get any work out of low grade heat, you can only do so with very low efficiency; that is, almost all of the heat will stay as heat.

When doing engineering with heat and other forms of energy, it's very useful to build up an intuition to distinguish between energy (the thing measured in joules) and exergy (the thing that gets work done). The form that energy is in, determines how much work it can do. Electricity can do huge amounts of work efficiently - it has very high exergy. Low-grade heat can do very little work - it has very low exergy.

Once you've created low-grade heat, you're already at the end of the line for exergy (useful energy). Almost all uses of energy end up at low-grade heat. It's the final form for pretty much every chain of energy conversions. And, at the cosmic scale, it is (as far as we can tell) the final form for every single joule, in the heat death of the universe.

Low-grade heat is the end of the road. If you want more work out of those joules, then get that work done before those joules are in the form of low-grade heat.

Answer 3

Not an answer per se. But worth a revisit.

The new generation of GPU cards can run at very high temperatures, though whilst gaming the temps vary, another area that displays constant temperatures is whilst crypto mining, where cards can run between 80c to 90c, since the load on the card is always 100% the temperatures stay consistent, depending on the clock settings set by the user.

With the previous answer, of an upper limit being 21% an educated guess would be around a consistent 30%

My interest would be using a TEG instead of any type of fan cooling. Which will inevitebly take core temps higher, however, because GDDR6 dissipates heat through the bottom of the card rather than the top where the fans are mounted, during mining, they are all but useless.