The ♨ case for Open Compute Heat Appliances
Towards truly sustainable data centers
Driven by climate concerns and the ongoing war in Ukraine, energy usage and the transition to a more sustainable and less dependent economy continue to be a hot topic. In Europe, political consensus has been reached to reduce the dependence on fossil fuel sources in general and Russian gas and oil in particular, with many countries committing to ambitious emission reduction targets (by 2030 and 2050). Policies and regulations to move things in this general direction are being drafted and put into law — but the question remains how these goals can be achieved concretely and specifically, from a realistic technical perspective.
The diagram above illustrates energy usage in The Netherlands for 2017. While some numbers have changed by now (2022), the general trends and flows remain similar. As indicated, the diagram shows how waste heat energy from industrial processes (including data centers) could be reused for domestic and commercial space and heating applications. In case of The Netherlands, a full transition could save approximately 33% of natural gas usage and associated emissions (420 PJ in 2017, 24 megatons(Mt) of CO2e at 56.6 kg/GJ CO2e ).
Heat source: Data centers (with a ♻️ twist)
Obviously, existing industrial sources of rejected heat are not currently deployed in ways that facilitate heat reuse. The numbers in the diagram show an ideal scenario of achievable gains, an order of magnitude and a mental picture to support the discussion.
The data center industry in The Netherlands only used 9.9 PJ of energy in 2017 (2.7+1e6 MWh, 308 MW installed capacity assuming 24*365 full utilization). So even if all data centers were upgraded to fully reuse their generated heat, only ~3% of natural gas used for residential heating could be replaced.
Data centers as sources of recycled heat producing compute elements
Data center operators are focused on efficiency, and typically replace their servers every 3–4 years. After this ‘primary usage’ the hardware components — think CPUs and GPUs — remain functional for many more years. In particular, they can continue to generate heat. By repurposing this hardware in heat pump appliances, additional gas used for residential heating can be replaced.
For sake of argument, let’s assume each refurbished heat appliance is a 4kW appliance filled with recycled compute hardware. 308 MW of capacity represents 308000/4 = 77000 heat pumps; a refresh cycle of every 4 years means 19250 electric heat pumps could be produced every year. These could contribute to the plan to install 100.000 hybrid heat pumps per year starting in 2024; additional refurbished parts could be imported as needed.
In a final state of (say) 1 million households equipped (out of 8 million total) with such electric heat appliances, assuming 1300 m3 of annual natural gas usage per household is replaced this represents 48 PJ of heating energy, or ~17% of natural gas usage in Dutch households.
Design requirements for Open Compute heat appliances
The transition from natural gas to reused heat from electric appliances implies specific design requirements. For example, the capacity of the power grid is limited and cannot simply replace gas lines. The charging of the heat appliances will have to be distributed over time in a coordinated fashion, which implies a need for storage and connectivity.
Note that many modern CPUs and GPUs can operate at higher temperatures than previous generations (most Intel and AMD CPUs can operate reliably at 95℃ for years, some GPUs reach junction temperatures of 110℃ ). Moreover, they have active thermal protection mechanisms that automatically throttle back their performance and heat dissipation in case temperatures outside supported ranges are reached.
As could be expected, the need for space heating is high when the sun and temperature outside are low. A heat appliance may be required to store up to 12 hours of heat, generated in the 12 hours before the actual need. For a 4kW appliance, this implies a minimum of 4x12=48 kWh of storage. For salt-based storage using K2CO3, at 0.84 GJ/m3 energy density this corresponds to 0.2 m3 (200 liters, see this link (in Dutch) for a comparable unit). Note that similar storage requirements would apply for a heat pump powered by local solar panels, independent from the grid.
The requirement for coordinated charging implies a need for network connectivity. Open Compute Heat Appliances are connected devices (also known as “IoT appliances” or “Smart Grid”). To a grid operator, 1 million devices with a total of 4GW installed capacity would represent an interesting option to offload surplus renewable sources from wind and solar, to balance demand with supply.
It is assumed that applications can be found to make use of the compute capacity provided by these open compute devices. The distributed, sustainable cloud could be leveraged for various purposes; existing cloud providers and innovative startups will have ample opportunity to put these resources to good use.
The Caldera Warmstone heat battery
For a concrete example of what this device might look like, consider the Caldera Warmstone heat battery
The device is connected to the Internet in collaboration with myenergi. It takes 5 hours to charge (using a full 100A circuit), and the core temperature reaches 500℃ when fully charged. The FAQ suggests the energy density is equivalent to 14 hot water tanks of 150 liter each, and the unit is positioned as a “heat battery” (as opposed to a “heat pump”) aimed at replacing existing oil heaters with regular room radiators.
The Sunamp Thermino hot water heater
The Thermino uses solid state phase change materials to store an amount of energy equivalent to a hot water boiler of 4x its volume. With an inlet temperature range of 65–80℃, this device is ideally suited to take in heat coming a CPU or GPU cluster.
A community driven modular systems approach
Examples of companies designing and selling heat appliances exist, even electrical ones featuring compute heat reuse. However, enabling cross industry use cases such as balancing the grid, and reusing compute elements from data centers at scale, require broad coordination and open standards. Dimensioning the storage capacity for a household yields different results when accounting for hot water usage only (x number of showers at 80–90 liter each, a hot bath at 150 liter) versus considering the needs to balance the grid, and to decouple heat generation from heat usage. Novel technologies such as salt-based storage or thermal acoustics may have specific requirements (like temperature ranges and materials used). The Open Compute industry forum is well positioned to drive these initiatives, to enable a broad and effective ecosystem with participation from various specialists. A uniform design will simplify installation and training of technicians.
The case for individual hyper scalers
Facebook/Meta recently tried to deploy a 200 MW data center in Zeewolde (near Amsterdam), but the project continues to face opposition. Imagine if Meta would come to the table and say: In addition to the 200 MW of annual direct heat, we commit to provide 40 MW worth of compute elements every year, in line with our 4-year replacement cycle. This hardware could be repurposed in the form of 10.000 4kW heat pumps, to electrify heating and move away from natural gas for an equivalent amount of households — every year.
This would address community concerns around affordability and availability of subsidized heat pump hardware; the well-tested supply chain would be very robust and predictable. Moreover, Meta (and other hyper scalers) could work with grid operators to optimize the usage of the resulting sustainable cloud, to maximize the economic benefits for residents and enable innovative monetization models at scale. It might be possible to provide electricity used to provide heat free of charge.
The refurbished compute elements would reduce the need for more hyper-scale facilities; they would effectively be replaced by distributed, small scale compute elements. And of course the same approach could be applied everywhere in Europe, perhaps even globally.