Data Centre Carbon Hackathon: Microsoft Azure (centralised) vs heata (distributed)

Feb 23, 2024

Data Centre Carbon Hackathon: Microsoft Azure (centralised) vs heata (distributed)

Feb 23, 2024
Chasing the low carbon energy:
Could heata beat Microsoft Azure on carbon, before factoring in heat re-use?

We’re focused on re-using heat, but we’ve often wondered whether the distributed nature of our network presents another opportunity to reduce the carbon footprint of our processing. Grid carbon intensity varies significantly by location and time, and with a geographically distributed network you could, in theory, place workloads where the low carbon grid energy is.

To give you an idea as a general rule the carbon intensity of the power in London is significantly higher than it is in Scotland, and there are variations across the UK. And at some times, in some locations, renewables will be delivering a greater percentage of the power – even an excess. (Perhaps all of the heata units should be in Scotland!)

A map of the UK showing the carbon intensity of different regions.

Carbon aware distribution is not an entirely new concept – for example the Green Software Foundation and Microsoft have developed tools for this very purpose. However, historically the focus has been on shifting workloads between centralised data centres, so we were keen to understand what sort of an impact a highly distributed system could bring to the table.

After presenting at a Lunch and Learn session at Microsoft, and fielding lots of interesting questions, an opportunity came up to work with Microsoft’s MAIDAP (Microsoft AI Development Acceleration Programme) team on an ‘AI for Greenops’ Hackathon. We thought this premise could be a great challenge for their bright minds, and we made it a bit more interesting by setting them the task of beating their own Azure data centres! 🙂

Who are the MAIDAP team?

MAIDAP is a programme at Microsoft for new graduates that provides the opportunity to collaborate with product groups across Microsoft to tackle some of the most exciting and challenging AI problems. Normally the team focuses on internal collaboration, but this was the team’s very first external collaboration, we felt very honoured!

This challenge was especially relevant to the MAIDAP team, as the goal and constraints led to a level of complexity that required AI / ML to resolve, and it was nice for them that the output was connected to the physical world, potentially leading to a tangible, physical carbon impact.

Could MAIDAP beat Microsoft Azure’s average carbon intensity in the UK simply by shifting workloads in space and time across the heata network?

The team were given five days to build a model that could automatically read current and forecasted grid carbon levels across thousands of heata nodes across the UK, and subsequently distribute workloads to the lowest-carbon options.

Key constraints:

  1. Moving data has an inherent carbon footprint: so the decision to move a workload incurs a carbon penalty that cannot exceed the carbon saving of moving it.
  2. As heata network provides heat to homes, we needed to maintain a minimum utilisation of >50% on each node in any given month (and all nodes must be used).
  3. Compute workloads needed to be delivered back to the customer within a reasonable timeframe; longer delivery times can impact service levels and that impact needed to be balanced against the carbon saving.


The MAIDAP team developed a solution based around three steps:

  1. 24 hourly submission of job list and metadata
    For the Hackathon, this meant synthesising data for the workloads based on job parameters like size (GB) and expected time to complete; the simulated jobs were randomised broadly following a daily and a weekly cycle. (for which the team used a non-homogeneous Poisson Process.)
  2. Optimisation of workload location
    To achieve this, the MAIDAP team applied Mixed-Integer Linear Programming (which took a lot of maths!) to find the optimal nodes based on the carbon intensity of the power supply of the households.
  3. Allocation of workloads to optimal nodes

Parameters and assumptions:

Number of servers:
10,000 servers, in heata’s case these were randomly distributed throughout the UK, in Microsoft’s case they were located in their UK data centres in London and Cardiff.

Grid power carbon intensity data:
Some of the most server populus regions also have high carbon intensity.
Different regions have different patterns of carbon intensity variation over time.
Grid carbon intensity data was obtained from the Carbon Intensity API.

Data transmission carbon penalty:
100g Carbon / GB.
This is not a very well understood area, and in reality depends on the locations the data passes through too. 100g is likely too high, but we had to give Azure a chance!

Synthetic Test Data Generation assumptions:

Arrival rate of the workloads changes over time:
Weekly trends: weekdays vs weekends
Daily trends: day vs night

Workload properties:
Data size in GB
Completion time in hours

Carbon intensity fluctuations by region
The team used the Carbon Intensity API
Did they manage to beat Microsoft Azure?

Of course they did!

Key outtake
Optimising data processing by location and time has the potential to reduce carbon emissions by as much as 42%, accounting for the impact of any additional transmission of data.
This is before you factor in the advantages of heat re-use and the associated reduction in cooling energy.
Carbon Hackathon: Gold medal for heata and silver for Microsoft Azure


1. heata vs UK Microsoft Azure data centres

Using the optimiser to run the workloads on the heata network, led to a 42% reduction in carbon emissions*

This demonstrates the significance of local grid carbon in any data centre’s carbon impact.

But my data centre uses 100% renewable energy?
This is a claim made possible thanks to Renewable Energy Certificates and Power Purchase Agreements. Referred to as ‘Market-based Reporting’, companies purchase renewable energy, though this renewable energy will most likely be completely unrelated to the power they are actually using at their location.

2. heata (today) vs heata (optimised distribution)

The optimised distribution also demonstrated a potential reduction in carbon emissions of 20% compared to the way heata distributes its workloads today**

A chart showing the relative carbon intensity of centralised vs distributed compute workloads when optimised in space and time.
* Based on the synthesised workload dataset spanning 48 hrs.
** Mostly-random assignment. Heata sends workloads to the cylinders where heat is required but that variable is independent of location, so as far as location is concerned assignment is random.
3 . Redistributing ongoing tasks

Another upside was also identified; ongoing jobs could be redistributed to new nodes, if after 24 hours the optimiser deemed it appropriate to do so (this took into account the carbon footprint of moving the workload). 

24 hours after the redistributed job was moved, the team was able to show that redistributing the job could reduce emissions by 28.5% vs the job that stayed in place.

Improvement before any heat re-use is taken into account

It’s worth highlighting that this improvement is based solely on the carbon impact of the power used for compute. The heata network also brings in additional benefits: reduced power overhead for cooling, and significant energy re-use (~66%) which offsets a household energy cost.

“It was great to work with such enthusiastic, bright people, and also great to see that an internal Microsoft team demonstrated that we could be substantially more sustainable than Azure!”
Chris Jordan, Co-founder, heata.

More sustainable compute infrastructure

Demand for compute is growing rapidly and putting significant stress on our power networks, reducing the energy load is a key goal as we build the infrastructure to meet this demand. As a general rule, using less power helps to reduce the carbon impact as it reduces the amount of power that needs to come from less clean sources.

Therefore improving the efficiency of existing data centres is key and Microsoft has made significant progress in improving the energy consumption of its data centres to optimise PUE (Power Usage Effectiveness) and maximise hardware utility.

But location also plays a significant role in the carbon intensity of any compute, and the results of this hackathon show that we should think more broadly about tackling this problem. Distributed infrastructure enables optimisation for carbon, using cleaner power to achieve the same end result.

What if?

Armed with this knowledge, it begs the question – if you can define the attributes of workloads suitable for distribution and then determine the volume of these workloads within a typical data centre (another Hackathon challenge!?)… why wouldn’t you choose to build new infrastructure in this way?

And this is before you factor in the additional social benefit of re-using the heat to deliver free hot water to families (offsetting a household heating energy cost), and the associated reduction in cooling energy (again reducing the overall power demand).

And before you factor in re-using physical infrastructure; the house, the cylinder and water, and power and connectivity are all already in place – which means you can avoid the incredible embodied carbon costs of building a brand new data centre, and its associated strains on local utilities.

Together it makes a powerful argument for thinking differently about compute infrastructure, especially as connectivity continues to improve.

Intelligent distribution

Getting back to the output of the Hackathon, the results exceeded our expectations and we’re looking forward to seeing how we can start to build this intelligence into the distribution of our workloads. And imagine then layering in fuel poverty criteria, whereby you also try to prioritise supporting those households who need the most help…

It’s exciting to imagine the potential of intelligent software working in tandem with innovative physical infrastructure to make a big difference to environmental and social goals.

“Microsoft’s mission is to empower every person and every organization on the planet to achieve more. When the opportunity to work with Heata came along, the MAIDAP team was thrilled to be able to have a positive social impact and use their AI expertise for Good.”

Big thanks

To Aarti, Chester, Derenik, Juhi, Natalie, Nick, Omisa and Soundar for engaging heata in the Hackathon. We were hugely impressed by the capabilities of the team and their enthusiasm for heata’s mission and vision.

Follow us on Linkedin to keep up with other heata updates, insights, and exciting developments in our journey towards a more sustainable future!

#Heata #GreenCloud #AIforSustainability #FinOps #GreenOps #DistributedComputing

If you’re an AI expert, or an advocate for sustainable infrastructure, we’d love to get your input: mail Charlie at

Continue reading