Carbon cost of computing on ARCHER2

With the new ARCHER2 national high performance computing facility in the UK, I thought it was time to update my old https://jarvist.github.io/post/2017-06-18-carbon-cost-of-high-performance-computing/ estimate of carbon cost.

ARCHER2 doesn’t seem to state it’s power consumption anywhere, but a smaller 128’000 core CRAY EX based on the same architecture is listed on the Top500 as 637.95 kW. This makes ARCHER2 about 3.7 MW for its 750’080 cores. Perhaps this is why the power consumption isn’t noted anywhere anywhere - that is a pretty power hungry beast! Though this is quite a modest 5 W per EPYC core, or 640 W per compute node. Both these figures seem sensible.

(ARCHER was 1.2 MW by comparison, and I calculated each 24-core Intel compute nodes used 250 W.)

The new ComputeUnit (CU) metric for job costing is simply a node hour. So one CU is 0.64 kWh of electricity, or enough to make 25 cups (250 ml) of Tea.

Our 6 month time allocation as a relatively junior computational group (via the MCC consortium) was 5000 CU, which consumed 3.2 MWh of electricity. (Another way to think of our allocation is that it’s almost exactly 1 node for the full 6 months, a 1/5860 share of the whole machine!) The UK carbon intensity has been falling (thanks to all those new renewables), but is still around 0.233 kg of CO2e per kWh of electricity. So our compute time was responsible for about 750 kg of CO2e emissions, roughly equivalent to a round trip flight to Boston to present the data at the Fall MRS conference.

I need to patch the checkScript command with the extra 1 CU = 0.15 kg CO2e metric!

Any improved estimates or corrections gratefully received; this blog post is very much back of the envelope.

Jarvist Moore Frost