The best machine to run PowerFactory

Day 33 / 60

The best machine to run PowerFactory is the one with the highest clock frequency possible.

I'd been avoiding this question for weeks: which machine should run PowerFactory?

I'll be honest: I assumed more cores and more RAM meant faster. I used the one recommended by default, a c2d-standard-16 with 16 cores and 64 GB of RAM. Never measured it.

So I did what I should have done weeks ago: a systematic benchmark. 5 machines, 25 power flows each, multiple runs. Same CEN database (~2,600 buses from the Chilean grid), same PowerFactory version, supported by Don Nelson.

The machines

All running PowerFactory 2024 SP1 on GCP, with Spark 0.3.1 executing the flows. I didn't get to collect full data for the n2-standard-4, but it's the slowest by far — the results are from the other 4. I tested c4-highmem-2 after publishing the post, details are in the update at the end.

Results

The chart speaks for itself:

The c4-highcpu-4 with 4 cores and 8 GB of RAM averages 2.19 seconds per flow. The c2d-standard-16 — the one I was using — with 16 cores and 64 GB averages 2.72. Four times the resources, 20% slower.

Why? PowerFactory is single-threaded for power flow solving. The only thing that matters is core frequency. More RAM doesn't help. More cores don't help. The 15 extra cores I was paying for did absolutely nothing.

Another data point confirming the thesis: the c2-standard-8 (8 cores) is only 3% faster than the c2-standard-4 (4 cores). Same frequency, double the cores, virtually identical performance. The extra cores are useless here.

Thermal throttling

The c2-standard-4 has a problem: it overheats. On the second run, the first 13 flows average 3.7 seconds, but from flow 14 onward they jump to 4.6 seconds. The CPU drops frequency to prevent overheating.

The red line is the c2-standard-4: you can clearly see the jump after flow 13. The c4-highcpu-4 (green) and c2d-standard-16 (blue) stay flat. The c4 isn't just faster — it's consistent.

This matters. When you run a full study with multiple scenarios, you need each flow to take the same time. If the machine throttles halfway through, your timing becomes unpredictable and your cost estimates are useless.

The other times

The power flow isn't the only thing that takes time. Before solving, Spark has to load the database and activate the scenario. These times also vary between machines.

Load time

Loading the CEN database (~2,600 buses) is pure I/O: reading the file from disk and parsing it into memory.

The c4-highcpu-4 with Hyperdisk loads in 8.3 seconds, 33% faster than the c2 machines with pd-ssd (~12s). Disk type matters here.

An interesting detail: in some runs the load time spiked to 80-130 seconds. That happens when the disk cache is cold — the first time the VM reads the file. In consecutive runs the OS keeps it cached and it drops to normal times.

Activation time

Activating the scenario (study case + operation variation) tells PowerFactory "I want to simulate this system state."

Same thing here: the c4 activates in ~1.95 seconds vs ~2.7s on the c2-standard-4. The absolute difference is small but it shows that CPU frequency impacts every operation.

The c2d-standard-16 had an outlier of 10.2 seconds in one run. No clear explanation — possibly contention on the shared VM.

Full pipeline

Putting it all together — load + activation + solver — shows the end-to-end difference:

The c4-highcpu-4 completes all 25 flows in 64 seconds end-to-end. The c2-standard-4 takes 107. The c4 wins at every stage.

Cost

In practice I keep the machines running by the hour. What matters is how much each dollar buys me.

The c4-highcpu-4 costs $0.17/hr and does ~1,642 flows per hour. The c2d-standard-16 I was using costs $0.73/hr and does ~1,322. I was paying 4.3x more per hour for 20% less throughput.

I was burning money. Literally.

Infinite compute

In a job application to OpenAI they asked: if you had infinite compute, what would you do? At the time I didn't have a good answer. Now I understand.

I would simulate the electrical grid in all its infinite possibilities to learn from it. Contingencies, protection settings, operation scenarios, demand variations. Every possible combination. A machine running 24/7 generating knowledge.

The c4-highcpu-4 I'm using costs $0.1701/hr. Running 24/7:

But if the only thing that matters is core frequency and I'm keeping the machine on 24/7, the best option isn't a VM — it's an Intel i9-14900K running at 6.0 GHz turbo. If the solver scales linearly with frequency, each flow would drop from ~2.19s to ~1.53s. 43% faster.

Compute isn't cheap. But it's what needs to be done. For an AGI to understand the electrical grid, someone has to generate the data. A PC running 24/7 simulating contingencies, running power flows, tuning protections. That's what I'm going to do.