I’m not sure the cpu load being higher is a bad thing, isn’t that basically showing that it is using less IO so it can use more of the cpu?

The throughput is way, way higher, so it’s using less cpu per transaction. If this were showing equal numbers of transactions the CPU usage would be lower.

Ideally, in a benchmark, I think we’d be seeing basically 100% cpu usage because that would mean the test hardware is being fully utilized and the software being tested isn’t being bottlenecked in some way.

Given the higher TPS I think you'd be better conceptualising it as "CPU utilisation" rather than "CPU load."

It is not about my "conceptualising", I was simply referring to the fact that the conclusions refer to something different than the graphs. The author himself uses the term "load", so I guess we should stick to what he meant. Still, graph is showing the absolute value of the load, not per-transaction value. Then it takes an extra effort to actually realize that maybe the author's claim is valid, but in a per-transaction context. Why then the graph wasn't made to plot per-transaction values if that was author's point? It adds unnecessary confusion. That confusion is perceived just after reading another inconsistency - that there is supposed 5x tps speedup while we see 4× (visible 80k divided by visible 20k is 4 not 5). So why 5x? Was this 5x based on a median perhaps, or on some percentile- then why such a median or percentile wasn't shown on the graph? ..and so on.

Please don't get me wrong. 4x tps speedup is nice achievement already. It's great enough to congratulate the author and be happy. But it's also presentation of the result that matters, if there are inconsistencies, or the author based his claims on a different measurements than what is shown, then it's natural that it can make one to raise in eyebrow. It doesn't solidify the trust, as opposed to presenting the conclusions matching the graphs exactly.

I think you might have misread the graphs. The graph is showing a 4x peak/~6x average improvement in TPS. Because of this the load is less I/O bound and thus CPU is able to be fully utilised. If you want to measure efficiency you would instead measure at constant TPS.

I suppose that more CPU load is because more cores are loaded by doing tasks in parallel. Less CPU load is actually indicator that performance is limited by something else (IO, locks, memory coherency) that couldn't be scaled with the ease of CPU's adding.

So bigger CPU load and bigger performance mean that parallelising of tasks is more efficient, which is a clear benefit.