Delays match almost exactly, but since power is aggregate the error tends to accumulate in one direction so it is off more.

Regarding Xyce being closer to Hspice than ngspice; in recent years we’ve spent a lot more time attempting to be similar to Hspice. We’ve never really targeted ngspice for comparisons, but in the old days (15-20 years ago) we used to compare with Spice3 a lot, and ngspice is basically fork of spice3. In general, I can’t say which of the two codes (ngspice or Hspice) is “better” in an absolute numerical sense. I know that Hspice takes a LOT of shortcuts to improve its speed, and I have occasionally observed them simply getting the wrong answer on some circuits. On the other hand, most commercial PDKs are setup to work with commercial EDA tools such as Hspice and Spectre, so they tend to be treated by most companies as the gold standard.

One other small comment; a lot of older SPICE models are numerically kind of bad. Lots of discontinuities, inaccurate derivatives, occasional simplifications to run more efficiently on 1970s era computers. They also often contain details that aren’t good or bad, but are undocumented. People generally expect those models to behave exactly the same as they always have. So, we coined the term “bug-compatible software”, to describe what we had to do there. Fortunately, most newer SPICE models don’t tend to have these issues (“newer” meaning the last 20 years). But the newer models are often harder for people to understand (the BSIM models have 100s of parameters, for example). As a result, the old “buggy” models are still popular for some uses.

PTM and Sky130 should both use relatively new MOSFET models, so my comment about the old buggy spice models is mostly moot. But I thought I’d mention it as it can be relevant for simulator comparisons.