PUMS vs. ACS

(Turns out, they don't split up households, as expected.)

Glad you found what you were looking for. I'm adding in some info for the record in case it's helpful to you or others. I find that I have a hard time remembering there things are located because the information is spread over so many documents. But I was recently trying to answer a similar question and most descriptive documentation I could find was in the 1-year ACS PUMS Accuracy document, on page 5.
www2.census.gov/.../2016AccuracyPUMS.pdf

"SAMPLE DESIGN
The 2016 PUMS was designed to include one percent of the housing units and one percent of the
GQ persons in the United States and Puerto Rico. The PUMS sample was selected from the full
sample ACS records separately for Housing Units (HUs) and GQ persons. The PUMS sample
sizes were based on the Population Estimates Program estimates for housing units and GQ
persons.
The PUMS sample of persons in households was selected by keeping all persons in selected
PUMS HUs. The systematic sampling method used sampling intervals chosen to yield the
sample sizes given in Table 1 and Table 2 by state, DC and PR. The sampling interval for each
state and HU/GQ sample is the ratio of the number of interviewed records available for sampling
and the required sample size (sampling intervals are not rounded to integers).
..."

The above compares with this link about overall ACS sample sizes and table B00002 (also B98001):
www.census.gov/.../
factfinder.census.gov/.../B00002