I also thought the sub-ms numbers can't be right — SQLite is fast but not that fast on millions of rows, but didn't look into it until now.

Turns out the benchmarking code is wrong: it didn't read the rows returned from db.Query, so the mattn version simply didn't wait for the results to arrive. Once you apply this patch:

  diff --git a/cgo/main.go b/cgo/main.go
  index 8796b3d..9a74a2f 100644
  --- a/cgo/main.go
  +++ b/cgo/main.go
  @@ -82,11 +82,15 @@ CREATE TABLE people (
        panic(err)
       }
      }
  -   fmt.Printf("%f,%d,insert,cgo\n", float64(time.Now().Sub(t1)) / 1e9, rows)
  +   fmt.Printf("%f,%d,insert,cgo\n", float64(time.Now().Sub(t1))/1e9, rows)
   
      t1 = time.Now()
  -   _, err = db.Query("SELECT COUNT(1), age FROM people GROUP BY age ORDER BY COUNT(1) DESC")
  -   fmt.Printf("%f,%d,group_by,cgo\n", float64(time.Now().Sub(t1)) / 1e9, rows)
  +   res, _ := db.Query("SELECT COUNT(1), age FROM people GROUP BY age ORDER BY COUNT(1) DESC")
  +   for res.Next() {
  +    var count, age int
  +    _ = res.Scan(&count, &age)
  +   }
  +   fmt.Printf("%f,%d,group_by,cgo\n", float64(time.Now().Sub(t1))/1e9, rows)
     }
    }
   }

modernc SELECT performance becomes pretty comparable, actually a little bit faster than mattn on my Intel Mac with high row count.

Not only that, modernc INSERT is noticeably faster on my Intel Mac...