That's about right. Half the TPS with half the power.

GLM-5.2 will be much more demanding tho

Which quants? I get these speed (only 20-30 TPS) at Q4_K_M with MTP for 27B on my framework desktop. I draw sub 130W for the whole machine.