Interesting choice of radix-8 booth multiplier that needs the x3 circuit. Seems like a area/perf tradeoff in order to push the fmax which means they had a latency cycles constraint since the same could be achieved via more pipelining.
You feed the partial products into a carry-save adder tree and choose a level which minimizes the delay and the extra flop area/power. I am not sure about the pentium technology but whatever the area, it might be comparable to the x3 multiplier
There is an implicit add in the x3 multiplier circuit that increases the delay which was deemed acceptable than the radix-4 delay. Considering all this, I strongly believe latency was a hard constraint (may be for SW perf).