How big of an advantage will the P4s ALUs become?

[RaynorWolfcastle]

I have read about many people praising Intel for their double-pumped ALU units. How big of a factor will this become as the P4 scales up in speed? a 100 MHz boost in clockspeed means a 200 MHz boost in ALU speed.

Does that mean that (theoretically at least) a 2.0 GHz P4 could run certain programs that are highly dependant on the ALU nearly 66% faster than a P4 1.5 GHz even though the clockspeed is only 33% higher than the 1.5 GHz processor (assuming everything else is the same.

I'm curious, how does the ALU unit impact the overall speed of a system? By that I mean, which applications benefit most from an ultra-fast ALU unit?

-Ice

 

[CTho9305]

integer stuff... so I guess mostly business applications. the irony is, there aren't many business apps that need more than a ~300 mhz processor anyway. games probably wouldn't benefit as much as they would from a good fpu

 

[Sohcan]

<< I'm curious, how does the ALU unit impact the overall speed of a system? By that I mean, which applications benefit most from an ultra-fast ALU unit? >>

Personally I don't think it's a huge factor...what Intel did was split the 32-bit ALU into two halves; one 16-bit ALU for the lower-order bits, another 16-bit ALU for the higher-order bits...therefore each half can be clocked twice as fast. An ALU operation is staggered, so it takes two fast ALU cycles to complete the entire 32-bit operation. Therefore, the latency isn't changed, but the throughput is theoretically doubled. If the instructions remain within the fast ALU units (for adds, subtractions, and logical operations), the P4 can execute them twice as fast (assuming no multiples or shifts are encountered), but since the P4 is still limited by how often it can issue and retire instructions, it's likely only useful in short bursts.

IMHO the reason they did the double-pumped ALU is so that the ALU is no longer holding back the overall clock speed. ALU theory hasn't changed at all in decades (everyone still uses the same carry-lookahead adders), and the add operation is the major source of the critical path length in determining clock speed. By splitting up the add operation, the ALU will always be clocked twice as fast, and cannot hold back clock speed.