Hier een stukje FAQ over Northwood vs Prescott :
The big question:
Why doesn't the Prescott with its twice larger L1 and L2 caches and architectural enhancements outperform the Northwood with ease? (refer to table above for numbers)
Prescott Advantages over the Northwood:
1. Twice larger L1 cache.
2. Twice larger L2 cache.
3. Integer Shift/rotate functions can run much faster due to improved ALU
4. Integer Multiply functions are much faster due to new dedicated integer multiplier (Northwood integer multiplies are done by the FPU unit)
5. New improved branch predictor to reduce pipeline stalls
Prescott Cache Disadvantages compared to Northwood:
1. Double the L1 cache latency
2. ~50% higher L2 cache latency
3. Half the L2 cache bandwith
4. 50% longer pipeline
ANALYSIS:
1) Although the cache sizes of the Prescott are twice higher, the cache latencies are also greatly increased, which somewhat offsets the increase in cache size. Caches generally work best when they are small and fast, not big and slow.
Also, according to the testing done by the guys at Digit-Life, the Prescott's L2 bandwith is only half of the Northwood. Perhaps intel reduced the Prescott's L2 bus width to 128bit (Northwood is 256bit).
2) The larger cache can barely offset the performance of the 50% longer pipeline (31 vs 21 stages). Longer pipelines means that a pipeline stall (due to branch misprediction) will be much more detrimental to the performance.
CONCLUSIONS:
Due to 1) and 2) above, the Prescott's performance is slightly lower than the Northwood in most situations, but they are roughly on par with each other.
An interesting observation made by the guys at Aceshardware is the power consumed per transistor of the Prescott. It is actually ~46% lower than the Northwood! But the immense number of transistors (more than double the Northwood) add up to make the total heat output higher.
So if you are looking for a P4 at the moment is get a Nortwood and stay away from the Prescott until Intel fixes the heat issue with the Prescotts, because the Prescotts feature a much higher heat output and lower performance than the Northwood.
This high power usage and heat output is due to the transistors leaking a lot of current ( 40% even in the OFF state!) with the new 90nm manufacturing process. The problem is happening with everyone's 90nm process (yes, AMD and IBM included) at speeds roughly above 2.5GHz.