slower clock speed, faster performance?

[mickeddie]

How is it that a P4 running at 3.4ghz is slower than an AMD 2.6ghz or Conroe 6600? Please explain in laymans terms.

Eddie

 

[RichPLS]

The P4 performs less work per clock cycle than AMD and even Conroe CPU's... Also, generally longer pipelines are used to ramp up MHz, but if the pipeline is not needed, or pre-fetch data is wrong, the pipeline must be dumped and re-loaded, wasting clock cycles...

Conroe running 400MHz slower than AMD's FX-62 beats it solidly across the board...

 

[angry_ducky]

Conroe running 400MHz slower than AMD's FX-62 beats it solidly across the board...

From what I've seen, the Conroe running 1GHz slower than the FX62 equals it.

Also, if you compare two identical CPUs from the same product family, but one has a higher GHz than the other, the one with the higher GHz will perform better.

Here's an example: The Athlon 64 3700+ and 4000+ (socket 939) are both based upon the San Diego core, and they have the same amount of L2 cache. But the 4000+ is a 2.4GHz CPU, rather than 2.2. If one were to overclock the 3700+ by 200MHz, then the two CPUs would perform the same, but at the stock speeds the 4000+ wins.

 

[deweycd]

Now, my question goes to why not do more work per clock cycle all the time?

Also, its the reason that AMD was able to be faster then Intel with different frequencies.

 

[SidVicious]

Now, my question goes to why not do more work per clock cycle all the time?

Also, its the reason that AMD was able to be faster then Intel with different frequencies.

Simply because IPC were'nt a priority in the Netburst architecture.

Designing a CPU is'nt something that happens overnight, it requires tremendous R&D ressources to create something that complicated, keep in mind that were talking about a few hundred million transistors etched into a surface that is less than a few hundred squared milimeters across.

 

[gOJDO]

Longer pipeline means less logic per stage. Each stage needs time to finish its task. With less logic per stage, the stage can finish its task faster so you can force it to do more tasks in a second and that means to raise the freqfency.
The penalty of the longer pipelines is when branch missprediction ocurs.
Anyway, no matter of the length of the pipeline, the K8 is more advanced and more efficient architecture than the Netburst.
Core2 Duo is even more advanced and it has 4 executing paths, while K8 and Netburst have only 3.

 

[1Tanker]

If one were to overclock the 3700+ by 200MHz, then the two CPUs would perform the same, but at the stock speeds the 4000+ wins.

True in theory, but in order to get that 200MHz, you need to increase FSB, which should have it running ever-so-slightly faster than the 4000+.

 

[mesarectifier]

True in theory, but in order to get that 200MHz, you need to increase FSB, which should have it running ever-so-slightly faster than the 4000+.

Reading a bit far into it, but if you run them both at DDR400 (FSB200) at stock speed and then increase the multiplyer by one for the slower CPU then it's same speed, same FSB (surely).

God, I really hate it when people get pedantic

 

[1Tanker]

True in theory, but in order to get that 200MHz, you need to increase FSB, which should have it running ever-so-slightly faster than the 4000+.

Reading a bit far into it, but if you run them both at DDR400 (FSB200) at stock speed and then increase the multiplyer by one for the slower CPU then it's same speed, same FSB (surely).

God, I really hate it when people get pedanticHere's more nitpicking for you. You can only lower the multiplier on A64's(surely).Only the FX chips have fully open multipliers.

 

[mesarectifier]

I think a call of touché is in order

 

[m25]

The P4 performs less work per clock cycle than AMD and even Conroe CPU's... Also, generally longer pipelines are used to ramp up MHz, but if the pipeline is not needed, or pre-fetch data is wrong, the pipeline must be dumped and re-loaded, wasting clock cycles...

Conroe running 400MHz slower than AMD's FX-62 beats it solidly across the board...

That's just it. With Conroe, intel has adopred an AMD-like pipeline (14 stages compared to the 12 of the A64s). Having almost the same pipeline, Conroe performs better ecause it does even more work even for less clocks than the A64s ; its got more processing units and they're larger.
That's how CPUs are evolving nowdays; those brains are just getting larger, it was useless for them to be just faster and the prescott was an avolutionary failure.

 

[angry_ducky]

If one were to overclock the 3700+ by 200MHz, then the two CPUs would perform the same, but at the stock speeds the 4000+ wins.

True in theory, but in order to get that 200MHz, you need to increase FSB, which should have it running ever-so-slightly faster than the 4000+.

You could set the 3700's base clock to 218MHz and have a 2398MHz CPU, or set it to 219MHz and have a 2409MHz CPU. Both numbers are close enough.