Running a 1x multi

[Digital Pimp]

first off if this dosent belong here please move it.

with todays increacing memory speeds what sort of performance gain/loss would you experience if you 'could' run your processor with a 1x (or close to it) mulitplier.

i see DDRII on the new GeforceFX will be running at 500Mhz (1Ghz DDR) if they eventually did make the same, or faster memory for system ram and motherboards started supporting 1x mulit's would there be any performance gain running the processor and memory in sync.

Digital

 

[Cluster]

There won't be a difference. The only thing that matters is the end result of the total CPU MHz. So a 1Ghz x 1 or 100MHz x 10 will have the same CPU performance. You will have tremendous amount of total processing power throughout your system though because of the increased memory bus.

 

[Cheesy Peas]

thats true to a certain extent, put the increase in memory bandwidth has the greatest increase in performance... that y most people overclock using fsb

 

[Black_Paladin]

For CPU's, 10x100 does NOT give the same performance as 1x1000. Look at any chart, the system with the greater amount of memory bandwidth is always the faster system (the system with the higher FSB, which in our example would be the 1x1000 system where 1 is the multiplier for the CPU and 1000 is the FSB Mhz) Just like Cheesy Peas had said, this is why people overclock using the FSB.

And on Athlon systems, people unlock their CPU's not to just get higher multipliers to be able to overclock but so that they can get LOWER multipliers. This is because if you use a lower multiplier for your CPU, you can increase your FSB on the motherboard and keep your CPU at the same Mhz if you wanted without any additional cooling. Your CPU would be running at the same speed so you only need to worry about your motherboard overheating.

Needless to say, you would never be able to run at 1x1000=Mhz with a CPU because there is no motherboard on the market that is even close to running the FSB at 1000Mhz right now. The example you just gave above was exagerated.

To give a more realistic example however, an Athlon system running 1400Mhz with a 14 multiplier and 100Mhz FSB will be crushed by an Athlon system running at the same Mhz but with say a lower multiplier like 8.5 and a higher FSB like 166 (166x8.5 will give you 1411Mhz which is pretty close to 1400Mhz). Memory bandwidth is definately something that is very important. I had made a chart on vr-zone forums about what kind of performance you can expect from a CPU when you overclock it using the FSB as opposed to overclocking it using the multiplier. I can look it up and put a link for it here if you are interested.

So, to answer your question Digital Pimp, a CPU with a 1x multiplier and a 1000, 2000 or whatever FSB on a motherboard would TOTALLY obliterate every CPU on the market right now. This is everyone's dream. A CPU like that, (let's say a CPU that is running 1x1000) would PROBABLY be much faster than a CPU running 15x100 because the enermous amount of increased memory bandwidth on the 1x1000 CPU would make up for the amount of extra CPU mhz on the 15x100 CPU. This example, of course is if we were comparing similar CPU arhitectures like a P4 Northwood 1x1000 vs a P4 Northwood 15x100. Comparing a 1x1000 Pentium 3 and a 15x100 P4 would be like comparing apples to oranges.

Also, the P4 architecture seems to like the increased memory bandwidth a lot. I am sure the P3 architecture likes it too but because of the fact that the FSB is quad-pumped for the P4, for every Mhz you increase the FSB on the motherboard, you get a 4Mhz total increase for the FSB. Because of this, you can increase your FSB by a pretty big amount on a P4. A 200 Mhz FSB set in BIOS will give you a total FSB of 800Mhz for example. This is way above anything a P3 could ever hope to achieve. The most a P3 could ever do is limited by the motherboard you are using (that is if you had an unlocked CPU and an EXTREMELY good motherboard) and that could be something like 250Mhz. Of course there is a big difference between 250Mhz and 800Mhz so I wonder what kind of performance a quad pumped P3 would
produce.

Anyway, the point is that some CPU's really like the increased FSB such as the P4 and will give you big performance benefits for every increase you do for the FSB up to a point where you start to get diminishing returns. I do not know what that point is. it is definately not 800Mhz though.

For your example, 1x1000 would translate into 4000Mhz FSB speed for a P4 (since FSB is quad-pumped). I would LOVE to see what that would do for performance.

Oh, another REALLY important note is that, to be able to really take advantage of that FSB speed, you would need to have some RAM that can handle that speed as well. So for our 1x1000 example, for a P4, you would need to have some kind of RAM that can handle 4000Mhz to be able to run synchronously with the FSB. Otherwise, your RAM becomes a big bottleneck for your system and you don't get a whole lot of improvement. Remember that right now, the fastest DDR on the market you can buy for your motherboard (you can't really use DDR for graphics cards on your motherboard) is 433Mhz. This speed is lower than the speed of the current P4's FSB which is 533 (133x4) and so memory becomes a bottleneck for your system. Dual Channel chipsets change this fact of course as on a Granite Bay (intel) or an Nforce 2 chipset (AMD), you can theoratically get your memory speed up to 433x2=866Mhz BUT then you run into other problems such as latency which are best discussed in a different thread.

Sorry for making an answer this long but I just wanted to be comprehensive.

 

[demon-eater]

woah.. paladin that was good, very good gramar as well not to mention very clear and to the point!

 

[Black_Paladin]

Thanks Demon!! I think my grammer is not that great actually but I tried to make my points as clear as possible.

 

[Digital Pimp]

yes, great reply black_paladin

after posting i realized that when running with a 1x multi, ram wouldent be memory anymore, it would be something like Cache memory, running at the same speed as the proc, so in theory you wouldent have to run 1Ghz+ to acheive high speeds (eg. Alpha processors and to a lesser extent Xeons). Kinda like having 512mb of cache . as an example running the processor at 500 x 1 =500Mhz, this setup would most likley blow any of todays processors out of the water, and would be possible shortly if motherboard manufacturers started supporting anything lower than a 6x multipliers.

but then of course new bottlenecks would pop up such as hard drive speeds as well as the PCI bus being able to supply the nesscisary information to the processor and memory.


Digital

 

[Penguin4x4]

That is, until PCI Express, aka 3GIO hits the scene next fall...

http://www.pcisig.com/specifications/pciexpress

 

[SBeaver]

I was thinking the exact same thing as you Pimp.
Multiplers is almost a sign of lazyness.

 

[ninthebin]

would there ever be a point with your 1x model that the extra bandwidth added would cease to give the performance gain you see in the extra bandwidth ?

Say if you were to plot the Bandwidth usage percent on a graph against clock speed of a the cpu. Would you see the bandwidth usage as a straight line constantly going up - or would it produce a curve that shows that eventually bandwidth becomes less effective...

obviously even if the bandwidth became useless increasing would eventually lead to need that bandwidth - but then if its on a curve it would always have catch up periods.

 

[RichardLizzard]

wow, erm well, read it think I got the tecnical side of it up,

in english: the model of 1:1 ram:cpu ratio would be preferable
only if indeed the ram had a decent lateney and such,
but if ram at the moment such as dual channel ddr 333
supplys the P4 at 533 fsb,
with enough bandwidth to do its job fully without hanging round waiting for the data too much (and I'm fairly sure that theoretically it should) would the multiplier matter that much going down if the ram performs as it should?

I think I'm saying that a 1x mutliplier really isnt needed as much as a decent memory interface is... maybe... (i seem to be ending alot of posts with maybe..)

 

[nealric]

I think the good old economics phrase diminishing returns would come in handy here.
You might see big ganes if you could run a 5x mult vs 10x (at same clock)
Going to a 2 would be less
going to a 1 would be still less
After a while, fsb ceases to be a bottlneck.

 

[larva]

The importance of memory speed is indeed real, but being overstated in this case. The important factor that has not entered the thread (yet) is the efficiency of the caches that exist for no other reason than to isolate us from the effect of the relatively slow memory subsystem. If you look at the internal L1 and L2 caches of modern CPU's they have hit rates in excess of 90% each. Because there are two cascaded the effective hit rate is high indeed. It's not that you don't have to read from memory at some time, but the incredibly astute optimization of the caching in modern CPU's actually does this before it is needed, keeping the pipelines full.

This is not to say that no improvement is possible. Indeed faster memory subsystems contribute to a faster system. But the gains are not as far reaching as indicated here, and the magnitude of the potential improvement has been overstated. Only if there was no caching on the CPU, or if the effectiveness of the caching were of a low level would the monumental difference depicted above materialize.

A good example is the last leading PC processor to run the memory as fast as the core, the i486DX50. People claimed that the later introduced DX2/66 chip would be no improvement, or even slower because the fsb and memory would only run at half processor speed. An enormous 2X multiplier would be utilized to derive the CPU's working frequency. In this era we see multipliers in the 7-26 range.

When the DX2 chips hit the market, it was easy to pit the 486DX50 against the 486DX2/50 and quantify the effect of the half speed memory. And lo and behold the DX2/50 yielded 97% of the performance of its more theoretically pure sibling. So while we strive to run our memory as fast as we can to bring it closer to the clock rate of the CPU whenever we can, we have to remember that the gains from doing so lie at a lower level than our gut instinct might predict.

 

[Digital Pimp]

larva, you bring up a good point, at some point there must be a "sweet spot" where lowering the mulit wouldent contribute to efficency. (eg. were a 4x would be better than a 5x, but not better than a 3x)

Digital

 

[slater3333uk]

the whole piont of running the cpu faster than the fsb is that the cpu has more work to do in a given time than the ram and other components in the system eg. the ram only has to transport and store info but the cpu has to atualy procsses it

therefore if we were to run with a 1x multiplyer the majority of the fsb bandwidth would be useless as the cpu could not keep up with the amount of information that it is being given from the rest of the system.

if you tryed to run a 1ghz chip on a 1x mulitiplyer this would mean that the fsb (Northbridge chip) would have to run at the same speed creating more heat and we know what heat dose to our systems. this would just be creating unessisery heat as not all of the fsb will be used.