Does a high FSB really make a difference?

[bgeh]

Originally posted by: mngisdood
IIRC the higher bandwidths are more benefitial to P4 processors (esp. p4c) than to AXP's. I'll try to dig up some numbers...

i agree, for the p4 architecture, higher bandwidths are more beneficial, while for the Athlon XP's, lower latencies are better

 

[TerryMathews]

Originally posted by: Jeff7181

Originally posted by: TerryMathews

Originally posted by: Jeff7181

Originally posted by: TerryMathews

Originally posted by: Jeff7181
Anyhow... the benchmarks I did were real world performance in real games. Not many people trust synthetic benchmarks to give an accurate depiction of performance anymore.

Right, but that's not a synthetic benchmark. Big difference.

May not be a synthetic benchmark... sure you can say more is better, and 2.5 GB/s is more than 2.1 GB/s... but what does that tell me about what it's actually going to do when I play a game? Not a whole lot... and even less to a novice.

By you're logic, people should question CPU freqs *in the same processor model* as well. They don't. People automatically assume more is better, which is correct. More is better. End of story. There is no down side to having more RAM bandwidth.

So then tell me, exactly how many FPS will I gain in UT2k3 if I had a P4 2.4 and switched to a P4 3.0?
You can't... and that's my point... REAL WORLD RESULTS are more useful than all the specs in the world.

As a matter of fact, I could. But that's not the point. The point is that more MHz is better.

Gragland asked if a higher FSB is better. I gave him a formula to prove that he'd gain about 20% more RAM bandwidth. This will show more when your application is RAM bandwidth limited and less when your bound by CPU or GPU. But, you will always (Or, almost always) gain some performance. And you'll never lose performance.

I'm not disputing your 'REAL WORLD RESULTS'. They're a great tool to show how a praticular application benefits from more RAM bandwidth. But if you go by benchmarks alone, you're a fool.

 

[Jeff7181]

As a matter of fact, I could. But that's not the point. The point is that more MHz is better.

Then tell me... prove me wrong... show me the math you use to do it, then we'll go look at benchmarks and see if you're right.

 

[Quackmaster]

Originally posted by: bgeh

Originally posted by: mngisdood
IIRC the higher bandwidths are more benefitial to P4 processors (esp. p4c) than to AXP's. I'll try to dig up some numbers...

i agree, for the p4 architecture, higher bandwidths are more beneficial, while for the Athlon XP's, lower latencies are better

I too agree. As demonstrated at Aceshardware (not to mention Anands )A lack of internal bitwidth is bottlenecking the Athlon. Just as the P6 Pentium core began to run out of steam at the end of it's architecural life-cycle, so to has the Athlon been running out of the headroom that existed at lower FSB and RAM speeds. There were a pair of nice articles that are easy to read detailing just this scenario and also here. They are rather old, and yet so prophetic it would make any microprocessor student take note (i sure did). Can't wait to see how this theory plays out with Athlon/Opteron's onboard DRAM controller. It ought to be very interesting.

People automatically assume more is better, which is correct. More is better. End of story. There is no down side to having more RAM bandwidth.

Terry, I agree that we all love more RAM bandwidth, but there IS a downside. When the saturation point is reached on any bus or interconnect, whether it be fsb, PCI, internal cache-to-execution-pipeline(bitwidth) or even Gbit LAN, we see a ratio of diminishing returns. As that saturation point is neared, the COST of producing silcon and interconnects (buses) able to handle those higher frequencies becomes increasingly less beneficial because the performance ratio is diminishing.

Think of it as having a huge enigine in a car being fed by too small fuel lines. Starvation is occuring. This is exactly what happened when the Pentium4 was first released. The only difference was that most of this starvation was occuring for the most part internally . With it's small 256k cache and it's 100MHz ("400") for you marketing junkies), the execution units were starved. Once Northwood (512k L2 cache) arrived, the improvement was huge due to the P4's heavy dependency on it's Trace cache in helping keep the pipline full (as well as other factors).

But now let's take that same "huge engined" car and give it massive fuel lines, carbs, exhaust etc. There will be a point at which the hardest thing to do is get the power from the engine to the ground (more friction). Also our fuel consuption will become amazingly large compared to our fuel economy (not to mention we'll be wanting 100 octane or better). Diminishing returns... So at what point and time will the P4 architecture run out of headroom. And will it lose it's ability to ramp it's bandwidth (internally or externally) first, or will the engine (P4 cpu) just poop out? The current .13micron process is about to run out of headroom, how far will .09 and whatever follows take us? Also, how much will it cost?

Intel's plan with the P4 architecture is to continually increase bandwidth everywhere, both within the cpu and externally on the buses, so that as the GHzs ramp up, the cpu is never starved. Prescott's (the upcoming P4 "3rd edition") larger caches (both L1, and L2) and the fsb being increased to 300MHz (1200) are examples of this. Gigabit LAN and AMD's Hypertransport are also similar in that respect except they are more useful to the networking equivalent of starvation (lets not even get into concurrency. *evil flashback to my old Professor :evil: from JPL crushing my brain*)

Perhaps a good example of this "bandwidth vs cost$ of technology" is Gigabit ethernet. The controllers for Gbit LAN are far more expensive than 100Mb, but can you actually USE the extra bandwidth provided? Or will the cost of multiple NICs, hdds, and everything else needed to fill that Gbit pipe prove to be prohibitive?

If all this is old hat to you, I apologize for the friggin' classroom session.

 

[DSE]

Originally posted by: GRagland
So let me get this straight:

No matter what chip I am getting (even if it is a bad oc'er), it would be benificial to get pc3200 because i can drop the multiplier and raise the fsb. And that would give me a substancial performace boost worth the extra money, without even actaully overclocking (mhz of the chips stays the same).

So the advantages of pc3200 over pc2100 is NOT that pc3200 it is better for overclocking, cause any ram will let you overclock the same distance mhz wise, but that the advantage is that once you REACH YOUR MAX OVERCLOCK, then you can adjust the fsb/multiplier ratio to achieve better performance. (and of couse better performance= higher fsb)

For the most part, yes. One thing you have to keep in mind is that dropping multiplier and raising FSB does not guarantee the same speed. CPU FSB maybe limited on it's own. a 266Mhz FSB CPU just may not be able to hit 400 or 333Mhz at all, no matter how much you drop the multiplier. So if you end up with a proc that'll only do 333Mhz FSB, the PC3200 would be an overkill.

 

[GRagland]

Of course more bandwidth is better, but i wanted to know if that extra bandwidth was WORTH the money. Thats why i wanted real world test results. Please PM eachother if you want to argue.

Okay, I'm getting a 2500+ and Epox-8RDA+, what ram would you recommend. 2x256 or 1x512 is what i want. I was thinking about Winbond CH-5 2x256 PC3200 which i can get for $94 shipped. But, I heard that the 2500+ has been having trouble at very high fsb speeds, so it may be pointless to get such good ram (like what DSE said above this post). Here's a review of the Winbond CH-5.

Also, since i have a 40gb and a 80gb (8mb cache) hard drive, would it be worth it performance wise to get a abit nf7 2.0 which has RAID ($40 more then the 8RDA+)?

 

[TerryMathews]

Originally posted by: Jeff7181

As a matter of fact, I could. But that's not the point. The point is that more MHz is better.

Then tell me... prove me wrong... show me the math you use to do it, then we'll go look at benchmarks and see if you're right.

Well, I've got better things to do with my time, but just to prove a point.

Take a sampe of two or more identical systems running the same benchmark, different processor speeds. Plot the CPU frequency on the x-axis and the FPS on the y-axis. Find a best-fit line between your given points and extend it out to the desired frequency. This graph will capture the diminishing returns, as there would be diminishing returns between say 2.4 and 2.8 GHz. I'm not going to plot the points for you, but any Algebra or better student will realize this graph would actually be very accurate given accurate samples.

Gragland: your budget is your own business, but I'll tell you right now that Athlon or P4, increased FSB and memory bandwidth will be one of the cheapest performance boost you'll get.

EDIT: Gragland, read your bit about the 2500+. I haven't heard any reliable information about Barton's not running at 200+ FSB. By reliable, I mean a board that's known to run at 200FSB, and a Barton running with it's mult lowered so that it's running at default clock speed and 200MHz FSB. Most of the complaints I've heard are people who can't just plug the Barton in and get it to run at 200MHz FSB.

This 'discussion' has come up before too. My point, which I'll make again, is that there is a part of the Athlon that deals with the FSB, it captures it and derrives it's internal clock off of it. It's impossible to imagine, however, that a chip that can run at 1000+ MHz has a component that won't work at 200.

 

[GRagland]

Im not sure if i should get the 8rda+ for $85, or the abit nf-7 for $125. The abit nf-7 has RAID, and i have a 40gb and a 80gb (8mb cache) hard drive. Would i benifit significantly from RAID?

 

[beatle]

When I bought ram, I didn't buy PC3200 for speed up front, I bought it to future proof myself for the P4's with 800mhz fsb and the athlons with 400mhz fsb. This way I can get another upgrade cycle out of my current ram. Consider picking up a 512 meg stick of Twinmos PC3200 for $90. Save your pennies until you can buy another stick. 512 should get you by for most things, plus, you'll likely save money when you buy your second stick. My Twinmos cost me $122 2 months ago. It's down to $90 now.

 

[m305]

yes it makes ur pc faster the higher the fsb

multiplier*fsb=clock speed say
11*150=1650

 

[infinite012]

Originally posted by: GRagland
Im not sure if i should get the 8rda+ for $85, or the abit nf-7 for $125. The abit nf-7 has RAID, and i have a 40gb and a 80gb (8mb cache) hard drive. Would i benifit significantly from RAID?

I thought you couldn't use RAID unless you had 2 drives that are identical in size and performance (RPMs, cache, etc, etc) so if this is true, then RAID [with 40 and 80GB hard drives] is false.

 

[GRagland]

maybe thats true, i dont really know about RAID, anyone know for sure?

 

[TerryMathews]

Originally posted by: GRagland
maybe thats true, i dont really know about RAID, anyone know for sure?

No, you can do different size drives in a stripe. You'll only get double the size and performance of the slowest and smallest drive, however.

 

[Jeff7181]

Originally posted by: TerryMathews

Originally posted by: Jeff7181

As a matter of fact, I could. But that's not the point. The point is that more MHz is better.

Then tell me... prove me wrong... show me the math you use to do it, then we'll go look at benchmarks and see if you're right.

Well, I've got better things to do with my time, but just to prove a point.

Take a sampe of two or more identical systems running the same benchmark, different processor speeds. Plot the CPU frequency on the x-axis and the FPS on the y-axis. Find a best-fit line between your given points and extend it out to the desired frequency. This graph will capture the diminishing returns, as there would be diminishing returns between say 2.4 and 2.8 GHz. I'm not going to plot the points for you, but any Algebra or better student will realize this graph would actually be very accurate given accurate samples.

Gragland: your budget is your own business, but I'll tell you right now that Athlon or P4, increased FSB and memory bandwidth will be one of the cheapest performance boost you'll get.

EDIT: Gragland, read your bit about the 2500+. I haven't heard any reliable information about Barton's not running at 200+ FSB. By reliable, I mean a board that's known to run at 200FSB, and a Barton running with it's mult lowered so that it's running at default clock speed and 200MHz FSB. Most of the complaints I've heard are people who can't just plug the Barton in and get it to run at 200MHz FSB.

This 'discussion' has come up before too. My point, which I'll make again, is that there is a part of the Athlon that deals with the FSB, it captures it and derrives it's internal clock off of it. It's impossible to imagine, however, that a chip that can run at 1000+ MHz has a component that won't work at 200.

You still didn't prove me wrong... I asked you to prove that you can tell me the EXACT FPS increase you'll get... you claimed you could do it.