Does CPU speed trump RAM speed in the real world

[mhw100]

I'm not a hardware genius as evidenced by my question.

I'm looking a several machines that very in their build. Both use the QX6700 quad chip. One build is water cooled and is stable in the 3.6Ghz range but the memory is 667Mhz but pushed to 750Mhz. The other build is air cooled and is stable at just over 3.1Ghz but the memory is ~1000Mhz.

My understanding that the memory speed is more important for gamers and bench markers but in the real world such as business applications it wouldn't manifest a noticeable difference between say 750 and 1000. The trading application I'm using it in requires a fast processor and fairly large draw on memory. It is 2D so graphics aren't heavy. In reading other threads it seems speed is king but how this equates with RAM speed I don't know or if I should even be concerned about the 750 versus 1000 difference.

So my question is which of these builds would be more appropriate for my trading application assuming all other things being equal in the system?

Thank you.

 

[AeroB1033]

CPU is still going to be the main bottleneck over memory bandwidth. But if you want to see big real-world performance gains, a fast HDD is the way to go.

 

[Scougs]

First of all, the term "real world" usually applies to anything outside benchmarks. There are many applications that would benefit from higher memory speed. It is difficult to say which of the systems you mention would be faster in any program since the clocks are so much different.

It might help people answer your question if you say what program you are planning to use. I can't imagine how any stock trading program (if that is indeed what you are talking about) would need bleeding edge performance. It would seem like your internet service provider would be the most important thing for you in the case of online stock trading. If I'm not mistaken, most business applications would run just fine on almost any computer.

The kind of performance that you are talking about would be more along the lines of doing research engineering work or top end gaming, etc.

 

[mhw100]

The application is TradeStation. It is very processor intensive as it has to do mathematical research optimizations as well as running live complex math calculations. It isn't just a online exection program but rather the user has the ability to program very complex math models to run within it. It currently uses 2 cores and is evolving to use more.

It appears to be quite memory intensive as well, but not necessarily for what I use it for on a live basis...for others it can be.

There is not a ISP or data bottleneck problem. Graphics are 2D so basic 256 cards are fine.

Of the systems I'm looking at they use either RAID 0 Raptors or SCSI.

Both systems come boxed and warrantied at ~ 3.2Ghz. So assuming I'm not going to raise that bar or tinker with the ram, I guess the real question should be; based on the application described above, am I going to see a difference between 667 and 1000?

Thanks for your help thus far.

 

[Ycon]

Another option would be a LGA771 Xeon workstation.

CPU power would not be a problem since you always have 2 sockets, if need arises.
BUT seeing that your software is heavily memory-dependent, the Quad-Channel memory controller might come in pretty handy.

The upgradability of the whole system will not be as good as with a desktop PC, but you can count on a higher stability and the price shouldnt be much higher (you can get 2 lower speed Xeon QCs which cost less than 1 QX6700).

 

[Mondoman]

One thing to consider is that water cooling is less robust than air cooling. The water-cooled system still needs plenty of internal fans for the other system components, but also has the risk of leaking. Leaking will either be just a mess, or could destroy some or all of your system. Water cooling will also likely require periodic maintenance to top off coolant, drain/flush/refill coolant, etc.
Surely these vendors have demo setups that would let you try out the systems to see which is better for your needs?

 

[mhw100]

Thank you all very much for your ideas and comments.

Here is quote from another person relative to this matter. Would you pros please see if you agree with this summary?

"...my quess is that the difference in overall throughput between 667-1000 will be on the order of 1, 2 or 3%. It's only a guess but I have spent nearly 40 years designing microprocessors and microprocessor systems."

"Memory is not holding your system back, at least not much to be concerned. 667MHz memory in dual channel runs at an effective 1334MHz, and at those speeds, memory is waiting for the CPU to catch up. So 1334 or 2000 is like talking about SATA 1.5 and SATA 3.0. The fastest hard drives (a RAID0 array of WD Raptors) can pump out just over 100MB/s. So 1.5GB/s or 3.0GB/s is moot. You can have 10 arrays and still not use up all 1.5GB/s, so going with 3.0GB/s will do nothing to improve your systems performance. The same applies to memory.”

When asked if this is true, then why do people even bother buying the higher priced 1000Mhz memory he responded with this;

“Well, its like anything in the enthusiast market - quantities drive prices down, and manufacturers need an edge to charge higher prices. Things like the Raptor drives and high speed memory help eek out those few extra points in 3DMark06, so those folks who are speed-obsessed will jump at the opportunity. There is, after all, a performance improvement (small as it may be) from using the higher speed memory. I just feel it’s not as important as other folks make it out to be. I can live with a 17,800 score in 3DMark06. 18,000, or 18,500 just didn’t justify me going for higher speed memory. Now, 20,000? THAT I would have considered!”

"The thing is that 667 and 1000 are numbers that determine the bandwidth of data between the processor chip and memory. If you looked in the fine print of the specs you'd probably learn that the latency of the two memories is the same or very nearly the same. Improved latency of memory will drop through to the bottom line in performance. But due to the way that today's systems are designed, they are not bandwidth limited so 667 vs 1000 is not very important."

 

[Scougs]

Your freind is correct if you assume stock clock speeds.

The reason for having faster memory is if you want to overclock. CPUs perform the best if the memory bandwidth matches or excedes the FSB bandwidth.

Since the FSB on the Core 2 is basically 128 bits wide at an actual clock speed of 266MHz (Intel multiplies this by 4 for marketing reasons based on technicalities.) and a dual channel memory setup is also 128 bits wide, optimal performance will be found with the actual clock speed of the memory matching or exceding the actual clock speed of the fsb.

Thus, the following are the matched speeds:

Stock
FSB 266MHz (Intel 1066MHz)
RAM 266MHz (DDR2-533)

Example overclock running DDR2-800 which would not work with DDR2-667 RAM
FSB 400MHz (Intel 1600MHz)
RAM 400MHz (DDR2-800)

If the multipliers are chosen such that the CPU speeds are close in both cases, the second system will perform better. I don't know how noticable it will be with your program, but I would wager that it would be noticable if it is as CPU intensive as you say it is.

 

[mhw100]

Thank you for this but I now see I mistakenly didn't provide you with all the information.

I just noticed that the FSB is 1066 on the unit with the 1000mhz ram and the FSB is 667 on the unit with the 667 ram.

All things other things being equal would you say I would notice a difference in this cpu intensive app between these units.

Sorry for being such a dweeb.

 

[Mondoman]

... talking about SATA 1.5 and SATA 3.0. The fastest hard drives (a RAID0 array of WD Raptors) can pump out just over 100MB/s. So 1.5GB/s or 3.0GB/s is moot. You can have 10 arrays and still not use up all 1.5GB/s, so going with 3.0GB/s will do nothing to improve your systems performance. ...

Just a note that your friend is off by a factor of 10 or so on the SATA, as their transfer rates are 1.5Mb/s (bits, not bytes) and 3.0Mb/s (bits, not bytes). Thus, Raptors are within a factor of 2 of maxing out 1.5Mb/s. His point remains valid, but is aging rapidly.

 

[Scougs]

I just noticed that the FSB is 1066 on the unit with the 1000mhz ram and the FSB is 667 on the unit with the 667 ram.

I don't think that can be true, because the only Core 2 quad only comes in 1066fsb as far as I know. The only 667fsb Core 2 systems are laptops or desktops based on mobile components if I'm not mistaken. The DDR2-1000 would definately be better.

 

[mhw100]

Sorry, you're right it is 1066. I had it wrong. I also feel a little better knowing that there isn't an appreciable difference in ram speeds of 667 to 1066.

One last question; is there a difference between SLI/EPP RAM versus RAM without these specifications. If so, would this make a difference between 667 and 1066 RAM speeds?

Thank you.

 

[Scougs]

As far as I know, SLI RAM is a special feature for AMD based systems. It is supposed to increase performance or something. I don't think that the SLI rating means anything other than the memroy has been verified to work with that feature. I am not speaking as an expert on this matter however so I could be wrong.

I'm guessing you mean ECC instead of EPP RAM. First of all, ECC RAM I believe has error protection of some sort. I don't know too much about it because typically only systems based on server grade hardware use it and possibly MACs, but I'm not sure about them. For your purposes, I don't think that you want ECC RAM. In fact, I'm not sure if it would even be compatible with your system. Again, I've never used ECC RAM so I could be wrong.

 

[Mondoman]

He does mean EPP. SLI/EPP are just different brand names (nVidia/Corsair) for storing extra configuration information, including voltage info, in the DIMM's SPD. Typically, this includes some factory-sanctioned overclocking settings. When plugged into a motherboard that recognizes and can make use of this extra info (so far, just certain nVidia chipset-based MBs), the MB can then use one of these extra configuration settings, enabling "automatic" factory-sanctioned overclocking.
So, the underlying memory DIMM has exactly the same performance as a non-SLI/EPP version; SLI/EPP is just an extended SPD system for those who don't want to set memory overclocking settings on their own.

 

[Scougs]

SLI/EPP is just an extended SPD system for those who don't want to set memory overclocking settings on their own.

What's the fun in not configuring you own overclocked settings?
Thanks for clarifying what SLI/EPP are. As I said I'm not an expert on that feature. Just trying to contribute as much as I can.

 

[mhw100]

A big thanks to everyone who helped me here. This seems like a terrific community for noobs like me to get aid from experienced vets.

 

[mhw100]

As a follow up, one of the machine's mobos has the FSB pushed up to 1270Mhz from the stock 1066. In the other machine the clock speed is the same at 3.2 Ghz but was bumped up by the mulitplier. If the clock speed of 3.2Ghz is the same in both cases, would is it accurate to say that there is a 20% boost in throughput or work or performance just becuase the FSB is raised up by that amount?

Thank you.

 

[Mondoman]

There's an increase in maximum memory throughput, but these CPUs don't seem to be too sensitive to that, so I wouldn't expect a big difference in overall performance. You'll just have to test them on apps that you use to see for yourself.