I have just bought an Intel core duo (conroe) based system for my friend and I was doing various bechmarks and I noticed the low memory bandwith efficiency. Specs first: Intel Core 2 DUO E6300 "LGA775 Conroe" 1.86GHz (1066FSB) Asus P5B (Socket 775) PCI-Express DDR2 Motherboard (intel P965) Corsair 2GB DDR2 XMS2-6400C4 TwinX (2x1GB) (800MHz) Western Digital Caviar SE16 250GB 2500KS SATA-II 16MB Cache Leadtek GeForce 7900 GT Extreme 256MB Enermax Liberty 500W ELT500AWT ATX2.2 Modular SLI Compliant PSU Microsoft Windows XP Professional x64 Edition inc. SP2 In the sandra 2007 memory bandwith benchmark it says that it gets: int- 4839 Mb/s float - 4847 Mb/s Maximum bandwidth possible = 8512Mb/s Current banwidth efficiency = 57% This is using a FSB : DRAM ratio of 4:5, this has set the memory frequency to 333MHz. If I got into the BIOS and change the DRAM setting from auto to 800Mhz to give me a FSB : DRAM ratio of 2:3 and a memory frequency of 400MHz, it pushes the banwidth efficiency up a measly 2%. Is there any other way to get the efficiency higher? For comparison my system achieves an efficiency of 91%. Edit - I have just learned that this ram when running at 800Mhz can have command rate of 1 (CMD 1 or 1T) but it is set to 2T in the bios. If I changes it to 1T would it make a differecnce?
First piece of advice I can give you... Don't use XP Pro 64-bit edition. Program incompatibilty and unstable performance is NOT worth the minor processing power upgrade. :/
All the programs are installed and I have seen no evidence of any compatability issuses. Most programs have 64 bit versions now, even things like Deamon tools.
Yea, I have XP x64 and most things are compatable. The only incompatability that bugs me is Abit not making uGuru for x64. Oh well, back on topic. Basically, Speed, command rate, and latencies make the differences. I'm not positive but I also believe it is in that order. Raise the speed as high as it will go stably and then try to lower that commmand rate. Then lower latencies.
CMD (Command Rate or Command Per Clock--CPC) will reduce the delay in the RAM refreshing, which can really help performance. I'm not sure if value designated RAM can handle it, but if you have high performance stuff, enable it. I'm not sure about efficiency, but do realize that nothing will be 100% efficient.
Well yea, I don't know what that would do to efficiency but I know it will help speed and thats all that really matters.
Are you overclocking your proc?.... I bet you are. Despite what the guy above said, I'd lower your memory timings first and then raise your clock speed. With that amount of memory, I'd concentrate on CAS and CMD timing.
Actually, the command rate is pretty significant, but some memory, typically of the value oriented stuff, doesn't handle it as well.
Well but it is high quality Corsair RAM. And, as I said, I wasn't sure whether or not the speed was more important than the latency.
There's nothing that says speed is better than latency or other way around. The thing with memory is that it interrelates with the proc and other peripheral devices. It's hard to tell where your bottlenecks are, which is why you have to continually test your overclocks to see what your best option is. Sometimes increasing the clock speed is better, sometimes lowering the latency is better - my method is usually this: 1) Overclock mem to highest clock speed I can get at default timing and test 2) Set mem to minimum latencies I can get and test 3) Set mem to minimum latencies I can get and increase clock speed as high as I can. Generally, the best result comes from the third method, but you should always compare to eachother. And you can't just rely on one benchmark software. You should test across multiple applications and "real-life" experiences.
Cheers for the help guys, I will give that all a try. One other question. I take it that the maximum bandwidth possible is what the mobo supports and its up to the RAM to utilise as much of that bandwith as it can, which in turn gives its efficiency?
Bandwidth and FSB interrelate. You should think of memory this way: You are at a fastfood restaurant. You go to the counter to order some fries. The guy working there takes your order, turns around to find the fries, gets the fries, and continues to turn to give you your fries (making a full circle). Him turning all the way around is a "clock" how many times he turns to get fries per second is his clock rate (measured in Hz - 1 time around = 1 Hz). If there is a line of people placing orders and he's not matching his turning to keep up with the orders coming in, then there will be a long line of people, getting longer and longer. There could also be a long line if he's given an order, but takes too long to find the fries (latency, which deals with your timings). The idea is to find the fries as fast as possible and take as many orders as you can. Now this was just a simple example and there are some flaws in my explanation, but I don't think it's too bad ad-hoc. Maximizing your bandwidth allows you to make more orders in at once. Minimizing your timings/latencies makes finding the information faster.
