Sign in with
Sign up | Sign in
Your question

The trend increase Mhz rating, increase timing?

Last response: in Memory
Share
June 16, 2005 2:50:29 PM

I'm looking at memory specifications and it sure seems like Corsair, GeIL, and others just keep increasing their ratings (i.e. PC4400 DDR550) by increasing the timing delay.

GeIL Ultra-X Series dual channel DDR550
CAS 2 5-2-2 at 400Mhz (actual 200Mhz)
CAS 2.5 5-3-3 at 500Mhz (actual 250Mhz)
CAS 2.5 7-4-4 at 550Mhz (actual 275Mhz)

Corsair TWINX PC4400 DDR550
CAS 2.5 4-4-8 at 550Mhz (actual 275Mhz)

Corsair TWIN2X PC8000 XMS2 Extreme
5-4-4-9

Can't any memory manufacturer just claim higher performance by increasing the delay/latency? I just recently discovered that my inexpensive "Value" series PC3200 memory was over agressively timed which was causing errors, I adjusted the timing values to those of the more expensive PC4000 and viola, the "Value" series suddenly performs like a champ?

I'm beginning to seriously think the memory game is really just a big money scheme were overclockers such as myself are the real suckers.

What gives? Anyone else done some memory timing comparisons and come up with the same conclusion?

Rob.
a b } Memory
June 16, 2005 6:25:47 PM

Higher timings lower performance and increase stability. When you overclock RAM, the easiest way to give it just a little more room for higher clocks is to use higher latency values. It's quite common for the same module to be programed as "PC3200 CAS 2-2-2-5" and "PC4000 CAS 2.5-3-3-8" for example.

When you slowed your RAM you made it stable, that's all they're doing.

<font color=blue>Only a place as big as the internet could be home to a hero as big as Crashman!</font color=blue>
<font color=red>Only a place as big as the internet could be home to an ego as large as Crashman's!</font color=red>
June 16, 2005 7:55:30 PM

Right, but the actual memory bandwidth remains virtually unchanged as one increases the FSB one has to increase the timing. So the end result is a faster CPU, but access to main memory is still slow. So why pay more $$$ for higher rated memory? Anyone can bump the timing up on the "cheap" memory to meet whatever PCxxxx spec they want, the bandwidth will remain virtually the same.

I can see the move to DDR2, but that is a physical difference which permits higher bandwidth.
Related resources
Anonymous
a b } Memory
June 16, 2005 8:23:20 PM

The FSB speed will affect the bandwith much more than the timming. That is especially true when you run your cpu 1/1 with the FSB.
Just look at my system for exemple, its 2 years+ old and it still performing really nicely.
I squeezed so much performance by using DDR500 ram even if my timming are pretty slow.
By relaxing the timming a bite more my CPU can do close to 3.4ghz really easily all that from a 300$can cpu bought 2-2.5 years ago!
Now if I bought some of the new ram, lets say OCZ VX(if I could supply the voltage hehe) I could run at the same clock speed but with 2-2-2-5 timmings, that wouldnt help me THAT much. Sure it would be faster but nothing like the bump I saw running at 250mhz instead of 200...

Asus P4P800DX, P4C 2.6ghz@3.25ghz, 2X512 OCZ PC4000 3-4-4-8, Leadtek FX5900 w/ FX5950U bios@500/1000, 2X30gig Raid0
a b } Memory
June 17, 2005 2:57:47 AM

First of all, not all memory can be substantially overclocked simply by raising the voltage or timings. Second of all, timings affect response time, not bandwidth.

<font color=blue>Only a place as big as the internet could be home to a hero as big as Crashman!</font color=blue>
<font color=red>Only a place as big as the internet could be home to an ego as large as Crashman's!</font color=red>
June 17, 2005 3:03:02 AM

Response time is integral part of the end result which is performance. Bandwidth is measured as data over time hence MB/sec. If memory has to wait 3 clock cycles before it is ready to give/take more data then that will adversely affect bandwidth (how much data can be transmitted over a give time).

Rob.
a b } Memory
June 17, 2005 3:07:21 AM

That depends completely on file size. Smaller files are more greatly affected than larger files.

<font color=blue>Only a place as big as the internet could be home to a hero as big as Crashman!</font color=blue>
<font color=red>Only a place as big as the internet could be home to an ego as large as Crashman's!</font color=red>
June 17, 2005 6:02:47 AM

How about games? I do know that when I set my ram to CL3 I get some slight dips in FPS when entering a room. When I set it to CL2, That dip is less dramatic if at all.



"Battling Gimps and Dimbulbs HERE at THGC"
a b } Memory
June 17, 2005 7:07:31 AM

Games use tons of tiny files.

<font color=blue>Only a place as big as the internet could be home to a hero as big as Crashman!</font color=blue>
<font color=red>Only a place as big as the internet could be home to an ego as large as Crashman's!</font color=red>
June 17, 2005 2:58:24 PM

Huh?? File size?? Files being loaded into memory via DMA (which requires very few CPU cycles) is limited by the Hard Drive or CD or DVD drive -- it is not that relevant since those devices are in magnitudes slower than memory. Where CL and timing counts the most is when the CPU wants to read/write to memory. If the memory address it needs is not found in any of the CPU caches, then it makes a trip on the FSB to go get it from main memory.

Like I said earlier, playing with timing is really just to make the memory work with a fasted FSB, it doesn't really benefit real world performance UNLESS the memory can remain stable with higher FSB and at lower timing delays.
June 17, 2005 5:26:45 PM

If a file is in memory then lower latency will help when the CPU requests the block of memory containing the first chunk of the file - kind of like seek time on a hard disk. The rest of the file will usually be stored sequentially after the first chunk, so latency has very little impact at all.

So as crash says, lots of little files (Lots of seeks) will benefit from lower latency, but anything larger will benefit more from the bandwidth increase of a higher clock, as it can (after finding the start of what it's after) grab the rest from memory in less time.

---
<font color=red>"Life is <i>not</i> like a box of chocolates. It's more like a jar of jalapeńos - what you do today might burn your a<b></b>ss tommorrow."
June 17, 2005 5:30:51 PM

Here's a question for ya crash - ever had memory that is stable at CL2.5 but not at CL3? Had that yesterday... I Tried installing an extra 512Mb (single stick) in a buddy's PC - nforce2, and he had 512Mb already (2x 256Mb sticks). I was hoping I could get DC going with 1x512Mb in one channel, and 2x256 in the other.

I had no joy, but after giving up for the evening and trying to return it to former settings, I just could not get it stable despite increasing timings and voltage.... Until I dropped CAS to 2.5 from 3, and then it was fine! The stuff was at 3 to begin with, which is why I started off there.

Seemed weird to me... This thread just reminded me about it. :eek:  Ever seen similar?

---
<font color=red>"Life is <i>not</i> like a box of chocolates. It's more like a jar of jalapeńos - what you do today might burn your a<b></b>ss tommorrow."
a b } Memory
June 17, 2005 5:45:26 PM

Thank you ChipDeath, you have such a great way of putting the confusing garbage in my mind into words!

<font color=blue>Only a place as big as the internet could be home to a hero as big as Crashman!</font color=blue>
<font color=red>Only a place as big as the internet could be home to an ego as large as Crashman's!</font color=red>
a b } Memory
June 17, 2005 6:26:18 PM

No, I've never seen that!

<font color=blue>Only a place as big as the internet could be home to a hero as big as Crashman!</font color=blue>
<font color=red>Only a place as big as the internet could be home to an ego as large as Crashman's!</font color=red>
June 17, 2005 6:40:42 PM

"Bandwidth" is one of the most confusing terms used in data communications/networking.

When we talk about the bandwidth of a communication link, we normally refer to the number of bits per second that can be transmitted on the link. We may say the bandwidth of an ethernet to be 10Mbps, or a memory bus to be 3200MB/s.

A useful distinction can be made about the bandwidth that is available on a link and the number of bits per second that we can actually transmit over the link in practice. THROUGHPUT is the term used to refer to the measured performance of a system.

Thus, because of various inefficiencies of implementation, a pair of nodes connected by a data bus or link with a bandwidth of 10Mbps (or 3200MB/s) might actually only achieve a THROUGHPUT of only 2Mbps (or 640MB/s).

Response time is an integral part of THROUGHPUT, but is not directly related to the bandwidth of the system (how much data is capable of being sent) as Crashman said above.

Quote:
I'm beginning to seriously think the memory game is really just a big money scheme were overclockers such as myself are the real suckers

For the most part, I could not agree more. Bunch of knuckleheads!

<font color=green>*****
"Memory with lifetime warranty? So, whose lifetime is that?"
<A HREF="http://www.brentcrowley.com/" target="_new">homepage</A>
<font color=red>AIM BrentUnitedMem<font color=white>
June 19, 2005 1:01:37 AM

Check these timings and bandwidth out. First OC on this rig. <A HREF="http://forums.extremeoverclocking.com/showthread.php?t=..." target="_new">Linky</A>

<font color=green>NED FLANDERS FOR PRESIDENT</font color=green> Its justa nother gansta PARTY!
Intel P4 Extreme(3.73)@<font color=green>5.6Ghz</font color=green>
Asus P5AD2-E-Premium
Crucial Ballistix DDR2 667@DDR2 855
June 19, 2005 5:53:24 PM

That is the image link, I didnt know that they had it setup so that you couldnt see pics unless you were registered.

<font color=green>NED FLANDERS FOR PRESIDENT</font color=green> Its justa nother gansta PARTY!
Intel P4 Extreme(3.73)@<font color=green>5.6Ghz</font color=green>
Asus P5AD2-E-Premium
Crucial Ballistix DDR2 667@DDR2 855
June 20, 2005 4:26:39 PM

To a degree you're right...but if you can clock it fast enough, the latency (in terms of nanoseconds) goes down (even if you raise CL, Trrd, Trcd, etc. slightly), which is what is important.

ns wise, DDR600 3-3-3-8 has the exact same latency as DDR400 2-2-2-5, but a MUCH higher bandwidth to the CPU (not to mention a faster HTT bus, if things are done correctly). If bandwidth weren't important and latency was the only factor, DDR550 1.5-2-2-0 would win, every time. It's not like that, though (although I'd be delighted if I had RAM that performed like that!), DDR640 at 2.5-4-3-0 is faster in most applications because of the added bandwidth. Of course, both these examples are at the extreme limits of DDR and not exactly consumer-level numbers, but they are real numbers.

Maxtor disgraces the six letters that make Matrox.
June 20, 2005 5:48:43 PM

Where can one find DDR640? I've only ever been able to locate DDR550.

Bandwidth is not the only game in town, as has been stated -- throughput is the name of the game. One may have super high bandwidth numbers, but if the CPU has to wait 2X, 3X, 4X as many clock cycles before it can get the data, then that will translate to stutters in games.
June 20, 2005 7:18:59 PM

Ever heard of overclocking RAM???? At least two vendors sell DDR600 and OCZ once sold DDR625 (still might, but I doubt it). Anyway, that's not the point.

I'm fully aware of throughput, which is VERY dependant on response time. However, response time, with 'higher' latencies might not be what you think it is. These higher speed memories run at the high speeds with high latency and DO perform better than low speed, low latency, regardless of what your little theory tells you.

Latency as we know it is in clock cycles, but we need to convert it into nanoseconds.

Let's say we have DDR400 at CL2 (just to simplify things, we won't get into the other latencies, but some have *generally* the same principal; others, not at all) and DDR500 at CL2.5 and DDR600 at CL3. The DDR400 will have bandwidth of 3.2GB/s, the DDR500 will have bandwidth of 4.0GB/s and the DDR600 will have 4.8GB/s. Now, let's add latency.

DDR400:
runs at 200MHz, or 200,000,000 clocks per second, therefore each clock takes 1/200,000,000 of a second, or 5ns. CL2 means that it will take 10ns (2 clock cycles * 5ns each) for the access, so take that much off the bandwidth and you get the throughput ((200,000,000-2*memory accesses per second) / 200,000,000 * 3.2GB/s [keep in mind there are MANY more types of latency involved, which is why throughput is often only 80% of bandwidth]). This still example equals about 3.0GB/s of throughput (with other latencies considered).

DDR500:
runs at 250MHz, or 250,000,000 clocks per second, therefore each clock takes 1/250,000,000 of a second, or 4ns. CL2.5 means that it will take 10ns (2.5 clock cycles * 4ns each) for the access, so take that much off the bandwidth and you get the throughput ((250,000,000-2.5*memory accesses per second) / 250,000,000 * 4.0GB/s). This example still equals about 3.7GB/s of throughput (with other latencies considered).

DDR600:
runs at 300MHz, or 300,000,000 clocks per second, therefore each clock takes 1/300,000,000 of a second, or 3.3ns. CL3 means that it will take 10ns (3 clock cycles * 3.3ns each) for the access, so take that much off the bandwidth and you get the throughput ((300,000,000-3*memory accesses per second) / 300,000,000 * 4.8GB/s). This example still equals about 4.4GB/s of throughput (with other latencies considered).

Now you tell me, which one is the fastest...heck, let's have DDR600 with CL5. You can run through the rest of the latencies yourself if you want, but they really don't take away a huge amount of the available bandwidth. Also, remember that some applications REALLY like bandwidth, other like lower latency to the memory...it all doesn't fall into a nice, clean number of throughput that signifies the end performance. Heck, the number Sandra will return to you under its bandwidth test in actually throughput, and we all know how much those mean for performance.

Maxtor disgraces the six letters that make Matrox.
June 20, 2005 7:23:41 PM

Also, the stutters in games are *NOT* because of memory latency!! There is a lot more to the performance than throughput, but let's be honest, the game accesses the memory so often and so predictably that upping the latecny a couple of clock cycles will not create stutters, but rather slow things down overall (not a whole lot, either).

Maxtor disgraces the six letters that make Matrox.
!