default header

Hardware

PC RAM

Moderator: JC Denton

PC RAM

Unread postby icycalm » 05 Jan 2010 21:38

I am a bit stumped in this area. What I DO know is that I need triple channel memory, because the X58 chipset requires it. From what I understand triple channel is the new thing on the block, and only Intel currently supports it, whilst AMD's chipsets work only with dual channel memory.

Now triple channel seems to mean you need to populate three slots on the motherboard, so you can go with either 3, 6, 12 or 24 GB of RAM -- no middle choices. In this case, and after scanning prices, etc., I think I'll go with 6 GB. There doesn't seem to be any reason to shell for anything more, and in fact I am inclined to think that I would have been just fine even with 3. However in that case I would have to use three 1 GB modules, which would have used up half the slots on the motherboard, so if I wanted to ever go above 6 I would have to throw the three modules away. That's why I am going with three 2 GB modules, so that I can then upgrade to 12 GB if needed without throwing anything away. My motherboard of choice maxes out at 24 GB, but yeah, lol, etc.

As for speed, I guess I might as well max out the motherboard and go for 2200 MHz. Problem is it seems hard to find, so I might have to go with something a bit slower. 2133 MHz modules seem relatively easy to find, so I'll probably get these.

I am still stumped on the whole timings issue. What I don't know is the relative importance of the timings compared to the speed rating. Because I have found modules that have very fast speeds but mediocre timings, and modules with mediocre speed but very good timings. I haven't seen anything that is top-rated on both counts -- though it has to be said I haven't searched very hard...

Let me make one thing clear: lower timings are better, correct? So, for example, 9-9-9-24 is WORSE than 7-7-7-20?
User avatar
icycalm
Hyperborean
 
Joined: 28 Mar 2006 00:08
Location: Tenerife, Canary Islands

Unread postby tackywoolhat » 05 Jan 2010 22:51

You are correct, timings are latencies, so the lower the better. And you're not likely to find memory that has both low timings and high clock speeds -- it's generally one or the other. I don't know exactly why.

Timings vs. clock speed (bandwidth) is a mainly theoretical issue. The benchmarks I've seen people run do not show a conclusive practical difference between memory run with fast clock speeds/slow timings vs. memory run with slow clock speeds/fast timings. Theoretically, fast timings should be better for games (which require quick access to a large amount of smaller chunks of data) and higher clock speeds should be better for applications (database software, Photoshop) that rely on tons of throughput and are less interested in jumping around.

It used to be that Intel processors tended to benefit more from low-latency memory, but that may have changed since I was last knee-deep in benchmarks trying to put together my own 'new' rig. I can't say which is better for them now.
tackywoolhat
 
Joined: 14 Nov 2009 00:16

Unread postby Nybble » 06 Jan 2010 17:49

Knowing what processor you got, make sure that the Voltage is 1.65V or less. Otherwise the processor will not work with the RAM.

I got 6 gigabytes as well (3 x 2GB). On the ASUS motherboard I got, you can add additional memory without another 3 sticks. You can add a fourth, fifth, and sixth stick one at a time. If I do get more memory (as Windows 7 uses about 1.5 GB at startup), I may just get another 3 sticks, as it is pretty cheap stateside.

Where are you buying your parts?
Nybble
 
Joined: 06 Jan 2010 02:03

Unread postby austere » 14 Jan 2010 20:34

tackywoolhat wrote:You are correct, timings are latencies, so the lower the better.


Timings are not exactly latencies, though you'd assume so after reading almost anything you find in tech magazines/product literature. This isn't helped by the fact that the first timing figure is called the CAS latency (CL). The definition of timing is precise, whereas the use of latency is usually vague. Latency is a measure of time duration as a general rule. With RAM the best way to define it is the average time taken from initial memory access to the first bit read/written.

The same SDRAM module running at a lower clock speed and the same voltage will allow you to use tighter timing, since the duration of each clock is longer. The latency of the module will remain constant, so you will sacrifice throughput needlessly. The only way to significantly lower the latency without losing throughput, is to use a higher voltage.

Let's demonstrate this by using icy's figures, but say the 9-9-9-24 timing is for DDR3-2000 and 7-7-7-20 for DDR3-1333. To get the time delay of each operation (Row access, Column access etc.), you have to divide the timing by the I/O clock speed. The I/O speed of DDR3-2000 is 1000MHz, since the 2000 figure is divided by two as data is read on either edge of the clock. This means that the time duration of each operation is 9ns-9ns-9ns-24ns. On the other hand, the I/O frequency of DDR3-1333 is 667Mhz, giving us 10.5ns-10.5ns-10.5ns-30ns. The latency of the DDR3-2000 RAM should be better than that of the DDR3-1333 RAM, even though its timing is worse.

For an example of confusion due to the lack of clear and consistent definitions, check this article: http://www.pcstats.com/articleview.cfm?articleID=873. Try not to cringe too much while reading their worthless analogies or you will develop wrinkles. if you check their gaming benchmarks and do some calculations like those above, you will see that the performance is improved with better latency not timing. Yet read their text and you'll see they're extremely confused.

tackywoolhat wrote:And you're not likely to find memory that has both low timings and high clock speeds -- it's generally one or the other. I don't know exactly why.
The reason why you won't find low timing-high speed memory is because the actual latency of DRAM modules were improving slowly over time in comparison to CPUs. This is because the latter relies on a completely planar process and considerably benefits from scaling. DRAMs on the other hand, have to scale both a transistor and a capacitor per bit. The latter takes up a lot of room, so they have to use a deep trench process to make DRAM viable. This is basically like folding a sheet and its much harder to scale than a transistor. You can't scale this the same way you scale a transistor since its dependent on process chemistry (and other reactions, interactions) rather than photo-lithography.

The only consistent benefit DRAM gets from scaling is shorter interconnects. Even this is negated by the fact that RAMs need to use lower voltages for higher resolution processes. Every other improvement has come from using different techniques, for example by putting the capacitor on top of the transistor. It's not trivial at all do something like this and a researcher can expect to spend more than half a decade of their lives making it viable. Most will fail. To remedy this fact, they created these architectures which increase the bandwidth of DRAM. L1, L2 and large L3 Caches on CPUs use SRAM and are much lower in latency so as long as you can fit the most accessed parts of your memory in the cache, DRAM latency will have little effect on your system performance. In theory. :)
User avatar
austere
 
Joined: 07 Dec 2009 22:50

Unread postby ontologist » 30 Jan 2010 23:26

Socket 7 motherboards, as used by the original Pentium CPU, were clocked at 66MHz (15 nanoseconds per clock tick.) The memory used back then, EDO RAM, typically had timings of 5-2-2-2, meaning that any memory access took five clock ticks to produce the first byte, and then produced three more bytes in the same 32-bit word, one every two ticks.

The first SDRAM (Synchronous Dynamic RAM) had timings of 8-1-1-1 or thereabouts; "Synchronous" referred to the *-1-1-1 timings. A couple of years later, SDRAM was clocked at 133MHz, with timings of 3-1-1-1 (known as CL3) and 2-1-1-1 (CL2). The original DDRAM (Double Data Rate SDRAM) would output a byte on the rising edge and on the falling edge of each clock cycle, so a typical rating of CL2.5 meant timings of 2½-½-½-½.

Dynamic RAM is a design that stores each bit of data on a tiny capacitor that is charged by opening a gate (a single transistor). Charge on capacitors will leak away; the contents of DRAM will be lost unless refreshed (read and rewritten) every few tens of microseconds, which obviously decreases the maximum possible bandwidth. However, the design is cheap. The only viable alternative was virtual memory on a magnetic hard disk, with a latency measured in milliseconds rather than nanoseconds.

Static RAM, which is faster but more expensive to manufacture, is used for the high-speed cache memory integrated on modern CPU dies. A few megabytes of cache memory are perfectly adequate; improving the capacity and price of DRAM are considered a lot more important than improving its speed.

I have difficulty understanding how the timings measured for SDRAM and DDRAM relate to the more complicated ones for DDR-2 and DDR-3; I suspect a degree of obfuscation. Nonetheless, the basic principle of a count of the clock ticks - a division of the clock frequency - still applies.

Now, the quad-core Phenom II is a competent replacement for the Core 2 Quad, avoiding the design problems that forced Intel to disable Hyper-Threading in the latter. The unique selling point of the Core i7 architecture is that the system memory controller has been integrated onto the CPU die, making it easier to clock it much faster than the motherboard chipset, thus ameliorating an obvious performance bottleneck.

This new System Memory Interface is described rather nicely in the datasheet for Intel's LGA-1156 Core i7 (PDF, 757kB); the feature list on page 7 of the datasheet (PDF, 866kB) confirms the details are similar for the LGA-1366 version.

The LGA-1366 Core i7 supports single, dual or triple-channel memory access, depending on how many of the three memory channels are occupied. If the channels contain different amounts of memory, it will work in three channels where it can, and use single-channel access for the overspill.

As a consequence of this direct connection, the CPU's I/O subunit is at the same voltage as the DDR-3, so excessive voltage can damage the CPU as well as the DIMMs. The official DDR-3 standards specify 1.5 volts; Intel have announced (PDF quoted above) a maximum safe value of 1.65V.

Officially, the memory controller operates at 1066MHz (PC3-8500) and at 1333MHz (PC3-10600). The only reason to consider faster DDR-3 is if you want to overclock. As this guide explains, the Core i7 is particularly elegant to overclock.

BCLK is a system-wide clock signal under motherboard control; every clock frequency in the computer is produced by multiplying BCLK. For example, the Core i7 920 (2666MHz) runs at 20xBCLK, and reduced power settings work by reducing that 20x multiplier. The memory controller lets you choose between 8xBCLK and 10xBCLK. To overclock the system, simply instruct the motherboard BIOS to increase BCLK.

For best performance, I recommend OCZ RAM. They select their best-performing RAM chips, attach a heatsink and boost the voltage, producing DIMMs that perform (reliably under a dedicated hardware tester, which is much more rigorous than Memtest86+) with much faster timings than are standard.

Their web pages load more quickly if you disable Flash.

All memory DIMMs contain a SPD chip - a ROM chip programmed with the optimal timings for that DIMM, to configure the BIOS automatically. The "Intel Extreme" and "NVidia SLI-ready" DIMMs have Intel's and NVidia's extensions to the SPD standard, that store several different overclocking profiles on the one chip; but I don't recommend them, because the timings they quote are the same as for OCZ's "Gold" Series. The timings given for their Platinum Series are noticeably better.

Their Blade, Flex XLC, Reaper and Platinum Series are the best ones; the only difference between them is the type of heatsink. The ones labelled "Low Voltage" were tested at 1.65V; the others were tested at 1.70V, above Intel's recommended safe maximum.

The bargain price of AMD's Phenom II caused a large upswing in sales of new computers over Christmas, which has led to a shortage of DDR-3. The price per gigabyte is about twice what I paid six months ago. Even though 12GB kits are just beginning to become available, I'd avoid them at the moment.

I've been told that 4.0GHz air-cooled isn't too difficult for a Core i7 (but check for yourself - don't take my word for it); this would mean setting BCLK to 200MHz, and a choice between 1600MHz (PC3-12800) and 2000MHz (PC3-16000) DDR-3. The best timings for OCZ DDR-3 are 6-6-6-24 for PC3-12800, and 9-9-9-30 for PC3-16000. The DDR-3 chips themselves are identical in both cases, but the PC3-16000 version's SPD chip has been programmed with much looser timings; otherwise it couldn't reach 2000MHz at all. You'd gain no DDR-3 speed advantage either way.

Then if you decide not to run the CPU at a 50% overclock, the SPD presets would be awful. You'd have to fiddle around with the BIOS, and change the timings back from 9-9-9-30 to 6-6-6-24 by hand.
ontologist
 
Joined: 27 Jan 2010 04:10

Unread postby icycalm » 26 Sep 2013 23:30

G.SKILL Achieves World's Fastest Quad Channel Memory Speed at DDR3 4072MHz
http://www.pcper.com/news/General-Tech/ ... R3-4072MHz

Jeremy Hellstrom wrote:Taipei, Taiwan – 18 September 2013 – No limit is too high for G.SKILL memory. In just a week after the official release of the new Intel Ivy Bridge-E Core i7 Extreme processors, G.SKILL memory is already testing the extreme limits of the Intel processors and broke the world record for fastest DDR3 yet again. This time a 16GB (4x4GB) G.SKILL TridentX memory kit is overclocked to a blistering DD3 4072MHz - the first instance of a quad-channel DDR3 memory kit to break the 4GHz barrier!

trix2.jpg


This astounding feat was made possible and achieved on the new Intel i7-4960X CPU and the ASUS Rampage IV Black Edition motherboard under LN2 extreme cooling.
User avatar
icycalm
Hyperborean
 
Joined: 28 Mar 2006 00:08
Location: Tenerife, Canary Islands


Return to Hardware

cron