6
\$\begingroup\$

I know that a DIMM is composed of a set of chips that contain control logic managing the decode and prefetching memory operations. According to a product specification, I found that newer RAM works at a high clock rate (> 1Ghz) that is comparable to some CPUs. And that's what made me wonder why only the CPU is equipped with a heat sink, and not also the DIMM, besides a certain high clock rate (and thus the amount of heat needing to be cooled).

\$\endgroup\$
  • 26
    \$\begingroup\$ Some DIMMs have heatsinks. Look at overclocker oriented DIMMs. \$\endgroup\$ – Lior Bilia Jun 17 '18 at 12:57
  • 2
    \$\begingroup\$ PC3-8500 does not (despite its common name of DDR3 1066, and despite many vendors suggesting otherwise in marketing materials) work at a clock rate > 1GHz. It has 1066M transfers per second, but because it's a Double Data Rate product that means that its clock rate is 533MHz and performs two operations per clock cycle (similarly to having a dual core processor). PC3-16000 products do operate at 1GHz, but seem to universally have heatsinks. \$\endgroup\$ – Jules Jun 17 '18 at 15:47
  • 4
    \$\begingroup\$ Umm… they are. Not all of them though, same as not all CPUs require heatsinks and you can find embedded low power ones without heatsinks. I suggest rephrasing your question, because it doesn't seem well-researched the way you posed it. \$\endgroup\$ – user1306322 Jun 17 '18 at 19:01
  • 3
    \$\begingroup\$ Why do you correlate Ghz with need for heatsinks instead of power consumption? I am not aware of any DIMM taking 250 watt. \$\endgroup\$ – TomTom Jun 17 '18 at 20:02
  • 2
    \$\begingroup\$ Yeah, but it is a little ignorant to go by frequency and totally ignore the thermal design specs. It is quite obvious that for example a TR4 socket CPU - which can use up to 250 watt in the newest iteration - will have a different cooling need than a DIMM that pulls - attention - around 1-2 watt, frequency or not. Frequency is not the only relevant element. At the end, it is about power consumption and thermal profile, and that is very differnt REGARDLESS of frequency. \$\endgroup\$ – TomTom Jun 17 '18 at 21:23
18
\$\begingroup\$

You're assuming that the power dissipation is directly related to the clock rate. That's true but there's more.

Suppose I have this chip A where only 10% of the chip area (die size) runs at the highest clock rate. Compared to a chip B of equal size where 100% of the circuits are running at the high clock rate, chip A would dissipate only about 1/10th of the power that chip B dissipates.

My point: not only the clock rate matters, also how much of the chip is actually running at that clock rate.

For a DRAM chips (PC DIMMs use DRAM) most of the area on the chip is DRAM cells (obviously) and these are run at a significantly lower speed than the external clock rate. The DRAM controller access the chips in parallel and in a sequence so that this lower speed is somewhat compensated for by parallelism.

On a CPU a much larger part of the circuits actually run on the maximum clock rate (depending on how busy the CPU is of course) so it is bound to dissipate a lot more power than a DRAM chip where only a small part of the chip is running very fast.

\$\endgroup\$
  • 9
    \$\begingroup\$ You can add another simple fact. the surface area of a DIMM is significant larger than the surface area of a CPU. Yes, it may not look so - but that is mostly because "the CPU" is mostly head spreading material, the acutal CPU is a lot smaller. \$\endgroup\$ – TomTom Jun 17 '18 at 16:06
  • 4
    \$\begingroup\$ It really is significantly lower. Most modern DRAM modules for consumer PCs refresh every 64ms so rather than operating in the GHz range, it nominally operates at a paltry 15.625 Hz. Of course, a heavily-loaded module will be quite taxed, but even then a CPU's built-in cache will reduce the need to repeatedly read from RAM. \$\endgroup\$ – forest Jun 18 '18 at 2:04
  • \$\begingroup\$ @forest, no as 1/64 ms = 15 Hz, for 15 kHz T is about 64 us. But 15 Hz is just the refresh rate, when accessed the DRAM cells will be accessed at a much higher frequency. The 15 Hz is only when idle (sitting there just keeping the RAM contents). \$\endgroup\$ – Bimpelrekkie Jun 18 '18 at 14:28
  • \$\begingroup\$ @Bimpelrekkie I said 15 Hz, not 15 kHz (15.625 where . is a decimal place). I said a loaded module will be more taxed. The majority of cells will be sampled every 64ms. \$\endgroup\$ – forest Jun 19 '18 at 2:59
13
\$\begingroup\$

DIMMs don't dissipate the same power a CPU does, so they don't need the same cooling. In addition, the power the memory and control chips do dissipate is much more spread out physically.

Power dissipation may be roughly proportional to clock rate, but that proportionality constant is quite different between a CPU and a memory. The CPU has many more transistors and gates switching at the clock transitions than the memory does.

Remember that for CMOS, by the time you get to current being roughly proportional to clock speed, the dominant current is charging and discharging all the little parasitic capacitors on the outputs of every gate. If you have fewer gates changing state, then there is lower current, which results in lower dissipation at the same clock rate.

\$\endgroup\$
6
\$\begingroup\$

You need a heatsink if your component produces more heat than it can dissipate through its own package. Heat is electricity converted to a change in temperature of some mass

Now, in a modern CPU, what uses electric energy is mainly the process of switching a transistor. Every single transistor switching costs energy, and the faster that switching has to happen increases the amount of energy per switching.

Now, for every clock cycle, your CPU does a lot complicated things like multiplying numbers, caclulating addresses, speculating what the next operation might compute before that actually happens, and so on. Those operations lead to a lot of transistors switching at once.

A DRAM chip (like the one on your DIMMs) is different in that there's no complex operations to do – it's just memory, which means that it basically has to switch about (word length)×(memory address bits) – so, really, less than 2000 transistors for a single chip (there's a bit of address and command decode overhead, but that's very "cute" compared to the complexity of a CPU). Sure, the things these transistors switch need more energy (because that charging and discharging relatively large capacitors, whose charge is the actual bit), but it's really very few transistors only.

Then, DRAM also needs to be periodically refreshed, but that happens every few milliseconds or so only, so only every couple million memory clock cycles – and hence doesn't contribute greatly to overall power consumption.

\$\endgroup\$
  • \$\begingroup\$ "Relatively large capacitors" in the nF range... \$\endgroup\$ – forest Jun 18 '18 at 2:06
  • \$\begingroup\$ I must admit I don't know how large these are. Probably multiple dozen picofarad! \$\endgroup\$ – Marcus Müller Jun 18 '18 at 10:59
2
\$\begingroup\$

Some types of DIMMs do have (and need) heatsinks. While the ones on gamer-oriented memory sticks are mostly for design/show reasons, there are e.g. FBDIMMs for servers which, due to their different architecture, require significantly more power (the last ones I used were roughly at 10W per stick) and thus need more cooling capacity than the bare plastic chip package can provide.DDR2 FBDIMM

\$\endgroup\$
  • \$\begingroup\$ "the ones on gamer-oriented memory sticks are mostly for design/show reasons" - I disagree. I needed to buy radiators for my PC memory few years ago, and it helped to prevent Windows BSoDs. Just anecdotal, I know,but still. \$\endgroup\$ – Mołot Jun 18 '18 at 10:55
  • \$\begingroup\$ @Mołot That's why I wrote "mostly".. there may be corner cases where memory sticks are overclocked hard enough (or misdesigned so much, or installed in a case with poor airflow) that they don't run stable without extra cooling anymore.. but the broad majority doesn't need that. I've just looked through a few Kingston gamer memory sticks, and they are all in the 1.5-2 watt area, which could easily be dissipated without the heatsink as well. \$\endgroup\$ – WooShell Jun 18 '18 at 14:15
  • \$\begingroup\$ To be honest, OC in gaming is no longer a "corner case", it's something manufacturers expect to happen. Also, I always do my best to buy the fastest memory my CPU can handle, and that makes a difference, I guess. \$\endgroup\$ – Mołot Jun 18 '18 at 14:23
0
\$\begingroup\$

DRAM is built from gates and capacitors. Think of glasses of water, some empty, some full. Pour some water in and periodically refill as it evaporates.

CPU cache is made of flip-flops. Think of faucets full blasting cold or warm water. You don't need refilling, but they do use a lot of water (energy).

Because of that difference DRAM chips usually don't need heatsinks (not much energy loss) but CPU does. Mind that CPU also does the calculations (another set of fire hoses) that contribute to the heat.

\$\endgroup\$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.