What makes CPU cache memory so much faster than main memory? I can see some benefit in a tiered cache system. It makes sense that a smaller cache is faster to search. But there must be more to it.
2What do you mean by main memory?– kedoskaMar 31, 2014 at 0:12
1@marcocs edited for clarity. I'm referring to CPU cache memory and main system memory, i.e. the removable dimms on most motherboards.– ConditionRacerMar 31, 2014 at 0:25
4What every programmer should know about memory– assyliasMar 31, 2014 at 21:09
2@ratchetfreak Uh...OP is pretty clearly asking for a proximate cause, not a final cause.– Kyle StrandMar 31, 2014 at 22:30
1Wow I didn't think I would ever use my computer engineering degree. The accepted answer isn't really an answer. It's because CPU cache operates at a much higher clock rate (the CPU clock rate, around 4GHz), while main memory operates at the bus clock rate (around 1600MHz). Not only that, but the CPU cache can read in 4 clock cycles, but system RAM might take 100 system clock cycles. Also, system RAM has contention with other peripherals like the disk or network, while CPU cache is dedicated. The longer the wire, the slower the clock rate it can operate at. tinyurl.com/zkdsstd– ChloeFeb 10, 2016 at 19:40
In the case of a CPU cache, it is faster because it's on the same die as the processor. In other words, the requested data doesn't have to be bussed over to the processor; it's already there.
In the case of the cache on a hard drive, it's faster because it's in solid state memory, and not still on the rotating platters.
In the case of the cache on a web site, it's faster because the data has already been retrieved from the database (which, in some cases, could be located anywhere in the world).
So it's about locality, mostly. Cache eliminates the data transfer step.
Locality is a fancy way of saying data that is "close together," either in time or space. Caching with a smaller, faster (but generally more expensive) memory works because typically a relatively small amount of the overall data is the data that is being accessed the most often.
Cache (Computing) on Wikipedia
30I think this is a good answer. However, it may be worth adding that there is also a cost hierarchy to cache: network storage is cheaper than a local hard drive, which is cheaper than RAM, which is much cheaper than on-CPU cache. If it wasn't for this we'd just make CPUs with 120GB of memory on-chip, and be done with it. Mar 31, 2014 at 0:26
1@ConditionRacer: Same reasons. L1 is faster, but more expensive, so they give you a little bit of very expensive L1, and more of the less expensive, but slower L2. See also stackoverflow.com/questions/3699582/… Mar 31, 2014 at 0:29
9"Why is L1 faster than L2" - the question becomes off topic for Programmers.SE, and should be asked on Electronics.SE– mattnzMar 31, 2014 at 0:42
5@ConditionRacer IIRC, it's to do with the technology used to implement the memory. I think L2 is on-chip DRAM, whereas L1 is SRAM or something like that; much more expensive, but much faster. It's over a decade since I worked in CPU design… Mar 31, 2014 at 1:24
4@CharlesE.Grant : while the cost hierarchy is a good example, there is another reason we don't have 120 GB of on-chip cache: addressing larger memory needs larger addresses, so either larger instructions or more CPU cycles. So to be the fastest, the number of registers needs to remain relatively small. Of course, one could have on-chip memory besides the registers, but the fastest tier has to remain small, otherwise it would not be that fast anymore.– vszMar 31, 2014 at 6:11
It is faster because both it is closer and because it is SRAM not DRAM.
SRAM is and can be considerably faster than DRAM the values are kept statically (the S in SRAM) so they don't have to be refreshed which takes away cycles. DRAM is dynamic, like tiny rechargeable batteries, you have to regularly recharge the ones so they don't drain away and become zeros. This steals cycle time in addition to how you have to access the bits, etc.
Being on the same die as or nearer the processor reduces the round trip, both L1 and L2 are faster than DRAM from an access perspective.
SRAM is faster to access than DRAM taken apples to apples, and the caches are usually on chip or closer or on faster busses than the DRAM making the access time faster as well.
1SRAM is also more power-hungry per bit, so you'd have to add radiators to it, and anything battery-powered would suffer.– 9000Mar 31, 2014 at 7:28
1Not to mention the fact that making SRAM bigger makes it WAY slower.– DarkhoggMar 31, 2014 at 11:30
agreed, more power, and slower relative to other sram. I think it is 4 transistors per bit for sram and 1 per bit for dram. dram relies on the capacitance in the fet, where sram relies on good old fashioned logic. so you have to power all of those transistors all the time and not just pump the charge periodically. Mar 31, 2014 at 13:46
4SRAM in caches generally is 6T, not 4T. Also, SRAM (certainly the 6T variant) is more energy efficient. The real power draw in DRAM is the refresh cycle, and SRAM just doesn't need that. SRAM transistors draw power when switching, the DRAM capacitor leaks all the time. If you replace the leaky cap of DRAM, you end up with EEPROM but that's not fast: if it's hard for the charge to leak out, it's also hard to get it out the normal way.– MSaltersMar 31, 2014 at 23:43
@dwelch CMOS FETs don't drain power if they are in open or closed state, so the argument about power consumption is invalid. Moreover, it's the CMOS logic switching, which makes CPUs drain most of their power - the current peaks in the intermediate state (non-1 and non-0), so the situation is inverse to what you say :)– RuslanApr 1, 2014 at 10:34
One thing that should be mentioned explicitly is the impact of the speed of light. In this video Grace Hopper shows a piece of wire about a foot long, which is how far an electrical signal can travel in one nanosecond*. If a CPU is operating at 3GHz, then that implies a distance of 4" per clock cycle. This is a hard physical limit on memory access speeds. This is a large part of why being close to CPU (as L1 cache is), allows memory to be faster.
EDIT *actually how far light can travel in a vacuum, the distance through copper/silicon is less.
2Hard limit for copper. This is a good point though, I didn't consider how ridiculously fast modern processors are. Mar 31, 2014 at 5:00
@ConditionRacer I just added a correction, 11.8 inches is light in a vacuum. Mar 31, 2014 at 5:04
2You need to halve that number since you need a round-trip to retrieve memory. So it's just 5cm not 10cm per cycle of latency in vacuum. Mar 31, 2014 at 9:53
2The speed of light is why the main memory bus clock stopped getting faster some years ago (2000ish?) It's not a direct effect -- it has to do with CMOS logic signals not being perfect square waves anymore at bus frequencies -- but c shows up prominently in the math.– zwolMar 31, 2014 at 17:10
Other answers already covered all the relevant bits: locality (and the associated data transfer cost, bus width and clock, and so on); speed of light (again, associated to transfer costs and bus width and throughput); different memory technology (SRAM vs.DRAM). All of this seen in the light of cost/performance balance.
One bit that was left out and it's just mentioned in Darkhogg comment: larger caches have better hit rates but longer latency. Multiple levels of cache where introduced also to address this tradeoff.
There is an excellent question and answer on this point on electronics SE
From the answers, it seems to me that a point to be highlighted is: the logic which performs all the required operations for a cache read is not that simple (especially if the cache is set-associative, like most caches today). It requires gates, and logic. So, even if we rule out cost and die space
If someone would try to implement a ridiculously large L1 cache, the logic which performs all the required operations for a cache read would also become large. At some point, the propagation delay through all this logic would be too long and the operations which had taken just a single clock cycle beforehand would have to be split into several clock cycles. This will rise the latency.
There are a lot of good points raised in the other answers, but one factor appears to be missing: address decoding latency.
The following is a vast oversimplification of how memory address decoding works, but it gives a good idea of why large DRAM chips ar generally quite slow.
When the processor needs to access memory, it sends a command to the memory chip to select the specific word it wants to use. This command is called a Column Address Select (we'll ignore row addresses for now). The memory chip now has to activate the column requested, which it does by sending the address down a cascade of logic gates to make a single write that connects to all the cells in the column. Depending on how it's implemented, there will be a certain amount of delay for each bit of address until the result comes out the other end. This is called the CAS latency of the memory. Because those bits have to be examined sequentially, this process takes a lot longer than a processor cycle (which usually has only a few transistors in sequence to wait for). It also takes a lot longer than a bus cycle (which is usually a few times slower than a processor cycle). A CAS command on a typical memory chip is likely to take on the order of 5ns (IIRC - it's been a while since I looked at timings), which is more than an order of magnitude slower than a, processor cycle.
Fortunately, we break addresses into three parts (column, row, and bank) which allows each part to be smaller and process those parts concurrently, otherwise the latency would be even longer.
Processor cache, however, does not have this problem. Not only is it much smaller, so address translation is an easier job, it actually doesn't need to translate more than a small fragment of the address (in some variants, none of it at all) because it is associative. That means that along side each cached line of memory, there are extra memory cells that store part (or all) of the address. Obviously this makes the cache even more expensive, but it means that all of the cells can be queried to see whether they have the particular line of memory we want simultaneously, and then the only one (hopefully) that has the right data will dump it onto a bus that connects the entire memory to the main processor core. This happens in less than a cycle, because it is much simpler.
One of the philosophies I studied was the obtain-maximum-throughput-in-minimum- hardware movement when we talk about any cache based memory, be it CPU cache, buffer cache or memory cache for that purpose. The basic motive is achieved when there is the least or no hardware movement for retrieving/reading/writing data and the operation is completed faster.
The data transfers from disk -> main memory (RAM)(temporary storage) -> CPU cache (smaller temporary storage near the CPU for frequently accessed data) -> CPU (processing).
The CPU cache is a smaller, faster memory space which stores copies of the data from the most recently used main memory locations.
The buffer cache is a main memory area which stores copies of the data from the most recently used disk locations.
The browser cache is directory or similar space which stores copies of the data from the most recently visited websites by users.
Reference: How Computer Memory Works
"operating systems philosophy is..." -- is this only your opinion or you can back it up somehow?– gnatMar 31, 2014 at 10:16
Was taught about this while studying Unix by Maurice Bach. Unfortunately no supportive documentation. Edited.. Mar 31, 2014 at 10:47
5howstuffworks is notoriously unreliable for getting reliable technical information. They're nice for gross simplifications that the average Joe and Jane can understand, but that's about it.– jwentingMar 31, 2014 at 11:27