Overclocking: all about overclocking processors v.5

Belomorkanal
well, she left me for 13k, so to speak. There is an option to take another one and also for 13k the same for cross. http://market.yandex.ru/product/10553644/spec?hid=91031&track=char
In general, I was in a hurry with the vidyashka. But it seems like everything is enough for now. So far, I struggle with a decrease in sponginess with a stable acceleration of 4.6 per constant. Above, so far there seems to be no sense, and there is no desire to give an extra one either. There is a desire to reset the valves to minimum. 1,420 while linpak passes. But now the weird thing has gone. I didn’t check it before, of course, but Win7x64, when checking the ratings at the very end, gives an assessment error. Moreover, both linpak and prime are happy with everything. But, I just noticed that it gives an error at the moment when the frequency percentage drops to 1.4 in connection with this cool and quit. The same test is like a friend of a procyclist with a spruce. I started it at 1.55V at 5GHz (of course I gave it too much, I'm dabbling for now), all the tests are also great, but after the 15th run, Aida shows 69C. The dropsy copes, even the valves are only up to 1500 revolutions, but it is better not to keep such springs permanently. I have it 5 GHz and holds it with less stress, I just wanted to see how much the temperatures would rise when both the frequencies and the voltage were too high. At 5.2 I will no longer risk driving it on this coolant.
A question for connoisseurs, tk. he himself did not bother with memory. Now I have left the memory in the drain (1866 at 1.5V 9-9-9-27). With an increase to 2133 and a voltage of up to 1.68 (below the memory tests), what are the optimal timings to set? By default, she drew me very not very beautiful numbers at this frequency - 14-13-13-34. The memory chips are micron-based. http://market.yandex.ru/product/7843017/spec?hid=191211&track=char This memory had to be chosen, because I was able to pull it out from the supplier at 2.2k rubles per die. Yes, and there were no Mushkins with g-skills.
-Sergey Viktorovich-
There is an option, of course, to send the r9 280x back. I have a great relationship with this supplier. She will take it for a reason - I didn’t like it, it seemed rather weak. But then you have to sit and wait on the old 560s. By the way, which would be better than R9 290 or GTX 970? I have not monitored tests for a long time. For work more with server systems now I come across, or with a budget office ...)))

7Rain7
I have 2 4 GB sticks on Sosnumgov chips, the frequency is 2133, timings are 10-11-10-27, I haven't tried to set them lower, stability is more expensive than a negligible gain of 0.1%. If the memory (of my brain) serves me correctly, the napruga is automatically set to 1.65, and maybe higher, but definitely not lower than 1.65.
Not a single error in the tests. For that, the secondary timings are simply horsey (compared to 1600 9-9-9-24), but this does not prevent her from getting a lot of numbers in tests (compared to 1600 9-9-9-24).

PS_ I still do not understand why Intel claims that the maximum supported memory frequency is only 1600 MHz? And then he declares that their controller can handle the bandwidth of 25 Gb / s.
1600 MHz and 25 Gb / s well, nonsense. Or maybe I'm such a fool?
At least my memory works great at 2133 10-11-10-27. And if you take into account 25 Gb / s, then this is not the limit.

Belomorkanal
I tried now to put on 10-11-10-27 at 2133. Loaded, but did not run the tests yet. Now, before going to bed, I'll check for stability when overclocking the processor at 4.8 1.45V. He already passed the principle of 10 runs at such a low voltage today. I'll put it on for a couple of hours with the CPU smart protector turned on)) So that it won't be sad in the morning.
All the best. Let's see what I can squeeze out of this protsik. Most interesting, tk. for 17 years of baking, this is my first AMD-shnik. It seems that I am not experiencing anything with him yet. And there is no disappointment with such an excellent price. Even rather, I'm still surprised that I don't see the difference. Until the next generations of prots I will live with him, and then we'll see what and with what I will use. Consider that the entire forced platform change cost me 8k per percent. Although i7 is certainly a pity ((I still take revenge on my wife, as the child falls asleep))

7Rain7
Everything will be fine, you'll see.

White channel
1600 MHz and 25 Gb / s well, nonsense 1600 MHz
- effective frequency
64bit - channel width 2-channel
operation mode
1600 MHz * 8bytes (64bit / 8) * 2 channels = 25600MB / s / 1024 = 25GB / s

And by the way, yes, most sub timings depend little on frequency.

Standard bin for Samsung (D) 9-11-10 at 1.5v. 1.65 should be ala 8-10-10, although trcd 10 may not pull. tras can be safely set to 24 and 21.

Belomorkanal
And about the memory of Intel, I also did not really get it. The controller is kind of like yes, it should only be friends with 1600. But in fact, in 2133, it stably passes. But, not all dies are eaten with such frequency. Now I'll look, the other day I was just reading an article when I was choosing memory. I gave it to my old brother and mother. He had something obscene in his shop assembly there.
Here is the reviewer http://www.thg.ru/mainboard/vybor_komplekta_pamyati_na_16_gb_test/index.html
By the way, I wonder who, there are G.Skill Ripjaws X F3-14900CL10D-16GBXL and Mushkin Redline Ridgeback 997121R
Yes, and the tests are A10 4770k and " ", here it is littered, somehow I did not feel respect for the A10 percent. By the way, based on the test for 2133, I can try 9-10-9-27. And on 1866, cut down xmp and set it by hand to 8-8-8-24

DartMaul
Oh yes, the bandwidth is indicated at 2 channels. Thank you, Your Majesty.
It turns out that I exceed it by as much as 8 GB / s. This is real?
And AMD is still weirder. How then do they fit into 21 GB / s if they officially support 1866 MHz? Is the bus width less than that of Intel controllers? It turns out that the same memory bar at the same frequency, on Intel and AMD's controller, shows different results in terms of the number of reads and writes per unit of time? It turns out that AMD needs 1.5 times more frequency to achieve Intel results? As well as for the cores.
Your Majesty, please explain to the fools. Bring a bright torch of knowledge into our humble dark barn, aka forum.

Belomorkanal
By the way, 9-10-9-27 at the start of the bsod at 2133 issued. True, the voltage left 1.68. 9-10-10-27 at 1.7V now delivered. Let's see how the test goes. OK. I wonder here, but I'm sleeping))

7Rain7
I would not chase too much for timings and left it like mine, 10-11-10-27. Frequency will always be prioritized over low timings.

White
channel The channel width of all DDR memory is 64 bits.
What is indicated in the PC12800 format is the maximum theoretically possible bandwidth at a given frequency. Technical limitation of the tire itself. The memory controller is responsible not only for the memory operation, but also for the algorithm of this work. It is different for all controllers. If you pay attention to the timings, you will see that the timings (not the values, but the timings themselves) are different on different platforms. This indicates the use of different algorithms. I have never been interested in the nuances of memory work on AMD platforms, I cannot comment.
As for overclocking: to date, the record for the frequency of DDR3 memory is set at around 2310 MHz (DDR3-4620 MHz effective). That is, more than 36GB / s per channel. Everything is possible)
Spoilerhttp: //hwbot.org/submission/2573743_sofos1990_memory_clock_ddr3_sdram_2310.1_mhz Aside
from the extreme, you can find DDR3 modules on sale with a declared frequency of more than 3GHz.
As you know, the desktop Haswell works with DDR3 memory, while Haswell-e works with DDR4. The haswell controller was originally developed for DDR4, and therefore it is not difficult to take high frequencies for it. I am absolutely sure that any haswell on a daily basis will work fine with memory at a frequency of at least 2666MHz.

7Rain7
I would not recommend you to boot into Windows with potentially unstable memory. If blue screens through the fault of the processor are relatively harmless for the system, then memory bsodes can easily collapse Windows, you will have to restore it later. Sometimes a very time consuming process.

DartMaul
Yes, in principle, the collapse of Windows is not terrible. I didn't put anything special on purpose. So, a couple of progs, firewood and codecs. I wanted to see what and what this system is capable of. Of course, the dies too)). So the killed Windows will not grieve too much. But I will know on what settings to leave the dice. All the same, I rearranged the entire system after changing the platform.

DartMaul
Almost 40 GB on a regular controller the same as mine? On the same processor as mine? Not bad.
But this is only one channel. And I get to load the controller on two channels up to 34 GB, 17 GB per channel.
I ran it a couple of times at 2400, I don’t remember which timings I set, everything worked without bsods, but I was confused by the lack of gain, relative to 2133. Does this mean that I ran into the transparent ceiling of the controller's bandwidth? Does the controller work for chtoli wear? Or are the memory chips already starting to bend over?
Wow, I haven't asked so many questions at once. Already a little tired. By the way, we nofftopped here very seriously, as if we didn't have to rake. Good night, Your Grace.

7Rain7 Do
n't overdo it.

PS_ Oh, we were moved to the appropriate topic! Thank you, dear -Sergey Viktorovich-. I thought that at least I would get it.

Belomorkanal Does
this mean that I ran into the transparent ceiling of the controller's bandwidth?
No. The memory controller overclocks with the memory. If he can no longer, the system simply won't boot.
Put some kind of MaxxMem or AIDA, they show the memory bandwidth.

-Sergey Viktorovich-
Thank you for the relocation. I just wanted to suggest moving into the topic from memory.
DartMaul
By the way, in Beavis, it shows that the controller operates at 2400 by default. If I understood correctly, then this is exactly how far it can use the memory?
By the way, 9-10-10-27 with a spill of 1.7 and a frequency of 2133 yesterday, on the third run, I caught an error with a linpak. I will test in the evening. Either I will raise it a little, or I will set the default frequency of the controller to 2400. Now the frequency has been dropped to 2200 with my hands. Or I'll play with the timings. Because before that, with a spill of 1.68V, 2133 took 15-20 runs (no longer drove) without errors. But there the timings were porn, set by the Beavis machine.

Belomorkanal
My second warning will result in a ban. Mate had a fight))). But there, without obscenities, the entire literary meaning of the phrase was lost.
You haven't buried yourself in the ceiling. It's just that according to many tests it shows that there is practically no real growth after 2133, and even FPS it does not add at all. Visually, we certainly won't notice these fractions of a second. By the way, I noticed it again during the tests with the linpak for stability. Changing the memory frequencies from 1600 to 1866 made it possible to notice even visually faster passing of tests by 15 seconds. From 1866 to 2133, the increase in the speed of passing was already 5 seconds. Maybe, of course, I confuse warm with soft, but I exclude the error, since the speed of passing the tests every time the frequencies were raised was higher. And linpak showed 1.5k more gigaflops.

DartMaul The
manufacturer of memory modules, not the chips themselves, but modules, claims 12.8 GB / s for one bar, and in all tests I get almost exactly 32 GB / s for both channels. It turns out that I exceed the declared bandwidth on each plank by 3.2 GB / s. And this is with complete stability, without a single error and without heating.
It turns out that the models more expensive than mine have the same memory chips, it's just that this unnecessary XMP stray is preset to higher frequencies. It turns out that manufacturers are deceiving us by selling the same planks at completely different prices. After all, my pair of planks fulfills the standard of planks, which are almost 2 times more expensive. And according to tests in Aida, my memory is generally one of the fastest in 2-channel mode.

Belomorkanal
I like described above where all these numbers come from?
12800MB / s is 1600MHz * 64 bits. Accordingly, this is the peak bandwidth of the interface at 1600 MHz. Accordingly, if you increase the frequency, the bandwidth increases.
The memory usually does not heat up at all, out of ~ 50 modules that have passed through me over the past 2 years, only PerfectStorms have been warming up on hypers.

It turns out that the models more expensive than mine have the same memory chips, it's just that this unnecessary XMP stray is preset to higher frequencies.
Not certainly in that way. The chips may indeed be the same, the only problem is that the quality of the chips varies greatly. At the factory, where the chips are molded onto textolite, the modules are selected according to their quality (i.e., overclocking potential).
Accordingly, the higher the XMP, the higher the bin, the more careful the selection, the better the overclocking potential. Of course, you can find cool chips on 1333/1600 ref modules, but the chance is 1-2%.
As for: After all, my pair of bars fulfills the standard of bars, which are almost 3 times more expensive:
The fact that your modules reached the XMP of higher bins does not mean anything. If you are interested in Samsung D chips, then they are selected on the basis of whether they can 2800 MHz with delays of 9-12-12, and if so, at what voltage. Good modules are made for 2v. Very selective ones do below 1.92v, and if the voltage is raised to the ceiling, then 3000s9 are given by some especially selected specimens. (this is with air cooling and room temperature, of course)
This is despite the fact that these chips are almost linearly scaled to ~ 2.2v
With a probability of 98% I can say that your 2800 9-12-12 modules will not do even at 2.2v. And those that have XMP at 2400c9 or 2666c10 usually do at least 2.05v.

(I by the way this month waiting for 2 sets of 4 strips each on Samsung 2400s9. will arrive poteschu unsubscribe)

And in tests in Hades
In tests Aida No saved overclocking results, ie there are no modules at a frequency above 1866 (except there are 2133 modules working on the platform with A10). So it's not surprising that you burst in there.

7Rain7
First, announce the model of the memory modules. To talk about timings, you need to understand on the basis of which chips the modules are.
On FXs above 2400, memory is tough. It is necessary to add CPU / NB at once. Put 1.2v to start, it should help.

DartMaul
"Accordingly, if you increase the frequency, the bandwidth increases."
And I was a fool thinking that this limitation is valid at any frequency. Now I realized that the declared bandwidth is tied to the declared frequency and timings.
I'll try to take 2800 today, otherwise I've been using it for half a year, but I still don't know the upper limit.

White channel
Bandwidth is the channel width * frequency (bandwidth = bus width * transmit freq). This is generally a general definition, valid for any bandwidth, no matter what.
That is, if you transmit on a 64-bit channel 10 times per second, you can transmit 640 bits. And if 20 times per second, then 1280bit.

DartMaul
DDR3 4Gb 1866MHz Crucial memory (BLT4G3D1869DT1TX0CEU) RTL Ballistix Tactical MT / s CL8
XMP version 1.3 Micron
chips, it is written on the box, as well as in aida. The standard timings for 1866 are 9-9-9-27 with a pressure of 1.5V. In fact, at this frequency, the mother even lowers the sprung to 1.49. XMP sets the same as it has already announced, only with the sprung 1.48V. I will not drive them in principle above 2133 on a permanent basis. CPU / NB spill in Beavis automatically at 1.219 somewhere like this shows. In the test that I posted recently, they are. There are also stable timings at frequencies. It's a pity I didn't find the Mushkins. Well, or ripjams. So there were these in terms of price / quality.

the mother even reduces the pressure to 1.49.
Natural drawdown from 1.5.

Chips are microns, the question is what. If D9P then there probably doesn't make much sense to twitch, and if D9Q then you can twitch. Then, on occasion, I will ask Semyon, maybe he will answer me for a mere mortal)
D9Q require high tRP, if at 2133 you can forget with timings 9-10-9 then most likely D9P.

In general, microns cannot reach high frequencies, so you will not go further than 2400, no memory will give you a percentage.