It is not the timing that guarantees the reliability, the clocking does, same idea as what CPU would respond to overclocking. If you can have a tight timing at 2400, keep that, this is considered reliable and best. I think 2666 is way over for DDR3. I have not heard anyone go that far when I looked in overclocking forums but that was quite a while ago.

Honestly, I feel safer doing at 2400.

Regarding USB 3.0 vs XP, interesting.
I'll recheck once someday I have chance to install XP Pro 32bit on the Z420 and see if it is actually USB 3.0. This is based on Ivy Bridge and is 2011 socket with DDR3 1866 memory in it.

Cheers,

pentiumspeed wrote on 2023-07-25, 23:16:

It is not the timing that guarantees the reliability, the clocking does, same idea as what CPU would respond to overclocking. […]
Show full quote

It is not the timing that guarantees the reliability, the clocking does, same idea as what CPU would respond to overclocking. If you can have a tight timing at 2400, keep that, this is considered reliable and best. I think 2666 is way over for DDR3. I have not heard anyone go that far when I looked in overclocking forums but that was quite a while ago.

Honestly, I feel safer doing at 2400.

Regarding USB 3.0 vs XP, interesting.
I'll recheck once someday I have chance to install XP Pro 32bit on the Z420 and see if it is actually USB 3.0. This is based on Ivy Bridge and is 2011 socket with DDR3 1866 memory in it.

Cheers,

I hope it's the combination of timings and clock frequency that guarantees stability or I have been doing it wrong this whole time! 😁 There are of course certain limitations on both depending on the model of memory IC and the quality of said chips and also of course the platforms involved.

.

.

.

When it comes to feeling safer at 2400 MHz that's certainly true for the integrated memory controller in my i7 4790K and I will show why. I'm not entirely sure the CPU alone is at fault. I found an indication that at least some some Asus Z87 motherboards have broken performance using the 2666MHz memory setting.

Looking at the results of this HWBOT review of the ASUS Gryphon Z87 with an overclocked i7 4770K using memory clocked at 2666MHz with 10-12-12 timings he got totally broken bandwidth results in Aida64 and less horrid but still rather crappy results for the memory spec in most other benchmarks.

https://community.hwbot.org/topic/77043-asus- … hon-z87-review/

Guess what issue I now have with my test system with Asus Z87-A and (now seemingly fully stable) 2666 MHz memory with 10-12-12 timings? Yea broken bandwidth results in Aida64 and also a PhotoWorxx score at 17.7K instead of the ~20K I expected, the latency is decent though. The general performance is just as for the HWBOT reviewer bad but not horrible.

.

I tested everything. different voltages, different timings and so on but the performance using the 2666MHz memory multiplier is consistently ranging from broken to worse than expected not so much depending on the memory timings but more the benchmark used.

.

A picture showing the memory kit (#9) used for this testing.

.

I got the be.. least bad performance at 2666 MHz with these timings but I tried everything from "Auto" to sensible to this.

#9 2x2GB single ranked Samsung 12800E 2Gbit D-die HCK0

Asus Z87-A + 4790K @4.9GHz.

Memory at 2666 MHz @1.7V (I also tried more voltage).

.

Primary timings

tCL. ___10
tRCD. __12
tRP .___12
tRAS. __23
CR. ____1

Secondary timings

tRRD.___ 4
tRFC. ___107
tREFI.___65535
tWR. ___14
tRTP. ___9
tFAW.___16
tWTR.___8
tCKE. ___4
tCWL.___9
RTL . ___"Auto" (43,44)

Tertiary timings

tRDRD. _____6
tRDRD_dr. ___1
tRDRD_dd.___1
tWRRD._____18
tWRRD_dr.___1
tWRRD_dd.___1
tWRWR._____6
tWRWR_dr.___1
tWRWR_dd.___1
DEC_WRD. ___Auto
tRDWR. _____10
tRDWR_dr.___10
tRDWR_dd.___10

.

Lets look at the not so great results. I will not add these results to the list in post number 3 as they are not a true representation of what the performance should be.

.

First the totally broken Aida64 results, only the latency is near the expected score. The memory bandwidth is 3/4 of what it should be. "Read" and "Write" should both be at around 40K and "Copy" at least over 30K. This is not what we see.

.

The Frybench x86 result is actually decent with 3m 31s render time just beating the kits score at 2400 MHz. At 2400 MHz tFAW was at 20, probably because CL9 @2400 @1.65V was just at the limit for the weaker of the sticks. Here tFAW is at 16 and the tertiaries set, CL10 is well within the modules limit for 2666 MHz at 1.7V. If everything would have been as it should I would have expected the score to be 3m 28s or 3m 29s with 3m 29s as more likely.

.

The story in SuperPI 32M is much the same as in Frybench, decent but worse than expected. The 2666MHz result is matching the score at 2400MHz but not beating it.

.

And once again in Doom III 1024*768 "Ultra" the 2666MHz result matches but do not beat the result at 2400MHz. In Doom 3 this setting should have scored at least 675 FPS.

.

The thing with these scores is that if you didn't benchmark and compare you would never kn0w anything was wrong. There could very well be a wide spread issue with the 2666MHz memory multiplier among budget oriented Z87 motherboards and few would notice or care if they did. I know some memory multipliers tend to give broken memory performance under certain conditions on Ivy Bridge-E and Haswell-E but I'm not aware of such bugs with vanilla Haswell.

The only theory when it comes to weak memory performance on Haswell at higher memory speeds I can find seems to be weak memory training. I can not say I saw a big difference in performance when the RTLs trained good verses when they trained less optimally which did happen. The other part of the theory is that the motherboard chooses weak tertiary timings and while true I tried a wide range of tertiary timings and the broken memory bandwidth (measured in Aida64) was rather consistent. I have a hard time seeing a 25% bandwidth loss from tertiary timings while the latency keeps decreasing.

I think I will give up on 2666 MHz with these sticks for now and get back to posting more results and adding more memory kits to the list. The number of tested kits is 24 at the moment so I have bit of a backlog. I will at some later time use another CPU to see if the performance issue at 2666 MHz memory speed is the boards or the CPUs fault.

This review is amateur but the scores tells alot but not enough tuning, hence amateur. This why I carefully choose review sites. Anandtech is very good review and they have lot of articles on memory tuning over the year, search for these.

Keep in mind, Timing is timing and this determines latency. If too much of this performance is compromised while stressing the memory modules, that might kill them, at so high clock, so much overclocked memory with too loose timing to make things to work is worthless than a reliable, long lasting, evenly tuned reasonable, clocked with tightest settings is best overall. Best memory are the ones with tightest timing at each clocked levels. If things start to wilt as you go higher and looser, then stop, step back down to 2400.

Go with 2400.

The best ones are actually the high end overclockers' memory like Ballistrix and others. Also high end board is a requirement. Ones that have one phase memory VRM is noisier than two or three phase memory VRM on the motherboard helps with stablity of the memory too. Same with CPU VRM phases, more the better. I was looking at Z87 boards and was horrified to see 150 to 200 used because I happen to have Haswell i7-4790K.

I strip off cheapest registered ECC memory modules with spring clamps then reuse the heatsinks to cool the memory modules if the best memory modules does not have one. PS: The one with spring clamps will have true thermal pads. The memory modules, those overclocker's too that are with no clamps tends to have wrong adhesive which is not thermal. Trust me. I had once one OCZ pair that had copper heatsinks with no spring clamps used clear adhesive tapes.

This why I liked Mushkin so much back in the day in DDR2 days, very good compatibility even for a third-party memory module and used spring clamps with thermal pads.

Remember this is not odd, not something wrong with CPU. CPU simply can be clocked lower or higher as you like and memory controller is also sensitive to this as well, and if you also clock that part for memory too far causes instability (crashing) and had to loosen the memory timing is not the solution, that is really band aid and damaging to the parts. It is combination of CPU, memory and rest of board determines the end result stability when clocked higher.

Going too close to the edge will damage faster or unstable. This why good overclockers pull back from last step that caused issues or drooping off at *least* 2 to 3 steps back for long term reliability.

Even used, they are not cheap still that why I rarely dabble in overclocking. The most I did was 2 successful overclocking was back in the day, Am486DX 40 to 50 and Pentium MMX 200 to 225. Cost was the major reason not to get best parts, even true to that day even now.

Cheers,

Now, talk about cost. Really. Overclocking success was fed on by oppostitsic makers who make profits on these people who wanted extra performance that amounts to 20% total "freebie" to them but those people was happy to buy for about 50 to 60% more cost. I'm not into these part.

If I have to spend 500 to 800 more or more to get reliable 15 to 20% increase over the normal expenses of from decent parts to the top of the line boards , then I write this off not worth it. That why I give hard pass to liquid cooling too.

GPU cards are never had overclocking in mind at all as these have tendency to have poor air cooled heatsink designs due to poor fan design and wrong heatsink design all had to be packed into small space. Blown VRM are all too common due to poor power supply and overclocking. Right now the Nvidia has the egg on their face on newer ATX 12V connector plug burning out (reason for this was soldered, not crimped), for the RTX 4080 and newer. The ones that is high end tend to be about 200 to 500 extra over the run of mill GPU due to decent heatsink, like all copper, triple slot, much larger fans. But you can get one decent one with 8 even 10 phase VRM for about 100 more, which is sensible but only give you about 10% increase in overclocking or reliability. I don't want to overclock a RTX 3080 due to cost, as they cost about 350 to 400 used, but improve the cooling.

Good example. EVGA had Classified editions GPUs and was extremely expensive same with Kingpin. Which I happen to have two GTX 580 EVGA Classified. of these. What a monster but it is for vintage gaming.

The majority of decent boards is mostly wannabe overclocking boards like 4 phase doubled to two pairs of 4 phases with memory VRM still at 1 phase maybe two. But these wannabe boards are good for reliable computing. Stock workstation boards is made for extra reliability for example HP Z420 have two phase memory VRM for 4 memory slots. There's two pair of 4 memory slots with two phase VRM each. The CPU VRM has 6 phase. I was not overclocking just extra reliability and they cost reasonable used. I can get a Z420 for 200 or so plus shipping extra. Which I did awhile ago.

Remember, this is lottery on who have the lucky chips. I cannot go back to the store and exchange 3 or 5 times to find a cherry CPU, store wil refuse that. I don't justify paying for $300 top of the line board. Not worth it for me, for liquid cooling set up. They tend to have higher probability of to leak and destroy things.

I cannot justify buying best memory modules which is all third-party, not OEM except Crucial.

And have to pay 250 to 300 for a 1KW 80 Gold or Platinum so I don't have to pay for utility bill and only choice is Seasonic which designs and sells theirs. Rest doesn't. Even liteon doesn't have one for gamers just run of mill barely specified. Even the most of power supplies are specified and designed by third party, rare few is quality and I don't want to game on their word of reputation by users.

Remember, the most feeling is overclocking the CPU and some from memory tweaking. GPU not so much.

Also you have to pay 200 for a windows 10 Pro or 11 Pro. Workstation and business computers is free due to COA key burned into firmware of the motherboard.

That family computer back then in Haswell era, was $400 for haswell i3, 2GB ram, but Asus motherboard, GPU was bought (HD7770) between I and mom but rest was paid by parents for their computer, was switched for one with 2 DDR3 slots with H41 chipset (horrid), at last minute without me knowing that due to lack of availability. That was in January. I supplied the COA from one of old computers. That was very reasonable.
It incrementally upgraded to better parts I can afford. Got i7, another GPU bit better than GTX 960 and finally got tower heatsink to keep noise down and 16GB max. If board was 4 slot memory, it would had ended up 32GB.

For mine in computing, I buy used workstations and Xeon CPUs and put ECC memory in them. They have 80 plus Gold rated PSUs but limited to 400W on Z220, Z420 increased to 600W PSU. There is solutions if one is careful is put in compact PSU like from the server for 500W (80 plus platinum) for 50 total shipping included, wired it up, and remote turn on together from motherboard just to power the GPU like GTX 1080 for example.
Yes need to modify the PSU to be open vented and use quieter larger fan. I intend to implement that.
Cheers,

The summer vacation was coming to an end and I had to focus on the stuff I was supposed to do on the house and in the garden and so on regardless of the bad weather. During my 4 weeks leave it rained every single day so I timed it perfectly.

Now I'm back at work so more time for non work! 😀 I will leave the test system at my temperature controlled home from now on though to keep the test conditions more consistent.

.

Memory kit #15 is this Adata XPG AX3U1600W8G9-DB 2x8GB set with 4Gbit ICs of uncertain origin. Likely 4Gbit memory chip candidates are Hynix AFR and Samsung D-die.

.

The best JEDEC spec these dual ranked 8GB modules support is 1333MHz with 9-9-9-24 timings. There is also an 1600MHz XMP profile with the same primary timings except for the command rate that is relaxed to 2.

Unless especially interesting I will from now on only do a single post for each memory set and only post the pictures showing the overclocked results. The results when running with the modules SPD spec and XMP profile (when available) will of course still always be added to the list in post 3 on the threads first page.

.

The best overclocked settings I could find for this (#15) A-data memory kit was 2400 MHz with 11-13-11-25-230-1 tRRD4 tFAW16 timings @1.65V. These timings are pretty typical for both Hynix 4Gbit AFR and Samsung 4Gbit D-die. I think I will be able to do a better guess which of the two after I have tested more modules with these IC types.

.

Lets see if the performance is decent even though the tRCD timing is on the weak side.

2400MHz 11-13-11-25-230-1 tRRD4 tFAW16 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

2400MHz 11-13-11-25-230-1 tRRD4 tFAW16 @1.65V. Frybench x86.

2400MHz 11-13-11-25-230-1 tRRD4 tFAW16 @1.65V. SuperPI 32M.

2400MHz 11-13-11-25-230-1 tRRD4 tFAW16 @1.65V. Doom III 1024*768 "Ultra".

.

The results are nice as the rather weak tRCD and tRFC timings doesn't seem to hurt the performance much at 2400 MHz. For a pure Windows XP system there are slightly better options but for a system dual booting Windows 10/11 and Windows XP sticks like these will do just fine.

Memory kit #16 is a dual ranked 2x2GB Micron 10600U set with 1Gbit D9KPT ICs.

.

The best JEDEC specification these (#16) modules support is 1333 MHz with 9-9-9-24 timings.

.

When it comes to overclocking the fastest setting I could find was 1866 MHz with 9-9-9-21-56-1 tRRD4 tFAW16 timings @1.65V. This result is very typical for what you can expect from this 1Gbit D9KPT/D9KPV Micron IC.

.

Lets take a look at the overclocked performance.

1866MHz 9-9-9-21-56-1 tRRD4 tFAW16 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

1866MHz 9-9-9-21-56-1 tRRD4 tFAW16 @1.65V. Frybench x86.

1866MHz 9-9-9-21-56-1 tRRD4 tFAW16 @1.65V. SuperPI 32M.

1866MHz 9-9-9-21-56-1 tRRD4 tFAW16 @1.65V. Doom III 1024*768 "Ultra".

.

Nothing spectacular, nothing really bad. These Micron D9KPT/D9KPV memory chips are more suited for socket 775, socket 1366 or perhaps socket 1156 rather than later platforms.

Memory kit #17 is a Corsair Dominator CMP4GX3M2A1600C9 ver3.1 kit using the same type of Micron ICs as the kit above.

This kit is also very similar to the Corsair Dominator v3.5 kit (#8) that has tighter XMP timings but in all likelihood also uses this type of Micron IC. The different version numbers (ver3.1/ver3.5) are a bit of a mystery but as you will see when comparing the overclocked results the IC type used in these kits is most likely the same.

.

The best JEDEC spec these (#17) modules support is 1333MHz with 9-9-9-24 timings.

.

The best overclocked setting I could find for this kit was 1866 MHz with 9-9-9-22-58-1 tRRD4 tFAW16 timings @1.65V. So far these D9KPT/D9KPV Micron IC seems extremely consistent when it comes to overclocking.

.

It would be easy to assume that this kit (#17) performs like kit #8 and kit #16 when overclocked...

1866MHz 9-9-9-22-58-1 tRRD4 tFAW16 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

1866MHz 9-9-9-22-58-1 tRRD4 tFAW16 @1.65V. Frybench x86.

1866MHz 9-9-9-22-58-1 tRRD4 tFAW16 @1.65V. SuperPI 32M.

1866MHz 9-9-9-22-58-1 tRRD4 tFAW16 @1.65V. Doom III 1024*768 "Ultra".

.

... and you would be 100% correct assuming so.

This unassuming dual ranked Corsair XMS3 2x4GB memory kit (#18) marks the debut of another once popular and still well performing DDR3 IC, the Hynix 2Gbit BFR.

.

The best JEDEC standard this (#18) Corsair XMS3 CMX8GX3M2A1600C9 ver5.11 kit support is 1333 MHz with 9-9-9-24 timings. There is also an XMP profile for 1600MHz with 9-9-9-24 timings but with the command rate (as is often the case) relaxed to 2.

.

The best performing overclocked setting I could find for this XMS3 kit was 2133 MHz with 9-11-10-22-94-2 tRRD4 tFAW16 timings at 1.65V. This is the first memory kit so far that did not want to overclock even half decently using command rate 1. Luckily the performance difference between CR1 and CR2 is hardly notable on this platform.

.

Lets take a look at the overclocked performance.

2133MHz CL9-11-10-22-94-2. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

2133MHz CL9-11-10-22-94-2. Frybench x86.

2133MHz CL9-11-10-22-94-2. SuperPI 32M.

2133MHz CL9-11-10-22-94-2. Doom III 1024*768 "Ultra"

.

The performance is mostly as one would guess. I would perhaps have expected the Doom 3 result to be 1-2 FPS better. I don't know yet if it's because of randomly unlucky benchmark runs or if I underestimate the impact of the command rate in Doom 3.

Memory kit #19 is a 2x8GB dual ranked Hynix 12800U kit with 4Gbit AFR ICs.

.

The best JEDEC standard these (#19) Hynix 8GB modules support is 1600MHz with 11-11-11-28 timings, the tRFC value is very loose @208.

.

The best overclocked settings I could find for this kit was 2400 MHz with 11-13-12-24-194-1 tRRD5 tFAW20 timings @1.65V. These timings are pretty typical for what you can expect from Hynix 4Gbit AFR at this speed. The tRFC value @194 is actually a bit better than the ~220 I would have expected @2400MHz. I'm not sure if Hynix AFR generally can run tRFC tighter than other 4Gbit Hynix ICs (MFR, BFR) or if it's just this kit, probably just this kit.

.

Lets see if the kit (#19) performs as expected.

2400MHz CL11-13-12-24-194-1. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

2400MHz CL11-13-12-24-194-1. Frybench x86.

2400MHz CL11-13-12-24-194-1. SuperPI 32M.

2400MHz CL11-13-12-24-194-1. Doom III 1024*768 "Ultra".

.

No surprises!

I think Hynix AFR is considered the weakest out of the Hynix 4Gbit ICs, behind MFR and BFR. 4Gbit AFR should on average overclock better than Samsungs 4Gbit D-die but worse than 4Gbit B-die. All in all these (#19) Hynix 12800U modules are good mid-tier 8GB sticks that (when overclocked) won't hold a fast dual booting XP/W10 Ivy Bridge or Haswell system back.

Memory kit #20 is also a 2x8GB dual ranked 12800U kit with 4Gbit memory chips, this time Samsung D-die.

.

The best JEDEC standard these Samsung modules support is 1600 MHz with 11-11-11-28 timings.

This "1342" (#20) kit actually consists of 6 identical modules from the same batch so we get some binning data for Samsung 4Gbit D-die.

Voltage needed for 2400 MHz CL11.

Module A. 1.64V
Module B. 1.7V
Module C. 1.64V
Module D. 1.54V
Module E. 1.55V
Module F. 1.6V

0 out of 6 sticks could run tRCD12 with full stability @2400MHz, all can run tRCD13.
4 out of 6 sticks can run tRP11 @2400, one (A) needs tRP12 for stability and the otherwise good module E sadly needs tRP13.
0 out of 6 modules could run tRRD4 @2400MHz with stability.

I ended up using these timings for the test using module D and F.

Primary timings

tCL. ___11
tRCD. __13
tRP .___11
tRAS. __25
CR. ____1

Secondary timings

tRRD.___ 5
tRFC. ___224
tREFI.___65535
tWR. ___12
tRTP. ___7
tFAW.___20
tWTR.___7
tCKE. ___4
tCWL.___8
RTL . ___"Auto"

Tertiary timings all (loose) Auto.

So all in all 3 out of 6 sticks can run the tested timings at the tested speed, 4 if you crank up the voltage.
6 out of 6 sticks can run 2400 MHz with timings that at least performs close to the tested spec.
The overclocking results for these modules with Samsung 4Gbit D-die are somewhat better than what the Internet thinks you can expect.

.

Lets see how this overclocked #20 Samsung kit performs compared to the kit #19 with Hynix AFR in the last post.

2400MHz 11-13-11-25-224-1 tRRD5 tFAW20 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

2400MHz 11-13-11-25-224-1 tRRD5 tFAW20 @1.65V. Frybench x86.

2400MHz 11-13-11-25-224-1 tRRD5 tFAW20 @1.65V. SuperPI 32M.

2400MHz 11-13-11-25-224-1 tRRD5 tFAW20 @1.65V. Doom III 1024*768 "Ultra".

.

No surprises! The Samsung 2x8GB 4Gbit D-die kit (#20) performs exactly like the Hynix kit (#19) with 4Gbit AFR both using the identical 1600MHz JEDEC timings and when overclocked. It seems the Hynix kits tighter tRFC timing makes up for the weaker tRP timing. The 2x8GB Adata kit (#15) is still a bit of a mystery. If it wasn't for the fact that it can run tRRD4 @2400MHz I would have thought it to be Samsung 4Gbit D-die and I still hold it as the most likely candidate but it is far from certain.

Memory kit #21 is a dual ranked Corsair XMS3 2x4GB kit with 2Gbit Hynix BFR ICs.

This kit is identical to kit #18 but manufactured half a year earlier if I have deciphered the date code correctly.

.

Just like the identical #18 kit this (#21) kit sports a 1333 MHz SPD profile with 9-9-9 timings and a XMP profile for 1600MHz with the same primary timings but @1.65V and with CR2.

I will still add the SPD and XMP results for this kit to the lists as it's a good showcase of consistency but if I end up testing too many 100% identical kits I will do some culling.

.

This kit (#21) overclocked about the same as the identical #18 kit, tRCD 12 instead of 11 and CR1 instead of CR2. The best setting I could find was 2133 MHz with 9-12-10-22-98-1 timings @1.65V.

.

Lets see if CR1 instead of CR2 makes up for the looser tRCD timing.

2133MHz 9-12-10-22-98-1 tRRD4 tFAW16 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

2133MHz 9-12-10-22-98-1 tRRD4 tFAW16 @1.65V. Frybench x86

2133MHz 9-12-10-22-98-1 tRRD4 tFAW16 @1.65V. SuperPI 32M

2133MHz 9-12-10-22-98-1 tRRD4 tFAW16 @1.65V. Doom III 1024*768 "Ultra"

.

The results indicate that going from Command Rate (CR) 1 to Command Rate 2 is a larger performance loss than going from tRCD11 to tRCD12. This seems logical as the performance impact of tRCD is minuscule while the Command Rate affects the latency.

Memory kit #22 is another dual ranked 2x4GB Corsair kit, this time Vengeance ver2.12 CMZ8GX3M2A1600C9(R).

This kit uses the same type of memory IC as the single ranked 2x2GB kit #3 and the dual ranked 2x4Gb kit #14, the Elpida 2Gbit BCBG/BCSE. If You think the two modules doesn't look identical, blame your eyes not me! I got these modules with a board and CPU combo, I guess they are "survivors". The motherboard does not care one bit that one module is Red so the XMP profile works as expected.

.

The best JEDEC standard these modules (#22) support is 1333 MHz with 9-9-9 timings. There is an XMP profile for 1600 MHz 9-9-9 with CR2 @1.5V.

.

Memory kits #3 and #14 using the same type of ICs didn't overclock beyond 1600MHz CL9. This kit with an XMP profile for 1600MHz CL9 should surly fare better?

Nah it's pretty much the same garbage except that these modules managed CL8 and tRP7 @ 1600MHz. The best setting I ended up with was 1600MHz 8-8-7-20-84-1 tRRD4 tFAW16 @1.65V. Compared to the completion (other memory ICs) at the time this "overclocking result" is not great.

.

Lets see if the tighter timings can help this memory kit at least beat the other 1600 MHz stragglers.

1600MHz 8-8-7-20-84-1 tRRD4 tFAW16 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

1600MHz 8-8-7-20-84-1 tRRD4 tFAW16 @1.65V. Frybench x86.

1600MHz 8-8-7-20-84-1 tRRD4 tFAW16 @1.65V. SuperPI 32M.

1600MHz 8-8-7-20-84-1 tRRD4 tFAW16 @1.65V. Doom III 1024*768 "Ultra".

.

Yes this kit (#22) does indeed beat the other kits stuck at 1600MHz, at least the other ones tested this far.

Memory kit #23 is yet another dual ranked Corsair 2x4GB kit.

This Vengeance kit has the same product number as kit #22 but version number ver5.11 which indicates that the kit uses Hynix 2Gbit BFR ICs.

.

This kit has a JEDEC SPD profile for 1333MHz with 9 9 9 timings and an XMP profile for 1600 MHz with 9 9 9 CR2 @1.5V.

.

The best overclocked setting I could find for this kit (#23) was:

2400MHz, 1.65V

Primary timings

tCL. ___11
tRCD. __13
tRP .___11
tRAS. __27
CR. ____2

Secondary timings

tRRD.___ 5
tRFC. ___108
tREFI.___65535
tWR. ___12
tRTP. ___7
tFAW.___20
tWTR.___7
tCKE. ___4
tCWL.___8
RTL . ___"Auto"

This is not a great overclocking result but this kit at least made it to 2400 MHz. I think to get a substantially better results with 2Gbit BFR I would have to bin all my sticks with these memory chips and combine the best ones (or simply buy a high binned kit). 2Gbit BFR is not like Samsung 1Gbit E-die, there is a lot of variance among these Hynix ICs.

.

Lets see how memory kit #23 performs.

2400MHz 11-13-11-27-108-2 tRRD5 tFAW20 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

2400MHz 11-13-11-27-108-2 tRRD5 tFAW20 @1.65V. Frybench x86.

2400MHz 11-13-11-27-108-2 tRRD5 tFAW20 @1.65V. SuperPI 32M.

2400MHz 11-13-11-27-108-2 tRRD5 tFAW20 @1.65V. Doom III 1024*768 "Ultra".

.

The performance is just as expected.

Memory kit #24 is a Corsair single ranked 2x4GB value ram kit with unknown 4Gbit ICs.

Sadly there are no version numbers on Corsairs value offerings. One thing to note is that this isn't a matched kit, one module was made week 15 2014 the other one week 44 2014. I did not get the modules from the same source and I have not owned the sticks true partners. There is a good chance these two modules aren't even using the same memory chip model.

.

The best JEDEC spec these single ranked 4GB modules support is 1333MHz with 9-9-9-24 timings.

.

Even though it's uncertain if these two (#23) modules are even using the same type of memory IC I know that I have had them overclocked to 2000MHz with 10-10-10 timings on the Socket 1156 platform with about 1.7v and at the time I suspected they used Micron chips. With this Haswell system the best settings I could find was 1866MHz with 10-9-9-23-168-1 timings.

.

The performance results for memory kit #24.

1866MHz 10-9-9-23-168-1 tRRD4 tFAW16 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

1866MHz 10-9-9-23-168-1 tRRD4 tFAW16 @1.65V. Frybench x86.

1866MHz 10-9-9-23-168-1 tRRD4 tFAW16 @1.65V. SuperPI 32M.

1866MHz 10-9-9-23-168-1 tRRD4 tFAW16 @1.65V. Doom III 1024*768 "Ultra".

.

The performance isn't great but this is probably what you can expect from poor clocking single ranked 4GB sticks needing very high tRFC. The 1333MHz (SPD) results are particularly bad and the slowest so far but 100% valid and a true representation of these modules performance.

Memory kit #25 is a dual ranked 2x8GB kit using Hynix best 4Gbit DDR3 memory IC, the 4Gbit MFR.

This kit from week 7 year 2014 consists of 4 sticks, for the test I used stick B and D.

.

The SPD settings for this (#25) kit is 11-11-11-28-208 @ 1600MHz.

.

Because there are 4 sticks and only two are needed we get some Hynix 4Gbit MFR binning data.

Voltage needed for CL10 @2400.

Stick A: 1.70V
Stick B: 1.61V
Stick C: 1.63V
Stick D: 1.65V

Only stick A failed to do 10-11-10 @2400MHz using 1.65V.

.

The best overclocked settings I could find for this (#25) memory kit was.

2400 MHz, 1.65V

Primary timings

tCL. ___10
tRCD. __11
tRP .___10
tRAS. __22
CR. ____1

Secondary timings

tRRD.___ 4
tRFC. ___222
tREFI.___65535
tWR. ___14
tRTP. ___8
tFAW.___16
tWTR.___8
tCKE. ___4
tCWL.___8
RTL . ___"Auto"

This is probably a slightly better result than you should expect from random Hynix 12800U 8GB sticks with 4Gbit MFR ICs.

.

Lets check out the performance.

2400MHz 10-11-10-22-222-1 tRRD4 tFAW16 @1.65V. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

2400MHz 10-11-10-22-222-1 tRRD4 tFAW16 @1.65V. Frybench x86.

2400MHz 10-11-10-22-222-1 tRRD4 tFAW16 @1.65V. SuperPI 32M

2400MHz 10-11-10-22-222-1 tRRD4 tFAW16 @1.65V. Doom III 1024*768 "Ultra".

.

There is a common belief in the PC hardware community that even though the Hynix 4Gbit MFR is the highest clocking DDR3 memory IC it isn't a very high performing one because of the loose memory timings 4Gbit memory chips generally need. These days if you want to be able to dual boot Windows XP and Windows 10/11 and be able to game even in the latter OS you need at least 16GB memory. If you want 16GB of really fast DDR3 memory for your Socket 115x system a kit with a nice bin of Hynix 4Gbit MFR is probably your best bet.

This is the best performing memory kit tested so far.