This (#11) G-Skill RipjawsX kit is the first memory kit so far that has an XMP profile that actually could achieve some decent performance.

.

.

The (#11) G.Skill RipjawsX F3-17000CL11D-8GBXL dual ranked 2x4GB kit has an XMP profile with 2133 with 11-11-11-30 timings. My Asus Z87-A motherboard did choose 2T when set to XMP so that is what is used here.

.

Lets see how this memory performs using its XMP profile.

XMP @2133MHz CL11-11-11-28-171-2. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

XMP @2133MHz CL11-11-11-28-171-2. Frybench x86.

XMP @2133MHz CL11-11-11-28-171-2. SuperPI 32M.

XMP @2133MHz CL11-11-11-28-171-2. Doom III 1024*768 "Ultra".

.

Not much to complain about.

While the G.Skill RipjawsX kit (#11) performed well using the XMP profile I'm sure manual overclocking can improve things further.

.

.

As this (#11) dual ranked 2x4GB G.Skill RipjawsX F3-17000CL11D-8GBXL kit uses 2Gbit Samsung ICs, in all likelihood E-die I expected that it would overclock reasonable well.

The best overclocked setting I could find was 2400 MHz with 10-12-11-24-156-1 tRRD4 tFAW16 timings.

.

Lets see how this fast dual ranked kit with Samsung 2Gbit chips stacks up against fast single ranked kits.

DDR3 2400 CL10-12-11-24-156-1 tRRD4 tFAW16. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

DDR3 2400 CL10-12-11-24-156-1 tRRD4 tFAW16. Frybench x86.

DDR3 2400 CL10-12-11-24-156-1 tRRD4 tFAW16. SuperPI 32M.

DDR3 2400 CL10-12-11-24-156-1 tRRD4 tFAW16. Doom III 1024*768 "Ultra".

.

In SuperPI and Frybench dual ranked memory with more relaxed primary timings beats single ranked memory with tighter primary timings. Doom III shows once more that it isn't very sensitive to differences between single and dual ranked memory.

~5% in Doom3 is a testament to good cache subsystem architecture in Haswell. In contrast modern CPUs are very sensitive to memory latency.

rasz_pl wrote on 2023-07-12, 12:47:

~5% in Doom3 is a testament to good cache subsystem architecture in Haswell. In contrast modern CPUs are very sensitive to memory latency.

Yes modern consumer CPUs are even more sensitive to memory performance in general, mostly because they have many cores and few memory channels. The tasks we use them for also likes to shuffle large chunks of information from here to there all the time.

During the Haswell era the myth that memory performance didn't matter flourished. Memory performance has always mattered to get consistent performance in games even when the average FPS is limited by something else.

The total lack of scaling with faster memory in 3dmark2001 might also have something to do with how Haswells cache subsystem is set up as I have seen scaling on other fast platforms.

There were only 11 matched kits in the scrap lot but by combining sticks with the same layout and ICs I got two more sets of memory.

This memory kit (#12) consists of a 2GB dual rank Samsung 10600U stick combined with a similar 10600E stick, both using 1Gbit E-die HCH9.

.

.

This (#12) is one 2GB stick of regular Samsung dual ranked 10600U memory with 1Gbit E-die chips combined with one stick of similar 10600E ECC memory. The best JEDEC standard these sticks support is 1333 MHz with 9-9-9-24 timings.

.

I know I wrote that I wouldn't do yet another post with dual ranked 2x2GB 1333 MHz 9-9-9-24 memory running its SPD profile but lets see if two unmatched dual rank modules with Samsung 1Gbit E-die yields the same result as a matched pair.

DDR3 1333 CL9-9-9-24-74-1. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

DDR3 1333 CL9-9-9-24-74-1. Frybench x86.

DDR3 1333 CL9-9-9-24-74-1. SuperPI 32M.

DDR3 1333 CL9-9-9-24-74-1. Doom III 1024*768 "Ultra".

.

Yes, yes they do.

How about overclocking the unmatched (#12) Samsung 10600U + 10600E pair.

Do they overclock much worse compared to the matched kits with 1Gbit Samsung E-die?

.

.

No it seems that Samsung 1Gbit E-die is a very very consistent memory IC.

The best overclocked setting I could find for these (#12) unmatched E-die sticks was 1866 MHz 8-10-9-20-60-1 tRRD4 tFAW16 @1.65V... the same settings as the other dual ranked 1Gbit E-die memory...

.

Will the same settings with the same memory layout with the same Samsung memory ICs once again yield the same performance?

DDR3 1866 CL8-10-9-20-60-1 TRRD4 TFAW16. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

DDR3 1866 CL8-10-9-20-60-1 TRRD4 TFAW16. Frybench x86.

DDR3 1866 CL8-10-9-20-60-1 TRRD4 TFAW16. SuperPI 32M

DDR3 1866 CL8-10-9-20-60-1 TRRD4 TFAW16. Doom III 1024*768 "Ultra"

.

Yes, the performance is once again the same.

The next memory kit (#13) is another unmatched 2x2GB Samsung pairing and the last memory kit from the scrap lot.

This time it's at least 10600U + 10600U and both single ranked Samsung 2Gbit D-die!

.

.

These very similar but slightly different (#13) 2GB Samsung 10600U single rank sticks support a speed of 1333 MHz with 9-9-9-24 timings.

.

We already know how single ranked 10600U 9-9-9-24 memory usually performs. Lets see if the sticks being slightly different makes anything slightly different.

DDR3 1333 CL9-9-9-24-107-1. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

DDR3 1333 CL9-9-9-24-107-1. Frybench x86.

DDR3 1333 CL9-9-9-24-107-1. SuperPI 32M.

DDR3 1333 CL9-9-9-24-107-1. Doom III 1024*768 "Ultra".

.

Everything looks just as bad as expected but not worse than other single ranked 2x2GB 10600U kits.

Lets see how this (#13) unmatched single ranked 10600/1333/HCH9 binned Samsung D-die pair stacks up against the 12800/1600/HCK0 binned and matched but otherwise similar kit (#9).

.

.

The best overclock for these sticks I could find was 2400 MHz with 10-11-11-22-96-1 tRRD4 tFAW16 timings @1.65V. One of the sticks wasn't fully stable at 2400MHz CL9 @1.65V but they could run tFAW16 otherwise the timings ended up the same as on the other single ranked D-die kit.

.

So how does this unmatched (#13) Samsung kit compare?

DDR3 2400 CL10-11-11-22-96-1 TRRD4 TFAW16. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

DDR3 2400 CL10-11-11-22-96-1 TRRD4 TFAW16. Frybench x86.

DDR3 2400 CL10-11-11-22-96-1 TRRD4 TFAW16. SuperPI 32M.

DDR3 2400 CL10-11-11-22-96-1 TRRD4 TFAW16. Doom III 1024*768 "Ultra".

.

Not being able to run CL9 at 2400MHz hurts the performance a bit but this (#13) unmatched Samsung kit is still very speedy.

With this being the last memory kit from the scrap lot I have to say that this was a nice amount of data for the ~25 euro spent. Not a single broken memory module is another nice surprise. Still it's not nearly enough data to draw any wider conclusions, luckily I have more memory...

The testing methodology improved with time and I needed to test the first memory kits I tested again to get more representative numbers.

Testing a few memory kits again was easily done but updating the results in the posts here took some time because of Vogons safe guards meant to hinder dis-grunt members from deleting / destroying their old posts. Well I understand the reasons and the updates are now done.

On average the scores do not differ more than half a second or in case of Doom III one FPS from a "true representative number". Some outliers will still exist but they are now far fewer and scores drift less from a "true representative number".

It's time to expand the testing beyond the scrap lot memory to random DDR3 kits that somehow ended up with me.

I guess most kits were either found with or bought with some motherboard and CPU bundle as I never buy memory (except when I do).

.

The first kit (#14) is a G.Skill Sniper F3 kit that actually is a true PC3-12800U kit other than having a 1600 MHz XMP profile.

.

,

This (#14) G.Skill Sniper F3-12800CL9D-8GBSR2 kit consists of 2x4GB dual rank modules with Elpida 2Gbit memory ICs. I find it very likely that the memory chips are BCBG/BCSE.

The SPD timings for this kit are 9-9-9-28-128-1 @1600 MHz. This is a pretty aggressive JEDEC compliant profile.

.

Lets see if this dual ranked 2x4GB kits SPD profile performs as expected.

DDR3 1600 CL9-9-9-28-128-1. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

DDR3 1600 CL9-9-9-28-128-1. Frybench x86.

DDR3 1600 CL9-9-9-28-128-1. SuperPI 32M.

DDR3 1600 CL9-9-9-28-128-1. Doom III 1024*768 "Ultra".

.

The performance is great compared to most other tested kits running their SPD settings, just as expected

With this (#14) G.Skill memory kit I ignored the XMP profile as it's worse than the SPD-profile and only there to ensure compatibility when running two kits with the same product name.
.

.

Lets see how this G.Skill Sniper F3-12800CL9D-8GBSR2 overclocks.

The best result I could manage was. 1600 MHz 9-8-7-18-88-1 tRRD4 tFAW16 @1.65V. The overclocking result (or lack there of) in combination with the week 34, year 2011 manufacturing date and Elpida identifier (640) in the module serial numbers makes me convinced that the memory chips are Elpida 2Gbit BCBD/BCSE.

.

Did the tighter timings brings improvements?

1600MHz CL9-8-7-18-88-1. Aida64 Cache & Memory Benchmark plus PhotoWorxx.

1600MHz CL9-8-7-18-88-1. Frybench x86.

1600MHz CL9-8-7-18-88-1. SuperPI 32M.

1600MHz CL9-8-7-18-88-1. Doom III 1024*768 "Ultra".

Yes, the performance is decent...

Nice results building up, thanks for tackling this. I knew there was a reason I had a soft spot for Elpida on Elpida and always grumbled to myself when I got a Kingston module.

There is also some data on on of the reddits, a wiki of DDR3 chips, think it's the buildapc one.

Got a small bundle of ones and twos myself which I have been kinda ignoring due to wanting 4s mostly for modern office/web boxen. Will have to check for anything interesting.

BitWrangler wrote on 2023-07-20, 18:46:

Nice results building up, thanks for tackling this. I knew there was a reason I had a soft spot for Elpida on Elpida and always grumbled to myself when I got a Kingston module.

There is also some data on on of the reddits, a wiki of DDR3 chips, think it's the buildapc one.

Got a small bundle of ones and twos myself which I have been kinda ignoring due to wanting 4s mostly for modern office/web boxen. Will have to check for anything interesting.

There are a few good compilations of data on memory ICs and modules. The best ones beeing the xtremesystems one and the second best the reddit one. I kind of hate reddit. 😁

My thought with this thread is to first test a lot of memory using the same test system to collect as much relevant first hand data as possible then when necessary add on data from other sources. If I go to the HWbot forum I can for example read that Samsung 1Gbit E-Die is shit because it can only do 2200 MHz CL9 @2.0V... This is not very useful information for most of us even if it is for some of the HWbot members.

I had never (as far as I know) tested Samsung 1Gbit E-die before and I was pleasantly surprised. 3 kits tested, one of them unmatched sticks and they all overclocked and performed exactly the same. To me Samsung 1Gbit E-die seems to be a good choice for a period correct, overclocked socket 1366/1156 system for XP gaming. Memory with Elpida BASE "Hyper", the really high performing memory IC from that time is very expensive.

No point getting Hypers unless you're overclocking on LN2.

2gb Elpida BBSE and PSC modules are still really cheap and perform 99,95% of hypers in regular use. BBSE / PSC were sold ie. XMP 1600mhz cl6-8-6 and 2200 cl9 commonly with good oc headroom.

Anyway, nice statistics 😀

timsdf wrote on 2023-07-20, 21:09:

No point getting Hypers unless you're overclocking on LN2.

2gb Elpida BBSE and PSC modules are still really cheap and perform 99,95% of hypers in regular use. BBSE / PSC were sold ie. XMP 1600mhz cl6-8-6 and 2200 cl9 commonly with good oc headroom.

Anyway, nice statistics 😀

Yea it's all relative I guess. I kind of shy away from paying for memory at all since I'm not a competitive benchmarker. On the other hand I have some nice sticks, Hypers will be featured as will 1Gbit PSC but I think I only have garbage bins of the latter, those varied a lot.

I have some nice dual ranked 4GB modules with 2Gbit Samsung D-die from some decent Corsair series that I kind of expect will do well but I put off testing them as they just as my Hypers are part of quad kits. For best results the modules need to be individually binned and that needs some seriously bad weather. I also just found 12 more single ranked D-die OEM sticks sitting in my spare EVGA SR-2.

I was going to write a long post post here about binning Samsung 2Gbit D-die! 😁

The universe wanted otherwise and I managed to close the tab with a T.L.D.R text written.

I'm not writing all that shit again but to sum it up.

1. Most of my single ranked D-die sticks kind of suck.
2. My i7 4790K has a really weak memory controller.
3. More SA voltage than ~1.0V do not seem to help. 0.95V seems to be the sweet spot most of the time.
4. Raising Analog and Digital I/O voltage more than 0.1V seems to have zero impact.
5. Extra Cache/Uncore voltage beyond the 1.25V I need for 4.5 GHz cache seems to have zero impact.

The combination of the above results in that it's impossible to get 2666 MHz stable at 1.65V memory voltage even though most of these D-die sticks can load Windows at 2666 MHz with only 1.5V memory voltage. Another particularity is that one of the involved parts (CPU/motherboard/memory) do not really like CAS 11 @2666. Stability when possible is found at CAS 10 using 1.7V memory voltage or more. It seems loser secondary and tertiary timings hurt more than help.

Trying to find a better 4790K CPU that both overclocks well and have a great memory controller would be an expensive endeavor. I was toying with the idea of getting a better motherboard, a boxed Asus Maximus VII Hero was up for sale on a local auction site. The motherboard ended up selling for ~65 euro so realistically it would have cost me at least 75 euro shipped, I do not regret leaving it be. I do have a Gigabyte Z87X-OC motherboard that is a decent memory overclocker but that board has a damaged memory slot latching mechanism so it's too tedious to use for testing loads of memory modules. I have two MSI Z87 boards, a G45 and a S02/G50 but I have no idea if they are better than the Asus Z87-A when it comes to memory overclocking. The MSI boards are just boards that ended up with me when I helped friends upgrading.

There are a few other issues with my Asus Z87-A board that makes it less than ideal for this even if I think that for ambient CPU overclocking it is as good as any Z87/z97 motherboard. Board flexing, every time I insert a memory module the paper thin motherboard PCB flexes... a lot. The BIOS is modded with Haswell micro code 7, a micro code from before Haswell was launched back in 2013. I did this mod to be able to overclock "non K" CPUs but I have no idea if this very old micro code hurts memory overclocking.

I have some i7 4770K CPUs sitting on my other Z87 motherboards but they are not (yet) delidded and I seriously doubt one of them will do more than 4.7 GHz with decently low voltage.

Lowering the overclock could perhaps help a bit but truth be told I want o to go in the other direction to 5.0 GHz. Running the CPU at 5.0 would make my results more comparable to other peoples results as this seem to be the most commonly used speed for Haswell memory benching. The other often used speed for memory benching is 4GHz but whats the fun in that? 😁

.

Here are some porn for now and with luck I will get to updating the thread with some more results later today.

Edit.

I replaced the damaged memory slot lock on the Gigabyte Z87X-OC. The board also intermittently suffers from the infamous "code 15" issue but I read somewhere that putting a solder blob over the small resistor at the inner corner of the inner black memory slot should sort that out.

If I get the Gigabyte board running without irritating issues (hard one being Gigabyte) I might switch to that board at some point in the future.

While squirreling about I found some Interesting but not necessarily well performing DDR3 modules like ones with Nanya 1Gbit A-die (zero info on the net) and some Corsair modules with Promos chips.

.

With all single ranked 12800U 2Gbit D-die (HCK0) modules binned it seems one of them is decent.

The typical voltage for stability at 2400MHz CL9-11-11-21 for these modules seems to be between 1.66V and 1.7V. The good stick only needs 1.63V and a few mediocre ones won't play nice with these settings regardless of voltage (= >1.75).

Most sticks can load Windows at 2666MHz CL11-13-13 at 1.5V but no stick is even semi stable with these loose primary timings at 2666 regardless of voltage or secondaries/tertiaries. A few sticks are bench stable at 2666 10-12-12 at 1.7V - 1.75V but I have not investigated stability much further.

At this point I think I'm binning my module mounting / board flexing skills and some modules being a slightly lighter or harder load on my CPUs crappy memory controller as much as binning the quality of the modules them selves. 😁

/Edit.

If you are running one stick, this is not worth it. Needs to be in dual channel mode in 2 or 4 modules to get the most bandwidth and going so far to get 2666 is too much and this one I don't like small margins.
Usually I see about 10% or 15% increase is normal with healthy margins for reliability, this is not a overclocking competition.

Otherwise junk all and go to DDR4 instead, this is lot more bandwidth, 2666 to 3200 is nice spot. Remember this is Haswell, DDR 1600 rated, but can most likely can do 2133 at most, and is firmly in windows 7 to 10 era.

Cheers,

pentiumspeed wrote on 2023-07-24, 19:28:

If you are running one stick, this is not worth it. Needs to be in dual channel mode in 2 or 4 modules to get the most bandwid […]
Show full quote

If you are running one stick, this is not worth it. Needs to be in dual channel mode in 2 or 4 modules to get the most bandwidth and going so far to get 2666 is too much and this one I don't like small margins.
Usually I see about 10% or 15% increase is normal with healthy margins for reliability, this is not a overclocking competition.

Otherwise junk all and go to DDR4 instead, this is lot more bandwidth, 2666 to 3200 is nice spot. Remember this is Haswell, DDR 1600 rated, but can most likely can do 2133 at most, and is firmly in windows 7 to 10 era.

Cheers,

I have to ask do you even understand the concept of binning memory modules after a certain binning criteria?

.

I will answer my self.

When you have a big bunch of similar memory modules that you would like to overclock it's good to get the know the modules on individual basis so you can pair them up appropriately. It's a huge waste to pair a memory stick that can do 2400 MHz CL9 at 1.65V with one that only can do 2133 MHz CL9 at that voltage.

I will also add just to counter misinformation that while Haswell indeed officially do not support higher than 1600 MHz memory speed the highest working memory multiplier for Haswell is generally the 2933MHz one. Stability with two dual ranked or four single ranked sticks is almost guaranteed at 2400MHz with the right memory kit on a decent motherboard and most Haswell CPUs is capable of memory speeds well above that if you put the work in. You can argue that Ivy Bridge was/is a more consistent memory clocker and you probably be right.

I have to admit that I do not know if Intel locked the memory multipliers on some of the lower end Haswell CPUs. I do know that my i7 4770S is not locked down when it comes to memory multipliers. The i7 4770S used to run 2x8GB OEM Samsung 12800U @2400MHz on the same Asus Z87-A I use for this testing, it was stable for years.

I meant that I'm aware the Haswell can have different memory speeds as part of overclocking and improve performance, sure thing. And this action of doing this *is* expensive enterprise in outlay of collecting lot of parts to bin a set of best stuff. I cannot do this and quality DDR3 motherboards costs much, so close to new current generation boards, I cannot justify that. Plus I would have spent lot to get windows COA especially windows 10 is about 299 for Win 1o Pro. Or use unsupported windows 7.

I'm looking at this from windows context. Haswell does not support XP windows, the minimum supported is windows 7. Last processor support for XP was Ivy bridge.

When you go to the point of running windows 7, you are essentially now have many choices: of windows 7, 8.x and windows 10 at that point since the programs supports all these 7 to 10, options opens up to many choices of processors not just Haswell. For this, I would just dump and go for DDR4 based computer. Easier to get set of DDR4 memory modules that overclocks and used motherboards is reasonable.

Overclockers' memory modules that you can get is still expensive even used even for DDR3. I couldn't find any that I like and difficult to find as third party overclocker's DDR3 modules are in short supply. In my case if I wanted to is Crucial's Ballistrix. But plenty of OEM DDR3 in 1600 and 1866 and I chose Z420 to use buffered 1866 fully filled to 128GB total.

Anyway,
The way I destroy extra expenses was buy a workstation or business line computer used. These comes with COA key built into the motherboard. Recently last spring, I spent on extra 16GB DDR4-2666 to bring up to 32GB, 1TB WD black NVME SSD and put into the $500 HP Elitedesk 800 G5 Mini (65W version) that have i5-9500 in it. Activated the windows 11 Pro on it. No extra expense. Much nicer and I had to spend more than I liked in order to learn windows 11 stuff bit early. This is for my work as we repair and support computers not just cell phones and consoles.

I have done same cost-sensitive concept with Z220, Z420 and upgraded them to specs I liked them progressively, yes that costs but affordable to me. Then got computers activated windows 10 Pro at no extra expenses.

In 2024, I plan to buy 2 more computers also third one for work, at my work my is HP 800 G1 tower which is also Haswell serving me well nearly 5 years now, rarely maintained. They will be workstations class to replace the parent's Haswell PC going for over 11 years by now and my main 7 years old ivy bridge Z220. All computers had to be replaced to support windows 11 by 2025.

Cheers,

pentiumspeed wrote on 2023-07-25, 16:30:

I meant that I'm aware the Haswell can have different memory speeds as part of overclocking and improve performance, sure thing. […]
Show full quote

I meant that I'm aware the Haswell can have different memory speeds as part of overclocking and improve performance, sure thing. And this action of doing this *is* expensive enterprise in outlay of collecting lot of parts to bin a set of best stuff. I cannot do this and quality DDR3 motherboards costs much, so close to new current generation boards, I cannot justify that. Plus I would have spent lot to get windows COA especially windows 10 is about 299 for Win 1o Pro. Or use unsupported windows 7.

I'm looking at this from windows context. Haswell does not support XP windows, the minimum supported is windows 7. Last processor support for XP was Ivy bridge.

When you go to the point of running windows 7, you are essentially now have many choices: of windows 7, 8.x and windows 10 at that point since the programs supports all these 7 to 10, options opens up to many choices of processors not just Haswell. For this, I would just dump and go for DDR4 based computer. Easier to get set of DDR4 memory modules that overclocks and used motherboards is reasonable.

Overclockers' memory modules that you can get is still expensive even used even for DDR3. I couldn't find any that I like and difficult to find as third party overclocker's DDR3 modules are in short supply. In my case if I wanted to is Crucial's Ballistrix. But plenty of OEM DDR3 in 1600 and 1866 and I chose Z420 to use buffered 1866 fully filled to 128GB total.

Anyway,
The way I destroy extra expenses was buy a workstation or business line computer used. These comes with COA key built into the motherboard. Recently last spring, I spent on extra 16GB DDR4-2666 to bring up to 32GB, 1TB WD black NVME SSD and put into the $500 HP Elitedesk 800 G5 Mini (65W version) that have i5-9500 in it. Activated the windows 11 Pro on it. No extra expense. Much nicer and I had to spend more than I liked in order to learn windows 11 stuff bit early. This is for my work as we repair and support computers not just cell phones and consoles.

I have done same cost-sensitive concept with Z220, Z420 and upgraded them to specs I liked them progressively, yes that costs but affordable to me. Then got computers activated windows 10 Pro at no extra expenses.

In 2024, I plan to buy 2 more computers also third one for work, at my work my is HP 800 G1 tower which is also Haswell serving me well nearly 5 years now, rarely maintained. They will be workstations class to replace the parent's Haswell PC going for over 11 years by now and my main 7 years old ivy bridge Z220. All computers had to be replaced to support windows 11 by 2025.

Cheers,

Haswell was supposed to have XP support and Intel did in fact host the drivers on their page for some time. Shortly before launch Microsoft probably nudged Intel and made them withdraw XP support for Haswell (pure speculation).

There were some odd Haswell systems that did get XP support like that "Gigabyte HTPC thing" and probably a few others. I used the driver package for the "Gigabyte HTPC thing" for this test system but it isn't really needed. The only issue I know of with using Haswell with Windows XP is the exact same issue as with Ivy Bridge, no support for Intel USB3. The USB3 ports still work in 2.0 mode so I do not really find this an issue at all. Drivers for various support chips can be found at the chip makers homepages. For the lazy there are auto updaters and even modified Windows XP install media with "everything" included.

When it comes to memory, much stuff that is valid for Haswell is valid for Sandy Bridge-E , Ivy Bridge and Ivy Bridge-E as well. Sandy Bridge "non E" is limited to 2133 MHz memory speed plus what ever little extra you can get by overclocking the system bus. Bloomfield realistically tops out at 1866 - 2000 MHz memory speed for everyday use depending on how well the uncore overclocks but Gulftown goes a bit further. I have not played much with memory overclocking with Lynnfield but 2133 MHz should be doable.

.

I do not think I ever bought any expensive overclocking memory at all. For my first DDR3 system and the only one I bought new (P6T Deluxe, i7 920) I bought 1333MHz memory. I have scored some nice used memory sticks in packages with motherboards and CPUs or when helping friends upgrading. The best source through the years has probably been complete boxes found in waste rooms or dumpsters. I have now moved to a small town and do not have access to electronic waste dumpsters or waste rooms any more. The only DDR3 memory sticks I have actively bought this side of 2010 is used 8Gb sticks for upgrading stuff, those I do buy when cheap enough. I'm also not afraid to pick up scrap lots with memory (and/or other stuff) when super cheap.

Most of these poor 2GB Samsung D-die sticks comes from boring office PCs that got upgraded some time during the winter 2012-2013 when they got upgraded from 4GB to 16GB memory and at the same time from Windows 7 to Windows 8. The old 2GB modules were seen as useless so they ended up with me, I think I had at least 40 of them at one point. I now think I know yet another reason/excuse my results with these D-die sticks are so poor. This is not the first time I bin these HCK0 D-die sticks, the first time was with socket 1366. This was shortly after I got them and I remember I just sorted them as good and not as good but I didn't mark them. My bet is that the "good pile" got separated and these are the not so good ones. With my luck the good ones probably ended up in systems I have sold, who cares about 2GB DDR3 sticks anyhow... 😁

.

I ended up testing the "kit #9" single ranked 12800E ECC modules with HCk0 D-die individually and these do not have an issue running the very slack "auto" timings at 2666MHz like the otherwise similar 12800U ones, I find this somewhat strange. One of these modules are just as good as the best out of the big pile and the other isn't bad either.

With a few modules that actually do seem to work at 2666MHz the next issue is performance. For some reason the performance is much worse at 2666 MHz CL10-12-12-23-107-1 with tight secondaries compared to 2400 MHz CL9-11-11-21-94-1. The "auto" tertiary timings at 2666 are pretty relaxed but the difference isn't night and day from the timings at 2400. I feel the performance deficit is too large to be the tertiaries only, I guess I will tighten them up and see what happens.