The PSC Report

[redux]

Hey all,

 

You may know me from the prodigious amount of PSC DDR3 that I binned and sold over the years, maybe some strong air results, but not so much from actual BOT submissions or scores. Although I subbed frequently when I first joined the bot, and went cold a couple of times, most of my time in OC became focused on binning for the sake of better understanding and measuring the limits of what hardware can achieve, not climbing the leaderboards. Basically, I tended to be more research-oriented about OC than competitive.

 

Anyway, my single biggest project was binning PSC, and I always had this idea of creating "The PSC Report," a gigantic summary of all my findings. It took me a long time to actually put it together, but recently I finished it! I published it elsewhere, but I didn't want these findings to only be behind a paywall, so I've put it up for free here:

[Public] DDR3 Elpida_PSC 1Gbit Rev.D Overclocking Test Report.pdf

[Google Drive mirror. Please let me know if you have issues with access for any reason!]

This report details the results of research conducted into the DDR3 memory IC PSC/Elpida 1Gb Rev.D (A.K.A. "PSC" or "BDDG") in the context of competitive overclocking and binning modeling. This includes a brief history, competitive context, and demystification of nomenclature and identification. This is followed by a detailed overclocking test report, containing statistical models gathered at various points across the IC voltage-frequency curve, percentile binning targets, and analysis of the tested memory modules and the effect of characteristics such as PCB model and speed rating on their performance. We hope that these findings will inspire further research into constructing behavioral models of overclocked computer hardware with statistical and scientific rigor.

 

The scientist in me must acknowledge the report is limited and imperfect, (years late to the party ,) and just barely scratching the surface of the world of PSC overclocking. There's so much ground I could not cover, in particular systematically testing subzero results and investigating whether any robust relationship could be measured between ambient and cold.

 

I'm not the first to conduct a mass binning, but to my knowledge, no single, cohesive work like this has ever been attempted in the history of OC. Even if you're not interested in this area of legacy hardware, I hope my work at least demonstrates that as messy as it can be, it is possible to systematically understand overclocking, and you can learn a lot by doing so!

 

For the "screens or it didn't happen" folks, you can find all of my screens, for PSC and otherwise, archived here.

 

Thanks to the Memelords (you know who you are) for their support, as well as all the other overclockers I've interacted with over the years. Now that The PSC Report is done, I can officially retire

Thanks HWBOT for the space.

Edited Friday at 12:18 AM by redux
Added mirror

[Don_Dan]

PSChD awarded! 👨‍🎓

[TaPaKaH]

Using public results has quite a bit of survivor bias. Most bad results are never posted in public.

Also, binning PSC for voltage on air is more of an epeen thing - it means nothing for LN2 (other than filtering hopeless rejects) and hardly beats Samsung 2Gbit D / 1Gbit G for performance and ease of handling.

[TerraRaptor]

Thanks. Though i think some of those 1300+ 6-x-6-x scores listed in intro pages could be cold bbse (not necessarily psc).

[Luumi]

We need similar on Elpida Hypers too. Sadly the report doesn't include my new ADATA result .

[SolidStateAlchemy]

Thanks @redux, this is some standout research in the field of overclocking.

 

I do have some questions though, is there any way to "follow" a single dimm through all the tests? What I mean is if a dimm wound up in say the 90th percentile in on one test did it continue to remain in that cohort throughout all the tests?

[l0ud_sil3nc3]

Damn bro 123 pages

Thanks for doing the heavy lifting, gonna take some time to read this 🦾

On 1/24/2025 at 1:55 AM, TerraRaptor said:

Thanks. Though i think some of those 1300+ 6-x-6-x scores listed in intro pages could be cold bbse (not necessarily psc).

I could never get BBSE to run c6 twcl 6 here, only c7 twcl 6 but maybe it's possible since binning BBSE on air has never been an accurate indicator of if they will be good on LN2.

[redux]

On 2/2/2025 at 1:01 PM, SolidStateAlchemy said:

Thanks @redux, this is some standout research in the field of overclocking.

 

I do have some questions though, is there any way to "follow" a single dimm through all the tests? What I mean is if a dimm wound up in say the 90th percentile in on one test did it continue to remain in that cohort throughout all the tests?

Good question! I don't think I talked about that too much in the report for the sake of time. As a rule of thumb, sticks tend to stay around the same percentile across tests.

Not always, though -- Table 103, breaking down results by PCB model, suggests 6-layer sticks can be above average at 1000MHz CL7, but the same sticks will be well below average at 1333MHz tCL8, due to bad voltage/frequency scaling. Anecdotally, one of the lowest-voltage 1000MHz tCL7 sticks I ever had was a G.Skill Pi 1600 6-8-6 sample on ST-G3U816-B07 6-layer PCB. However, because that PCB kind of sucks for scaling, it was nothing special at 1333MHz (and needed 2T on top of that).

But for most, say, G.Skill PSC on 8-layer PCB, you can expect similar results percentile-wise regardless of test.

 

Also, added a mirror to the post.

Edited Friday at 12:48 AM by redux