Re: finding (or alternatives) 16GB 3866MHz DDR4 SDRAM UDIMMs?

On 4/5/2024 2:30 AM, David Chmelik wrote:
> On Tue, 2 Apr 2024 03:44:26 -0400, Paul wrote:
> 
>> On 4/1/2024 4:58 AM, David Chmelik wrote:
>>> Can you still buy 4x16GB 3866MHz DDR4 SDRAM UDIMMs, such as two G.Skill
>>> F4-3866C18D-32GTZR preferably without horrible 'RGB' LEDs (though if
>>> only option and OpenRGB can turn off, we'd use this model)?  I searched
>>> Google,
>>> New Egg, Amazon, Best Buy, Ali Express, but closest are 3733MHz,
>>> 4000MHz. We want to upgrade old PC--GigaByte GA-Z270-HD3P--with fastest
>>> RAM, which GigaByte.com says is 3866MHz... can you even run 4000MHz at
>>> all--stable at 3866MHz--or best switch down to 3733, 3666, or 3600MHz,
>>> etc.?
>>>
>>> Internet Relay Chat (IRC) gave unhelpful replies like 'use
>>> DDR5'--GA-Z270-
>>> HD3P can't--and someone saying that technology generation doesn't
>>> 'scale' well with RAM (whatever that means) but we're not getting rid
>>> of it nor using slower just because of vague/subject-changing IRC
>>> statements.
>>>
>>> Of course, if/when there are decent/better DDR5 SDRAM system-/logic-/
>>> main-/mother-board options (see my 2024-1-4
>>> alt.comp.hardware.pc-homebuilt thread asking about CPUs) I'll build a
>>> new PC (for me, not replacing GA- Z270-HD3P (family PC) so saying 'use
>>> DDR5' is irrelevant)... that is, if/ when a new DDR5 full/extended ATX
>>> system-board has plain PCI slot(s) and preferably can run powerful AMD
>>> Ryzen (7950X(3D) Zen4 or successor Zen5) or maybe Intel Core i9 almost
>>> as powerful (threads) for computer programming/science (particularly
>>> compiling large projects, which still, primarily needing larger number
>>> of threads, only AMD is best at)... entirely unrelated to upgrading old
>>> PC for the many years we plan to use.
>>>
>>>
>> https://www.reddit.com/r/intel/comments/115th3p/
> kabylake_i7_7700k_32gb_ram_support/
>>
>>    https://benchmarks.pugetsystems.com/benchmarks/view.php?id=105243
>>
>>        Intel Core i7-7700K CPU @ 4.20GHz Memory  64GB  (2x32GB)  3000MHz
>>        Motherboard    MSI Z270 GAMING M3 (MS-7A62)
>>
>> That says, basically, that there is no hardware barrier to using 2x32GB.
>> I've had boards before, they simply do not have support for the highest
>> density chips, even though the hardware is agreeable (enough address
>> bits),
>> and without support the timing settings are all wrong. For example, on
>> my 4Core board from Asrock, the thing actually supported 2x2GB RAM, when
>> even the VIA web page (the chipset maker!) reported it only supported
>> 2x1GB.
>> But without proper BIOS support, I would have needed to do 256 trials of
>> memtest, to figure out the necessary settings. Not gonna happen...
>>
>> On speed, I would stay at no more than 3600.
>> Some people have experienced "BSOD-city" at anything over 2133,
>> but that could be due to older generations of RAM chips.
>> 3600 is much easier to find, and is "standard speed" for current gen.
>>
>> What we don't know, is if the Gigabyte BIOS handles 32GB modules.
>>
>> And I don't know if any computer shop support, will run 2x32GB DDR4-3600
>> on XMP speculatively, and compare to 4x16GB double-sided DDR4-3600
>> (which uses chips having half the density and might be more agreeable to
>> the Gigabyte BIOS parsing). Any shops I've been in, their "RAM cabinet"
>> is filled with garbage RAM. They never break open packs of boutique RAM
>> and offer a testing service to identify a working module. You know,
>> one shop, they'd never heard of memtest, so I fucking well had to give
>> them a copy :-) Like, who should not be in the computer store business.
>> Now we know. The shop that didn't know what memtest was.
>>
>> It's the usual ugly situation, where the current generation DDR4 does
>> not give a rats-ass about details, and older tech might indeed care
>> about the details, as it's not quite as easy-going.
>>
>> One other aspect, is automatic adjustments the BIOS makes, and Gigabyte
>> is one of the more aggressive manufacturers (uses too much VCore when
>> MSI uses less). There can also be bumping of PLL voltage when a higher
>> speed RAM is inserted. The result is a 90C CPU, and difficulty avoiding
>> throttling. If the CPU throttles, then all that "juicy" memory
>> performance improvement, is tossed out the window like trash. These are
>> things an owner can attempt to hand tweak (as Enthusiast boards expose
>> the settings and bad choices can be overruled). The 7700K did not have
>> very good TIM inside the thing (even though the curved lid makes it look
>> like a soldered lid). It's hard to keep a CPU cool,
>> if it's being tortured on voltage (90C versus 70C, if it were done
>> properly).
>>
>> It's a multi-variable problem. And in addition, some of the details we
>> need, may not be on the RAM manufacturer web site. Even whether RAM is
>> SS or DS is not necessarily on the master spec page. And since several
>> densities of chips are available, and the RAM makers select which one to
>> used based on daily spot price... it's just a trip to the "hardware
>> casino". I can go to the computer store, buy two packs of RAM, one pack
>> has sticks with 16 chips,
>> the other pack has sticks with 8 chips (you can read out the details by
>> dumping a CPUZ text report using the last pane in the interface).
>>
>> I would try to borrow a 32GB stick, and "see what happens" :-)
>> As a starting point. If it reports 32GB, woohoo!, 2x32GB here we come.
>> 2x32GB, if it reports 64GB as expected (minus video card RAM decode
>> perhaps), the signal integrity, and ability to hit 3600 when you flip on
>> XMP, should be super-easy. I am a little more concerned about how easy
>> the 4x16GB case will be, as some 16GB sticks will be DS, and if a 16GB
>> stick is SS, it's using the same DRAM chips as a 32GB DS stick would be
>> using. This means most likely, all the 16GB more or less, will be DS and
>> using chips of half the density of the 32GB modules.
>> If you needed to sell off 4x16GB, both 3200 and 3600 should still have
>> some resale value, with the 3600 being easier to move. The box I'm
>> typing on, uses 3200.
>>
>> If flipping on XMP does not work, and your new config crashes,
>> be warned that it takes *one week* of fiddling and testing,
>> and the end result is most likely your 3200 or 3600 kit runs at 2400.
>> That XMP setting saves a lot of work... when it passes memtest.
>>
>> I'm not a RAM expert, because I don't know what all the settings in the
>> BIOS do. As a hardware designer, I know what the "turn time"
>> is, the time of flight to the module and back again. I know the
>> theoretical concepts. But some of the labeling in the BIOS eludes me :-)
>> RAM actually contains features that PC compatible computers do not use
>> (some of the cycle abort features are not used).
>>
>> As for CAS, some of the chipsets/generations have a lower limit on CAS.
>> Normally, the modules can't quite make it that low, but there have been
>> generations where indeed, you could buy a CAS3 and the BIOS would only
>> go down to CAS4. And you'd be angry as a result.
>> CAS is much higher now, and the minimum value is likely out of reach,
>> so one less thing to worry about.
>>
>> Summary: 2x32GB DS modules DDR4-3600 ? I'd gamble on that as the SI is
>> better.
>>          But if the BIOS stupidly reported only half the RAM ? Hmmm.
>>
>>          4x16GB DS modules ? Getting XMP to work may be out of reach on
>>          that generation, and then there's a week of fiddling to do and
>>          whatever the intended speed, a lesser speed will result.
>>          Current gen hardware laughs at the four DIMM case, the older
>>          systems... less so.
>>          Some people have run 4x8GB DDR4-3600 on the 7700K, but
>>          generally they did not write up the details of the experience.
>>
>>          Each case has different known-unknowns. [...]
> 
> Thanks!  So, if I still do this, I might try DDR4-3600MHz as seems highest 
> widely-available, though in fewer cases DDR4-3733MHz is also.  However, 
> I'm having second thoughts about overclocking RAM: some people 
> knowledgeable about hardware said if you do, the system-board won't last 
> as long and some capacitors go bad.  This is exactly what happened with my 
> BioStar X470GTA this year: a capacitor popped off after almost three years 
> usage, so likely I should've used fastest non-overclock RAM.  Also, though 
> I had the GA-Z270-HD3P in an online shop cart and was going to checkout, I 
> finally decided not to rather than maybe keep the GA-Z170XP-SLI it still 
> has, and maybe don't overclock that either.  I've also found some our old 
> PCs/servers could use RAM upgrades so may make another post asking about 
> these (mostly/all not type of RAM in subject line).
> 
> I don't understand all the technical terms--just some of them--but the 
> detail is pretty great.  I used to look at CAS & RAS speeds in the 1990s 
> but maybe one doesn't matter anymore and they seem to have increased 
> rather than decreased (which was happening for a while and which was 
> better).
> 

I'm hoping the calculus is pretty simple.

1) The results are gauged by error rate.
   If the CPU memory controller is really that crappy,
   the signal small, the eye closed, it will throw errors.
   If you flip on XMP and it is error free, it should
   also be (relatively) stress free.

2) The RAM does not change. At this density, there is
   likely only one basic design. The different between
   3200 speed and 3600 speed, is the fill in the SPD chip.
   The chips are still the same. And at this speed, virtually
   every DIMM passes. These aren't squidgy products with
   only 5MHz headroom.

You might find 3200 and 3600 as popular values, check your
motherboard manual for the supported speeds (it would be
comforting if the RAM XMP speed matched one of the multiplier
choices).

If you buy the 3200 the CAS will be lower. The actual
internal analog delay on the chip itself does not change.
The delay is measured in clock ticks. So the difference
between 3600 and 3200 on latency, the 3200 could be
a tick or two better on the CAS value. This means
there isn't really a difference in that respect. The
analog delay is the same.

The speed difference changes the absolute time to
transfer a cache line (say, burst of four) from
CPU to RAM. Which results in a tiny improvement
on the 3600.

If you buy the right brand, it's likely tested :-)

Puget Systems would not likely make a listing
if the thing was marginal. If you run a computer
long enough, you get some idea whether it's flaky
or not, with that combo.

I think the acceptance criterion will be simple.
Put 2x32 on the end of the bus. Good SI because
it's at the end of the bus. The loading could
only be better, if the chips were double the
density again, and it was 32GB SS. These will be
32GB DS DIMMs, 16 chips, double sided.

You plug it in. It comes up at 2133. You can run
one RAM test at that speed if you want. It should
easily pass. If you're happy with the performance,
you could leave it like that.

You can switch on XMP and test at 3200 or 3600
(as a function of what you bought).
See if the error rate is zero. If it is, leave it.

The reason for using XMP, is you don't want to be
spending a week tweaking it. But, you could resort
to that if necessary, picking a speed between
the top (XMP) value and the 2133 or whatever, nominal value.

Otherwise, just switch off XMP and go back to
eating your lunch :-)

Summary: The biggest stressor, is whether Puget is correct or not.
         That 2x32GB is supported in hardware on a 7700K.

         Select something like a GSkill. They seem to test for stuck-at.
         I doubt any manufacturer does a four hour test on this RAM.
         If you're nervous, or suspicious of the higher speed,
         just leave the XMP off.

         The trick with buying the RAM then, is making sure the XMP
         value on the tin, is a listed value in the motherboard manual.
         You get two operational speed choices, XMP off, or XMP on.

         With the CPU listing a 64GB limit, there's not much incentive
         or temptation to put 4x32 in there. That's unlikely to work, and
         while the system will come up that way, good ole Intel will
         likely sheer off the top of memory with their address decoder
         and the second two sticks would be pointless. If your video card
         is 8GB, you might see after installing 2x32, you got 64-8=56.
         That's a sign that Intel is using the address decoding as a policeman.

   Paul