real world experiences with 16GB DDR2?

[cheechi]

Every few years I buy identical PCs for my wife and I, hers folds occasionally and mine is folding 99% of the time (as in, I rarely use that PC). The parts I retire I often consolidate and rebuild for family members but this round I have rebuilt my 'server' and was hoping to upgrade it. Currently it's folding with the following configuration

GA-MA790x Gigabyte board
Phenom II 965
1050 Ti
6GB DDR2

I have the CPU OC'd (not in front of it to tell how much, but it's been stable at that speed for 5+ years and the one folding stable) but the RAM is running DDR2 800 no OC. One of these boards is currently running a PC now, with 'only' 6GB it's keeping up with light browsing and serving older games.

I bought some 4GB modules to upgrade these boards. 8 sticks of 4GB since the board's specs list the max RAM at 16GB.
Each stick clears multiple passes of memtest 86+. Some I tested individually, some I tested as a pair. I am reasonably confident each individual stick is good. I used different slots on the board to test as well since most of their lives, only 2 of the slots were used due to the height of the (OCZ) ram's heat pipes interfering with the Thermaltake 120XP.

Any one stick the board will POST & pass the test no issue. Any pair (can't say I have tried every permutation with 8 sticks total) no issues. Any 3, seems to be maybe I had one or two failure along the way but nothing repeatable. With any 3 sticks, seems that memtest will throw errors, but not consistently at the same address. 4 sticks the board will display the post logo but won't go past it.

I remember vaguely many years ago there was a reason I set slots 3 & 4 to 195ns and 1 & 2 to 105ns, but I don't remember why now and I wound up using only 3 & 4. SPD sets it at 105. I tried all 4 at both settings didn't change POST with 4 sticks. the manufacturer's specs specify 6-6-6 timings, SPD sets it at 5-6-6-18. I set manually at 6-6-6-18. Doesn't matter which setting. However, if anyone else remembers why this setting was important back then it could be helpful because I'm pretty sure it was relevant to RAM stability back then. The ram I had before was running at 2.2v, this ram stock wants 1.8v, would increasing the voltage to 1.9v without OC the ram itself increase stability do you think?

The board has a setting for using ganged/ungagned (single or double channel) which doesn't seem to make a difference either. The BIOS has a note that their testing had better performance with single channel, and I'm pretty sure I ran it that way for many years. That setting doesn't matter that much to me honestly but neither made a difference.

If this succeeds one of these boards will be running a VM server, haven't decided on what software or all the details yet but it will be folding also. The other will be a steam in-home streaming server, may also run a VM appliance, through win 7 64 bit. Will be folding when not gaming. I hope I don't have to try mixing and run 2x8 and 2x4, I haven't tried it yet but it's my backup plan.

All testing has been done on only one of the boards as the other is in use, but I may try with that other just in case the 5+ year old board is the actual problem and the other 5+ year old board is in better shape. I'm a bit skeptical of this being the case as they are the same age and the test board is the one that has led an easier life.

TLDR Has anyone actually had a PC running 16GB DDR2 actually successful and if so what kind of suggestions that I haven't tried yet? I really hope there's some dumb thing I've overlooked and can get this working properly. Thanks in advance.

[bruce]

Foldingforum is not an overclocking support site; we deal with the specifics of FAH's software and hardware is an incidental side-effect. You probably won't find anyone able to delve into such details here.

(FAH specifically makes no recommendations regarding overclocking except do so at your own risk.) I suggest you ask your question elsewhere.

[cheechi]

Thanks bruce. Not asking about overclocking, asking if anyone has a better recollection of what steps they may have had to do several years ago with this era of system to get parts that seem to be individually working and on paper compatible to work stably enough to get a folding machine up and running.

The RAM is and will remain stock, no plans to overclock it nor increase stock voltage to it unless that actually helps it run stable stock.

The reason I asked here is there are a lot of setups that may be good enough for a certain game or application that would not be stable enough for 24/7 folding machine. As such advice from other places is usually lesser. Thanks for your help though.

[bruce]

FAH is designed to run on typical home computers (and especially on hardware preferred by gamers). FAH makes heavy use of those components so any problems that FAH contributes to are typically heat or power related. If your RAM is overheating, you might try getting more airflow or trying the heat-spreaders that are sometimes used on RAM. I'd also check that the power supply can meet the power requirements of your hardware. (no promises ... just suggestions)

[ProDigit]

If you're running any operating system from Windows Xp to windows 10, FAH requires you to have at least 1GB of RAM, but preferably 2GB of RAM (or more).
If you're using it just for FAH, and browsing, more than 4GB of RAM would be overkill, and slow down performance by a few percent.
Disable swap file, and run everything from RAM.

RAM latency timing isn't very important anymore. The most modern RAM will have lower timing, but higher Mhz.
If you do the math, eg: running 333Mhz with latency timing x,x,x,8 will have the exact same timing as running them at 666Mhz x,x,x,16.
The latency doubles, but so does the frequency, so the overall latency is exactly the same in both cases.
Just that the 666Mhz ram will be faster.

If you want to fold faster, reduce ram to max 4GB of the fastest ram sticks (and run a 32bit operating system).
Performance should average about 10% faster reducing RAM and changing to a 32Bit OS.
The fastest setting on your motherboard, would be by running 4x identical, 1GB sticks of the fastest RAM in it, at the penalty of slightly higher power consumption.

If you buy modern DDR2 ram, can run as fast as 800Mhz (dual channel, so that's 1600Mhz), and costs only a few bucks on ebay.
Your motherboard only supports up to 1333Mhz, so it'll be running the RAM at 666Mhz.

Second best option, is to run 1x2GB stick in the yellow slot, and 1x2GB in the red one.
If your motherboard is modern, it'll run it.
If it has an ancient setup of requiring 2 DIMMS right next to one another, you'll be limited to run the 2 sticks in either the yellow, or the red slots.

The third best would be to run a single DIMM of 4GB.
It's most efficient, but is slightly slower, and only works on modern DDR2 mobos. THe older ones require 2 sticks next to one another.

It's better than running mismatched RAM.
The slowest ram can actually slow down the faster RAM, which is why it's better to run fast ram only, and ditch the slow ram.
It's not going to do your system any good, running RAM you don't need. It'll only allocate more resources to the extra (unused) RAM.
But despite the faster speeds recorded in Benchmarks (I suppose maxing out the RAM at 4GB, running a 32bit OS, and use 4 slots), recording as much as 15% of performance improvements, you'll only notice a few percent (lik 1-5%) speed improvements in FAH, as it's mostly CPU based.

Fourth option to consider is to just get a new system altogether.
If you keep your system running 24/7, it probably will cost you in excess of $35/month on electricity.
Get a modern system, and it's not only faster, but might even half the bill for the same performance, or keep the same electric bill, but increase your folding capacity by a good 2-20x.

[Joe_H]

If you're running any operating system from Windows Xp to windows 10, FAH requires you to have at least 1GB of RAM, but preferably 2GB of RAM (or more).
If you're using it just for FAH, and browsing, more than 4GB of RAM would be overkill, and slow down performance by a few percent.
Disable swap file, and run everything from RAM.

...

If you want to fold faster, reduce ram to max 4GB of the fastest ram sticks (and run a 32bit operating system).
Performance should average about 10% faster reducing RAM and changing to a 32Bit OS.
The fastest setting on your motherboard, would be by running 4x identical, 1GB sticks of the fastest RAM in it, at the penalty of slightly higher power consumption.

A couple comments.

Current GPU folding cores and projects can use more than 8 GB of memory, real or virtual, during the unpacking phase and during validation passes. So if you are doing folding on a GPU having at least 8 GB of RAM is becoming essential. Disabling swap on less than 12-16 GB of RAM will result in failures due to insufficient VM.

As for running a 32-bit OS, that will limit what folding you can do. Pretty much everything folding related is headed towards requiring a 64-bit OS. At this point 64-bit is required for GPU folding on Linux, and the number of projects that can fold on 32-bit Windows is limited.

[cheechi]

It will be a 64 bit OS just undecided between 7 pro or which Linux.

This is one of two identical boards, one of which is running a pair of 2GB and a pair of 1GB for 6GB so I know that all 4 slots can be populated simultaneously. Plus that one has an old 60GB SSD separate from the boot drive with only the swap since it's too small to be useful for anything else (for me). Prior, I have had many windows PCs folding that had no page file, I don't notice any difference with this one having it on the SSD. This one that is running is doing a respectable 188k ppd according to fah control. Yes I could get better for the watts but this won't be replaced before my last Q66 based PC does.

This board colour codes the yellows as a dual channel pair and the reds as a dual channel pair. The RAM is set in the bios to run at 800MHz with the manufacturer's recommended 6-6-6 timings and although SPD would have set it tighter I manually set it instead to 6-6-6-18 since the manufacturer's site didn't specify the last I went by the SPD. There is a note in the bios the manufacturer testing had faster results with the RAM unganged, ie single channel, and the board that is running I believe I left that way. We're talking DDR2 here so it's not going to be a massive powerhouse machine no matter what I was just hoping for a little extra with a RAM upgrade.

This has been an off and on project throughout the weekend. I'm not sure if having so many RAM chips to deal with means either the NB or the RAM itself is being starved for enough voltage, so it's taken a lot of iterations to try to get more stable and I think I've just run up against a lot of false positives. It's looking like my best hope might be running 3 sticks for 12GB although I do have an order for another brand RAM to see if that may make a difference.

[bruce]

If you're running FAH on the CPU threads, some folks have recommended UNDERclocking RAM (or at least not overclocking) because they were able to boost their CPU overclock while reducing the RAM speed as for the same heat dissipation, CPU overclocking is more important to FAH than memory overclocking. YMMV.

[ProDigit]

If you're running any operating system from Windows Xp to windows 10, FAH requires you to have at least 1GB of RAM, but preferably 2GB of RAM (or more).
If you're using it just for FAH, and browsing, more than 4GB of RAM would be overkill, and slow down performance by a few percent.
Disable swap file, and run everything from RAM.

...

If you want to fold faster, reduce ram to max 4GB of the fastest ram sticks (and run a 32bit operating system).
Performance should average about 10% faster reducing RAM and changing to a 32Bit OS.
The fastest setting on your motherboard, would be by running 4x identical, 1GB sticks of the fastest RAM in it, at the penalty of slightly higher power consumption.

A couple comments.

Current GPU folding cores and projects can use more than 8 GB of memory, real or virtual, during the unpacking phase and during validation passes. So if you are doing folding on a GPU having at least 8 GB of RAM is becoming essential. Disabling swap on less than 12-16 GB of RAM will result in failures due to insufficient VM.

As for running a 32-bit OS, that will limit what folding you can do. Pretty much everything folding related is headed towards requiring a 64-bit OS. At this point 64-bit is required for GPU folding on Linux, and the number of projects that can fold on 32-bit Windows is limited.

Yes, my experience stems from the NaCl, I don't GPU fold yet.
CPU folding on both the desktop and native client doesn't use much ram at all.
The Gigabyte motherboard he has, has only 2 slots for GPU, so it's not likely he'll need 8GB of RAM.
If the client truly uses a full 8GB for GPU folding, then all you need is 10GB of RAM.
There's no reason why the system should need more than 2GB of headroom on Windows 10.

I do wonder how VRAM affects RAM in FAH?
I guess, as soon as my motherboard arrives, I will do GPU folding, and know.


Didn't know the desktop client ran better on 64bit systems? It's unusual, because most programs run slower on 64 bit systems, not faster.


I never run my systems with a swap file, because I have SSD drives. Swap files wear out SSD drives very quickly.



DDR2 ram is so cheap on ebay, you might as well buy 2 sticks of the fastest DDR2 RAM, and run it from that.
The fastest 1066Mhz DDR2 ram can be found between $2-2.5 per GB, is really not that expensive.
But like said, changing latencies is only going to affect PPD by a few points per WU. Not really worth the trouble.

[bruce]

Didn't know the desktop client ran better on 64bit systems? It's unusual, because most programs run slower on 64 bit systems, not faster.

Speed has very little to do with it. Yes, there's a slight speed difference, but the amount of virtual memory that can be addressed by the OS is a lot more important.

A protein with lots of atoms needs more RAM than a protein with a lesser number of atoms. If you run a 32-bit OS, you simply will not be assigned projects that need more RAM than that OS can process. The FAHCore has to be able to address enough RAM to process all of the atoms associated with the protein. Windows has two versions of FAHCore_xx, one which runs in 32-bit memory and another which runs in 64-bit memory (if you have it).

FAHClient runs in 32-bit memory but as soon as a WU is assigned, it will invoke whichever FAHCore that protein needs.

[ProDigit]

I think in his case it would barely matter.
Unless he's got some huge graphics cards in there, I think a system running on DDR2 RAM will more than likely be assigned the smaller units anyway, seeing that both CPU, and RAM will be a few generations old.

[bruce]

I think a system running on DDR2 RAM will more than likely be assigned the smaller units anyway, seeing that both CPU, and RAM will be a few generations old.

What information about your system do you suppose is used by the logic in the Assignment Server?

I do know that the Bit-ness of the OS DOES MATTER becuse it influences the amount of VIRTUAL RAM. I do know that the the GPU type and Species DO matter because that influences the number of ahders (and GFLOPS achieved). I know that the number of CPU threads and the instruction set supported by the CPU do matter (SSE2 vs. AVX).

As far as I know, for FAHClient, the Age of the CPU and RAM do not matter -- though I could be wrong. I don't think the Assignment Server even knows if it's SSE2 memory or something else.

The NaCl client does detect the number of CPU threads, but probably nothing else.

[gordonbb]

I was running my RTX 2070 on a almost 10 year old Athlon II processor with 1 2GB DDR2 and 1 1GB DDR2 module in a PCIe2 x16 slot with no noticeable performance difference compared to the Z370 system with 4GB DDR4 and. Pentium G5400.

For the new system I borrowed the 4GB module from another Z370 system and dropping that one from 8GB to 4GB and from dual to single channel again there was no measurable performance decrease.

On Linux (Ubuntu 18.04.1 LTS Desktop or Server) I see 2.5/4GB free on a system with 2 GTX 1060 6GB and 1.5/4GB free on another system with a GTX1070 Ti and the RTX 1070. Both systems have 2GB swap which is not used.

[bruce]

I was running my RTX 2070 on a almost 10 year old Athlon II processor with 1 2GB DDR2 and 1 1GB DDR2 module in a PCIe2 x16 slot with no noticeable performance difference compared to the Z370 system with 4GB DDR4 and. Pentium G5400.

For the new system I borrowed the 4GB module from another Z370 system and dropping that one from 8GB to 4GB and from dual to single channel again there was no measurable performance decrease.

On Linux (Ubuntu 18.04.1 LTS Desktop or Server) I see 2.5/4GB free on a system with 2 GTX 1060 6GB and 1.5/4GB free on another system with a GTX1070 Ti and the RTX 1070. Both systems have 2GB swap which is not used.

That's not surprising. FAHControl and FAHClient and even FAHCore_2* (for the GPU) use maybe 0.1 or 0.2 GB each in short bursts so the characteristics of your CPU and main RAM will see only minor performance changes. If you're running a CPU slot (FAHCore_A*) which does all of its work on the CPU you might see some differences, but what's important here is the number of CPU threads and its GHz. FAHCore_2* does all its heavy work on your GPU so the characteristics of the GPU are important, It typically appears to use one CPU thread heavily, but it's a spin-wait so it's not really consuming as much as it appears to be.

For both types of FAHCore, the amount of main RAM used will depend on the particular WU assigned but it's easily pushed out of main RAM into virtual RAM when a higher priority task comes along (and FAH is designed to run at low priority, so anything else that you run on your computer is higher priority.

Bottom line: The compute characteristics of your GPU and/or your CPU are very significant. The characteristics of your RAM are not -- except if you're seeing errors.

[ProDigit]

Seeing that a GTX 1050 runs fine on a PCIE 1x slot at nearly full PPDs, I don't think the PCIE bus on older computers are a limitation.
At least, if it's running 4x even a GTX 2070 will probably not be limited by it.
Just as long as your CPU can feed it with enough data.

Running my 1060 + 1050 together, takes just over 1 core, and 1 hyper threading core at 1,91Ghz.