Not a question, just excited to post my new hardware story here! :D


In a prior post, I posted about ARM boards for crunching.
They're working fine, though run between 1 to 8 hours per task, but the unit got plenty of threads.
I spent about $500 for 25 boards running 4 cores each at 1,5-1,7Ghz + 100 for an ethernet router + gadgetry (cables/hubs/...), comes down to $750 for the whole system.

Aside for it being a pain to set up, maintenance (each board goes offline about a few times a week), and having lots of parts, I now invested the same amount to get a Ryzen 3900X ($420 with air cooler for the CPU. $150 for the motherboard, and $40 on RAM), specially for speeding up projects like QuChemPedIA.
The 3900x seemed like the best bang for the buck at this time, and their 7nm 12 core 24 threads could potentially match or exceed my ARM server in performance; plus the motherboard allows for extra GPUs.

AMD stuff is not only efficient at 7nm,but great to buy, since a year or more from now, these CPUs will be half the price, and newer CPUs will be made, fitting their standard AM4 sockets, like they have been for many years now!


It uses 40 Watts more than my 7th gen Intel Core i5 7400 system; or 40 watts more than the E5 2650L V2 Xeon system it's also replacing (which has 2cores/4 threads less, running at half the frequency), or about 60 Watts more than my ARM server.
But it should be faster than the Xeon and Core i5 system combined.


I did some math on both higher and lower CPUs than this one.
The lower CPUs (2900 series) weren't as efficient at 12nm, so they weren't an option for me.
Besides, the price difference was only a few hundred dollars, which is minimal compared to the increased power usage of older hardware, running 24/7.

For a moment, the step below Ryzen 7 3800X had the same TDP as the 3900x, was on my list to buy,
But looking over the specs, it was running only 8/16 c/t, at a much lower boost frequency. The performance difference was just too big between both CPUs, especially for crunching data 24/7. (some benchmarks reporting a ~30% difference, for a 35% increase in price).
So I'd say the 3900x is just where you want it to be!
Plus, AMD is constantly improving their binning process, to run better at lower voltages, allowing higher overclocks.
So the more modern your CPU will be, the higher the bin will be (usually).


On the higher end, I couldn't justify the added cost of a 3950x (the next step up, which is ~$750 without cooler, $890 with water cooler, that's more than double the price of the 3900X WITH cooler) at this time.
I wouldn't recommend that chipset just right now, as it's placed on the upslope in price (meaning not much more performance, for a lot higher price.
And the 3960X is priced even higher on that upslope (you'll get twice the cores of a 3650, uses twice the power, but costs $1000, or more than twice the price).

The 3950 runs slightly lower boost clocks on all cores than the 3900x, causing both the 3900x and 3950x to use about the same power, and perform similar results (within 10% of one another).

However, the 3950 does better on overclocking, due to a higher bin (better silicon lottery).
As time goes on, the difference should diminish between the two in terms of overclockability.
The problem with overclocking is that you don't want to do this for sustained loads, simply because the system will run less efficient (meaning A LOT more power consumption, for only marginally faster performance, like seen in benchmarks below).
https://www.youtube.com/watch?v=ThUPENcqeqI

While the stock CPU runs at 105W, the 3950 on a full overclock uses around double that.


These are the reasons why I went with a 3900x.

So I had to replace the garbage cooler that came with the ryzen, and replaced it with a closed water-cooling loop.
The 3900x did about 3,1-3,2Ghz on all cores at 65W, and would thermal throttle at 105W with the stock cooler,

But with the water cooling, does ~3,5Ghz on all cores at 65W, and ~3,8 to occasionally over 4 Ghz at 105W.
Depending on ambient temperature.
I also vamped up the fan speed running at max speed all the time. The boost in performance, justifies the few Watts of energy I lose on higher fan speeds.

The 3900x did about 3,1-3,2Ghz on all cores at 65W, and would thermal throttle at 105W with the stock cooler,

The base clock speed of a 3900X is 3.8GHz. Why is yours underclocking?

But with the water cooling, does ~3,5Ghz on all cores at 65W, and ~3,8 to occasionally over 4 Ghz at 105W.

Mine runs Universe on 6 cores at the moment at 4,224 MHz, SMT in BIOS off. This with a Corsair watercooling AIO at setting silent, max temperatures 67C.

When I turn BOINC off, clock speeds on all cores drop to 3,724 MHz. If yours are lower, you aren't running the 3900X at correct base speed. You may require a BIOS update, but at least check in the BIOS if all is set correctly, and otherwise load factory settings before trying again.

---

I run 24 threads, so it's only normal the CPU frequency lowers compared to only 6 cores no ht.
Though my cooling should be below 50C, fans set to 80% (100% at 60C).

I could run close to 4,2-4,3Ghz if only a few cores were running,but a bios update I might need to look into that..

I figure 6 tasks at a generous 4.7Ghz is less work done, than 12 tasks even at a lowly 3Ghz (let alone running all 24 threads).


I found a way to get between 50-100Mhz higher clocks, topping out between 4086 and 4092Mhz on all cores, by adding additional heat sinks to the VRMs. They were running in excess of 60C. Seeing that boards of competing brands, use the same shallow vrm sinks, I would think it would benefit people to actively cool them, or paste on extra heat sinks to draw out more heat.

I run 24 threads, so it's only normal the CPU frequency lowers compared to only 6 cores no ht.

You have a base clock speed of 3.8GHz on the CPU. So when it runs idle with no Windows processes at all running, it runs at (close to) 3.8GHz on all cores, no matter how many are available, SMT on or off. But since there's always some processes running, speeds will be shown higher in Task Manager Performance.

Specially for you I enabled SMT, have 23 cores running Universe now. 1 core for GPU.
Speed is 4,048.81 MHz on all cores. Without any tricks, without overclocking. Just basic settings in the BIOS, only SMT on.
Temps are then 83C. By the sound of it, fans are full on. Quite some heat being moved out of the top of the machine. [/quote]

Off:
https://www.dropbox.com/s/cfc5gnc5hv0hijc/boinc_full_off.jpg?dl=0
https://www.dropbox.com/s/3pc4799trbkbv7x/full_off.jpg?dl=0
https://www.dropbox.com/s/1gk2hjk90urfz9u/full_off_lazy.jpg?dl=0

On:
https://www.dropbox.com/s/izrf6usruzv6fgv/boinc_full_on.jpg?dl=0
https://www.dropbox.com/s/smmglk0mjr8u6um/full_on.jpg?dl=0
https://www.dropbox.com/s/ihxlcmijvsg60fs/full_on_2.jpg?dl=0

(no admin account, thus also the too many links problem. Solved in two extra posts. ;-))

The motherboard still was quite hot around the ryzen, so despite it having water-cooling, I placed a 80mm case fan on top of it, not to cool the CPU, but the board and the VRMs (I've added stick-on cooling fins on top of the VRM heat sink).

I also updated my bios, and that seemed to give me an extra 50mhz or so.

Now I reach 4.106 - 4.155 Ghz, at 78 ambient on all cores.

The only thing I don't know yet, I have a Gigabyte x570 UD motherboard.
The ram seems to always revert back to 2133Mhz stock values, even though they are Corsair 3200Mhz rated.
My first question is, if increasing the ram speed, will increase CPU projects by a lot?

Second, ram speed Currently is set to auto. I'm having a hard time finding a basic voltage adjustment. It seems locked to 1.2xxV,but from the manual, I should be able to get up to 1,35V, and 3200Mhz.
With the current settings, ram doesn't seem to increase by anything.
Even 2400Mhz, the next logical step, gets me errors.
The voltage is greyed out. The only other control I have, is to adjust the CAS and other latency values, but I know next to nothing about it.
My module says:
1.36V,
16 18 18 36,
3200Mhz

Set the correct XMP profile (Extreme Memory Profile), this will set all the memory settings correctly for you. (Page 22 and 23 of your motherboard manual).

---

Thanks!
Turns out I had to enable xmp in the bios.
Initially CPU frequency would be higher, but now, running the same project, it seems to run at 4.042 < f < 4.101 Ghz.
So it seems to have lost 50Mhz overall.

I can only assume that the faster ram feeds the CPU quicker, and it therefore needs to lower the frequency by a bit.

No noticeable difference in crunch speed...

dmidecode --type 17 does report 3200Mhz, but at a base 1,2V.

The ram modules are handwarm, just like before, which means it somehow is able to maintain 50% more ram speed, at the same power levels.

*Edit: they're getting warmer now, seems like dmidecode might report wrong ram voltages.

*Edit 2: seems like CPU frequency is project dependent. Now I have astroids running, and it seems to tax the CPU more (frequencies down to 3.725Mhz on identical settings).

I am glad you are happy with your Amd 3900x. I was very happy with mine. After I upgraded to a 3950x I couldn't afford to keep all of them running so I sold the 3900x to a Seti Friend.

---

Tag.

Your "it".

I am glad you are happy with your Amd 3900x. I was very happy with mine. After I upgraded to a 3950x I couldn't afford to keep all of them running so I sold the 3900x to a Seti Friend.

I now run a 3950x as well, aside the 3900x.
Both of them are running at 3,6Ghz on all cores.
The 3950x, because it's in a rather warm ambient environment (80F), and because it has more cores.
While I occasionally see the 3950 blip in at 4,11Ghz, most of the time it's running at between 3650-3880Mhz.

The 3900x, I installed inside a mini ATX case. Only a single fan watercooling system fitted, so I had to reduce this one to eco mode (65W), otherwise the CPU would thermal throttle (92-94C).
At 65W it runs at a nice 75C.

My question is, on both Eco and standard mode, the speed of the 3900x is nearly identical.
Eco mode, because of voltage/power limitations.
Standard, because of hitting the thermal threshold.
I wanted to know, what setting I could set it, to increase CPU to perhaps 75-80W.
I'm sure a 10-20 extra watts will make a lot of difference in performance, without hitting the thermal limit.

the problem is that my bios doesn't accurately display VCore voltage (I think I need to adjust this?)
It shows VCore voltage of idle. But not under load.
I fear changing this value could cripple my setup.
But I do see a parabola, when I look at the performance curve (CPU Frequency) from ECO to Normal setting.
Either Eco or Normal is not a good setting in my scenario.
The best setting is somewhere between.
I use a Gigabyte Aorus pro Wifi on this system, with latest BIOS update.
There are so many voltages to choose from, and I don't know what their stock settings are, as a VCore voltage of 1,225V seems quite low.


PS: On the 3900x, the difference on setting change from 65W to 105W (which should measure 40 extra watts), I measured 100Watts at the wall.
Just running the ryzen 9 3900x on eco mode, with passive GPU, Wifi, and water pump, runs at 270W on the wall. That's quadruple the rated TDP!
Running it in normal mode, and the load registers at around 365W.
PSU is 500W, so it should handle it.

I would consider moving the 3900x into a standard sized midtower case.

If you want to increase your CPU power consumption you are going to need the most space and best cooling you can afford.

I think there is a Windows app from AMD that will help you tell what kind of temperatures your system(s) are running.

There are some for Linux too but not from AMD. My AMD 3950x is running Linux because there was a super efficient gpu app for Seti@Home that I was/am using.

From a "bang for the electrical buck" point of view the econ mode might be the very best you can do.

Tom

---

Tag.

Your "it".

I think there is a Windows app from AMD that will help you tell what kind of temperatures your system(s) are running.

I use Core Temp 1.15.1: https://www.alcpu.com/CoreTemp/, although it only shows the temperatures for all the cores in the system tray.

AMD System Monitor (https://www.amd.com/en/support/kb/faq/gpu-100) is ugly but may do the job. Although, it shows all my cores at 0MHz. Oops.

---

I've done something to my system, which left me baffled!
I had 8B of RAM running at 2400Mhz, but saw the system used about 1GB of SWAP file for BOINC.
So I upgraded it to 16GB (3200Mhz).
I figured 8GB wasn't enough for 24 threads.

I also found a flaw in the housing, allowing for hot exhaust air to re-enter the case.
Once I taped off the hole, I would have expected a few degrees lower temperatures, but not that much....

For reasons yet unknown to me, after the change, the power output on the wall has dropped from 225W (CPU in ECO mode running at 3,5-3,6Ghz, lm-sensors at 70-75C), to 182W (CPU running in eco mode, 4,15-4,16Ghz, lm-sensors at 60-63C).
Doesn't make sense that a small adjustment in ducting, and a RAM swap could be responsible for that much power drop!
Though I can guess that an additional factor could be the project I was running (World Community Grid); as it may not fully utilize the CPU, and may be the cause of the higher CPU boost speeds, and lower power consumption.

I also bet that the RAM is newer, lower power RAM, that also runs at a higher speed; is able to feed the CPU better with data, and the system now doesn't need to access the disk continuously; that this is also a contributing factor.

For anyone needing this info, it is my experience that large CPU projects, like World Community Grid, Cosmology, or QuChemPedIA need about 600MB of RAM per thread.
So a 24 threaded CPU like the 3900x needs 16GB of RAM (I see it use between 8,5 to 11GB of RAM on Linux); and the 3950x I often see eating 19-20GB of RAM (I supplied it with 32GB).
Windows running Boinc in a virtual environment, probably ill use a bit more RAM, but 16/24GB of RAM should be enough for a 3900X/3950x.

For anyone needing this info, it is my experience that large CPU projects, like World Community Grid, Cosmology, or QuChemPedIA need about 600MB of RAM per thread.
So a 24 threaded CPU like the 3900x needs 16GB of RAM (I see it use between 8,5 to 11GB of RAM on Linux); and the 3950x I often see eating 19-20GB of RAM (I supplied it with 32GB).

Tasks currently running on my boxes for CPDN are using about 1.4GB RAM/task. From memory (my own not RAM) LHC has been reported to be very heavy on RAM usage. And when the new openIFS tasks eventually come on stream we could be looking at up to 5GB/task which is what some of the tasks used in testing.

Best not run Rosetta either as their (8 hour) tasks use up to 1.7GB per task.

---

So apparently my guessing was correct,
Gamer's Nexus confirmed that for certain (older, less known) memory modules, without xmp profile configured, Ryzen CPUs can run hotter, due to the memory controller being on the CPU.

Boards using xmp profiles, best run modern, popular RAM sticks, from known and popular brands.

https://youtu.be/9IY_KlkQK1Q

If you want to know everything about memory configuration and overclocking on Ryzen Zen2, you should start reading the dedicated OCN threads.

(Gigabyte X570 AORUS Owners Thread)
[Official] AMD Ryzen DDR4 24/7 Memory Stability Thread
Ryzen DRAM Calculator (overclocking DRAM)