Safe temp for SSD

[Wilson__]

What is practical safe temperature to aim by adjusting fan speed (vs fan noise) for using Samsung 990 PRO NVMe SSD? 

My fans are powerful enough to drive temperature quite low but noise is high.  So, my question is what temp you normally aim for?

Official spec is 70 degrees Celsius https://semiconductor.samsung.com/consumer-storage/internal-ssd/990-pro/   Presumably, it is best practice to run with some safety margin.   I believe every 10 Celsius lower means half failure rate/double useful service hours.

[Lindley]

Are you using a Heatsink on the NVme , they say its better for cooling when its under load.

You could always try fitting a smaller 40mm  fan to cool  just the NVme / heatsink.

[Ranayna]

I remember this "10 degrees C lower doubles the life" rule of thumb from electrolytic capacitors.
I don't think it applies nearly that much for other components, and modern SSDs don't use electrolytics anymore. (Some older models had electrolytics as an energy buffer to enable a safe shutdown).

As with CPUs, SSDs will throttle their performance when they get too hot. There are many benchmarks showing this behaviour, where SSDs slow to a crawl when inadequately cooled.

Consider a modern gaming PC: Often, the most performant M.2 NVMe slot is directly below the GPU. If a gaming GPU is intalled, the slot is often completly covered. Modern gaming GPUs can take hundreds of watts and get hot, with hotspot temperatures reaching above 100 degrees C. And even then, it is rarely an issue for the SSD located directly under a GPU like this.

So i doubt additional cooling, except maybe a heatsink as Lindley suggested, is required. If in doubt, you can always keep an eye on the temperature of the SSD with SMART monitoring.

[Nominal Animal]

What is practical safe temperature to aim by adjusting fan speed (vs fan noise) for using Samsung 990 PRO NVMe SSD?

As the official specs say 0°C – 70°C for the 1.5Mh MBTF, I'd say anything below 70°C should not shorten its life from the 5 year or 600 TB written per GB of size.  I'd prefer 50°C myself, with heatsinks on memory and NVMes, and turbulent airflow inside the enclosure.  That way I don't need to analyze the point temps under load with a thermal camera through a thermally visible enclosure window (what I believe off-the-shelf computer manufacturers do).

The reason for turbulent airflow instead of laminar is to ensure the surface components get cooled also.  I've even used a low-speed (800 RPM quiet) internal fan, just to ensure this.  Slight overpressure (more intake than exhaust) ensures a slight pressure gradient on intake, and that helps slightly reduce dust buildup inside.

Back when we used spinning rust, people wondered about this as well, until Google found that running them at 40°C (internal reported temperature via SMART) didn't statistically reduce their service life.  Many found this surprising.  My last silent enclosure with two spinning 1G disks a decade or so ago ran the various fans in a overpressure configuration with HDD temps in the 33°C-35°C range at 20°C room ambient.  But that CPU's TDP was just 65W, overall draw on the order of 120W at full load.

[Berni]

Keep in mind that NVME drives typically report the temperature of the controller chip, not the flash chips.

By far most power hungry component is the controller chip since it is where the very very fast PCIe 5.0 lanes go into and where all of the processing happens to check ECC codes, bounce data between cache and flash, command reordering...etc So this is the chip that is actually getting very hot. It won't die prematurely even if you run it at >80°C. It just detects it is getting too hot and slowly things down to protect itself fro becoming unstable due to overheating. These things never have any built in heatsinking or even use a BGA package that allows for heatsinking (so only cooling is trough the balls), hence it can get fairly hot despite not actually burning that much power.

Yes flash degrades faster at high temperature, but the actual flash chips are often quite a bit cooler, so in most cases it is not a problem unless the M.2 SSD is mounted in an unfortunate area of the motherboard where it is near other hot components (the worst typically being powerful graphics cards)

That being said it never hurts to add some heat sinking to your SSD. Lots of gaming motherboards include a M.2 cooler cover. Otherwise you can separately buy M.2 heatsinks (typically they come in 2 pieces that hug the SSD from both sides), you can easily pick one up for online for 10 bucks.

[Wilson__]

Many thanks for the advises.  The mother board has build-in heat sink for the SSD,  A piece of 2mm thick aluminum, flat and no fin. 

Monitor software shows SSD reports its temp at three points, 46, 36, 36 Celsius.  Seem recall it was 3 chips on the SSD.

No GPU video card.  Just motherboard and then plug in CPU with video, RAM and SSD

I just added one fans. Now 3 fans.  For same temp, noise is now lower than previous 2 fans running at higher speed.

Monitor software shows temp of SSD and CPU.  Both throttle up and down by a few degrees Celsius even at 30 to 50 Celsius range.  Seem auto speed adjustment is continuous instead of on-off at 70 degree "trigger point"

I did controlled cpu load test to drive os to 100% CPU utilization. Took 5 seconds to rise to 70C peak, stayed at peak for 5 seconds then automatically drop.  It is a small heat sink for a small case.  That heat sink does not have much thermal "mass". Lesson learned.  Big heat sink could have allowed CPU stay at top speed for longer and more "output" from the expensive cpu

[BradC]

So, my question is what temp you normally aim for?

I set hard ceilings of : 70 degrees on the MB, 65 on the SSD and 80 on the CPU in my fan control daemon. I'm happy for temps to drift around but if any of those components get near their respective ceiling then the fans ramp pretty hard. There's a pretty large thermal time constant in there though to allow the machine to absorb any short load spikes without getting loud.

[Wilson__]

How long is your thermal time constant and how big or heavy is your heat sink?   My small downward wind blowing CPU heat sink is 5 seconds to jump from 3x to 70 Celsius

Could have been better for a heavy 1.5 kg unit but I can hear the fan noise level with adjustable RPM curve vs. temp in bios.  Will see how it go.


Technical Specification

    Product Dimension: 129 x 138 x 160mm
    Heat sink dimension: 127 x 110 x 157 mm
    Net Weight: 1456g
    Heatpipe: Ø6 mm x 6 pcs
    Fan Dimensions: 120 x 120 x 25mm
    Fan Speed: 500~1850 RPM±10%
    Fan Airflow: 68.99 CFM
    Fan Air Pressure: 2.19 mmAq
    Fan Noise: ≤28 dB(A)
    Fan Connector: 4-pin PWM
    Bearing Type: Fluid Dynamic Bearing
    Fan Rated Voltage: 12 VDC
    Fan Rated Current: 0.12 A
    Fan Power Consumption: 1.44 W

[bw2341]

This AnandTech article is relevant.

https://www.anandtech.com/show/9248/the-truth-about-ssd-data-retention

Apparently, high temperatures during write operations reduces stress to the tunnel oxide that holds the charge on the cell, improving data retention. The opposite is true when the drive is idle. Lower temperatures reduce charge leakage, improving data retention.

Since it's not possible to optimize the temperature for both cases, I would aim for enough active cooling to keep the controller from throttling under heavy usage to retain performance. For idle retention, I prefer hibernation or full shutdown of the computer over sleep mode if it's not too inconvenient.

[Berni]

With CPU cooling there is a different problem.

Yes modern CPUs can easily burn >150W for a speedy one. But more of a problem is that the vast majority of consumer market CPUs are handicapped by having pretty crappy thermal contact between the silicon die and the metal heat spreader. The gap between the die and lid used to be filled with special indium solder, but is now just thermal paste. It is only the huge expensive high core count server CPUs that still have soldered heatspreaders because they would overheat without it (they can burn >250W).

So what tends to happen with my CPUs is that despite having a big hefty heatsink once i run a heavy load across all cores the CPU temperature shoots up to 80°C in less than 1 second. Heatsink is still completely cold, fan goes to ridiculous speed even tho there is no heat to remove from the cold heatsink fins. But the only way to keep the cores going at the highest clock frequency is to keep the top of the CPU as cold as possible so that most of the temperature delta is across that thermal interface inside the CPU (where the thermal bottleneck is). These days running a CPU even up to 95°C is considered normal in laptops

This problem is non existent in GPU since they don't have headspreaders. If you take off the heatsink you can see the bare silicon die. Put some thermal paste directly on there, slap a watercooler on it and i can keep the GPU at only 50°C while it is burning 400W of power. If you want to go silent you can run  hot 50°C water trough the waterblock and it still stays under 70°C. This gives you nice hot 50°C heatsink fins that only need a light breeze trough them to dissipate tones of power into the ambient air.

People do go at ripping off the heat spreader from a CPU to dramatically improve thermals, but that is a rather scary risky procedure that has a decent chance of killing the CPU in the process.

[BradC]

How long is your thermal time constant and how big or heavy is your heat sink?   My small downward wind blowing CPU heat sink is 5 seconds to jump from 3x to 70 Celsius

It's a small 90mm Noctua heatsink, but even when it's wound up it's quiet. I control the fans with my own fan daemon. The case has 3 150mm intake fans, so it's built for bulk airflow. The intake fans idle at just over 100rpm and the cpu fan at about 200rpm.

Thermal time constant on the intake fans is 30 seconds unless something hits the limiter. The cpu fan ramps up quicker but is geared toward noise rather than thermal performance.

[BradC]

People do go at ripping off the heat spreader from a CPU to dramatically improve thermals, but that is a rather scary risky procedure that has a decent chance of killing the CPU in the process.

Nah, I did that years ago with a 3770k in about 10 seconds with a hammer. Piece of cake.

[Ranayna]

People do go at ripping off the heat spreader from a CPU to dramatically improve thermals, but that is a rather scary risky procedure that has a decent chance of killing the CPU in the process.

Nah, I did that years ago with a 3770k in about 10 seconds with a hammer. Piece of cake.

Try that with a modern CPU, and the CPU will be toast. Many heatspreaders are soldered nowadays.

[SiliconWizard]

There are tools to do this relatively safely. https://www.amazon.com/cpu-delidding-tool/s?k=cpu+delidding+tool
A hammer? 

[Simmed]

there are some research papers about SSD reliability
iirc 1 of them result show some detriment when SSD goes over 50C
i cannot remember the error rate
you can google it

[BradC]

People do go at ripping off the heat spreader from a CPU to dramatically improve thermals, but that is a rather scary risky procedure that has a decent chance of killing the CPU in the process.

Nah, I did that years ago with a 3770k in about 10 seconds with a hammer. Piece of cake.

Try that with a modern CPU, and the CPU will be toast. Many heatspreaders are soldered nowadays.

Some were back then also. Oddly enough you didn't delid those ones because the solder is an effective heat transfer medium. The factory TIM under the IHS wasn't.

There are tools to do this relatively safely. https://www.amazon.com/cpu-delidding-tool/s?k=cpu+delidding+tool
A hammer? 

Yep, those tools were developed after I demonstrated that an IHS could quickly be removed by shearing the adhesive rather than half an hour with a razor blade.

[Nominal Animal]

By the way, it is very easy to control 12V PWM fans using a microcontroller.  The PWM output is nominally 25 kHz, although anything between 21 kHz and 28 kHz should work absolutely fine.  You can find the official Intel specifications here.  To interface safely to the fans, I use two NX138AKR N-channel MOSFETs in SOT23 per fan (one for PWM, the other for the tachometer return signal, two pulses per rotation); they're similar to ubiquitous BSS138 MOSFETs, but slightly faster due to lower total gate capacitance, and work fine down to 2.5V logic.  At 2.5V, the Rds(on)≃10Ω, and the gate can handle ±20V wrt. source, so it also works as a "safety" against harming say 3.3V MCU I/O pins.

I mention this because if you want a silent enclosure, and you have multiple fans, you'll want to keep their frequencies apart or slightly varying, to avoid "beat" and enclosure resonance.  For example, Noctua already provides pairs of fans for their CPU heatsinks where one is configured to be slightly slower than the other at the same PWM duty cycle, exactly because of this (although there are also other benefits in a push-heatsink-pull config like that).  Another option is to use different fans: on intake (with filters), ones with better pressure, and on exhaust, ones with more flow, with something in between for any attached to heatsinks.  If you use all same fan, expect suboptimal results!

The older-style 3-pin 12V fans are voltage-controlled.  You can drop the voltage using a diode, or by using suitable resistance.  As a typical fan draws on the order of 100mA at 12V, and run from something like 6V upwards –– these obviously varying from fan to fan! –– you can control their speed using a 1W 100Ω potentiometer.  They're not nearly as nice to control from an MCU.

[Wilson__]

Interesting that you write control daemon.  3 items I did not do.  May be I can try.

a) I believe temp is by reading some of the Linux virtual file. 

b) There is a software driver/module named GPIO or similar on Raspberry Pi which has GPIO pins.  How and where you get a GPIO pin on pc motherboard?  Some bigger m-ATX boards have the original pc parallel printer port and have a dozen pin, some in some out.  My mini-ITX board does not have old printer port. May be those DTR, etc. of old serial port can be used?

c) how to 'convert' a normal user space program to daemon?

[Nominal Animal]

You didn't ask me –– it was BradC who wrote about a control daemon in #6 and #10 ––, but I prefer to use microcontrollers with native USB interfaces instead.  Control is typically via USB Serial (USB CDC ACM) or USB HID, often with a physical knob (rotary encoder) for me to select between Quiet and Cool.  This way, as long as the MCU has power, it will control the fans.  One can even read the PWM duty cycle and fake tachometer outputs for motherboard fan connectors.

In Linux, typically your built-in sensors are wired to lm-sensors, so you can read them by running sensors (or sensors -j if you want it in JSON format).  If storage devices support S.M.A.R.T and have temperature sensors, make sure you have the drivetemp kernel module (5.6 and newer) loaded so that lm-sensors will report those too.  Similarly for the jc42 kernel module for JEDEC JC 42.4 temperature sensors used on many DDR3 memory modules.

Hwmon sensors are exposed as pseudofiles in /sys/class/hwmon/hwmonN/; the /sys/class/hwmon/hwmonN/tempM_input contains the temperature reading in degrees milliCelsius, i.e. 1000 = 1°C, 2000 = 2°C and so on.

[Berni]

I actually have a project planned for doing fan control using a MCU, but i haven't gotten around to actually building it.

My main problem is that i got watercooling and i would much prefer the respective fans to be controlled by water temperature or air temperature, some fans cool other stuff like VRMS...etc But at the same time id still want ability to select different fan modes (like for casual use, or for when number crunching at 100% load for hours on end)

The plan is to connect a MCU as a CDC serial port to one of the motherboards USB headers. This way the fans work normally regardless of the host OS being up, but it can be controlled from the OS using a simple app.

[Nominal Animal]

My main problem is that i got watercooling and i would much prefer the respective fans to be controlled by water temperature or air temperature, some fans cool other stuff like VRMS...etc But at the same time id still want ability to select different fan modes (like for casual use, or for when number crunching at 100% load for hours on end)

The plan is to connect a MCU as a CDC serial port to one of the motherboards USB headers. This way the fans work normally regardless of the host OS being up, but it can be controlled from the OS using a simple app.

Yep, exactly!  There are many quite sensitive I²C temperature sensors one can use, too.  For water, you could even use a 3-wire Pt100 sensor via MAX31865: there are many stainless steel and copper "stud"-type ones with M3 and M4 threads, for direct measurement in the water pipe (using e.g. a modified coupling, adding the threaded cross-hole for the measuring stud).

I particularly like control schemes that control the RPM and not just the PWM duty cycle.  That way, as the fan ages or whatever, or there are changes in airflow near the intakes/exhaust (something like window open vs. closed during the summer), it will not affect the actual fan RPM.  For temperature sensor based schemes, you of course control the temperature; but even so, during e.g. quiet startup/wakeup, it is nice to use the tachometer output to verify the fan has started (at sufficient speed) rather than do the one-second full-blast pulse (at zero or 100% PWM duty cycle), just when everything else is starting up too.  And, if you do have multiple fans in push-pull configurations, you can control them as a group, with RPMs suitably apart, via RPM, based on temperature changes, to avoid "beat" and resonances and maximize the group's effectiveness.

As to MCUs, there are lots that work just fine, from AVRs to ARM Cortex-Ms, to RISC-V. I even have one- and two-fan controller designs using ATtiny85 and a linear pot for adjustment, for non-computer devices.  ARM Cortex-M3/M4/M7 (ARMv7e-m) are nice in that they have the ARM_DWT_CYCCNT cycle counter for accurate measurements for the tachometers.  At two pulses per rotation, around 1000 RPM the pulse frequency is just ~ 33 Hz.  I like a scheme where I count cycles over a number of tachometer pulses, adjusting the number of pulses so that the measurement time is approximately half a second to a second, depending on how "tight" you want your control loop (considering thermal inertia of the system).  The rotation speed is then clockrateconstant×pulses/cycles, noting that this calculation needs only be done 1-2 times per second per fan.

[Wilson__]

Many thanks from advises.

May I have a list of heat sink, fan brand/model. fan power, RPM, size in diameters, etc.  that you use?

How many fans  sucking/blowing air in/out of the case?   

How many fans circulating air inside the case (I find this mode is very efficient in heat removal from CPU while little noise is induced)

What is your heat time constant?   After giving all CPU/thread  100% load, how many seconds it takes for  temperature rise to 70 Celsius and CPU slow down to protect itself.

I did not tried them in yet.  On market, there are 'super fans', 10 times higher DC input power.  Really strong as compare with normal computer fans.  Real ball bearing, UL safety listed .  I wonder if they are made for professional 1U, 2U rack computer blade server or they are aimed for small handheld vacuum  cleaner?  I wonder if It may work well blowing extra strong wind at the metal plate where heat sink meet the CPU top plate.

[Nominal Animal]

I'm almost a decade out of date; I built my last silent workstation almost a decade ago now.

My favourite fan was 120×120×25mm Scythe Gentle Typhoon, aka Nidec D1225C12B5AP, but I've also used various Noctua 80×80×25mm and 120×120×25mm fans. Noctua NF-A12×25 PWM is considered the best all-around fan right now.  For CPU heatsinks, I like huge Noctua towers (NH-D12 and NH-D15 series).  Again, not up to date on what's best right now.

For "silent" operation, I run these fans at 500 RPM to 800 RPM.  The larger the fan, the better: the area of the airflow is larger, so that for the same volume of air moved, the velocity is lower.  Smaller than 80×80mm fans often whine –– often simply because of the velocity of air in the small cylindrical area –– to move useful amounts of air, so I like to design my cooling around larger fans.  I only trust reviews so much; it all bogs down to testing different fans in the same setup, to see which works for me best.

I didn't build my cases scientifically, and instead approached it empirically.  I used the largest heatsinks and fans I could fit, then scattered lots of thermal sensors inside the enclosure, and measured the temperatures as I reduced the fan speeds.  Today, I'd definitely get a thermal camera for this, using something that is transparent at the 3µm – 14µm wavelengths for one side of the enclosure; PMMA and Plexiglas are not that, they're quite opaque at 3µm to 25µm.

If you use an application to control your fans, you can have them predict extra thermal load from increased CPU and I/O load.  Other than that, having enough thermal mass (which is just mass for pure-metal heatsinks) smooths out any sudden spikes.  You'll want a thermal sensor at the base of your CPU heatsink (assuming a large tower-like heatsink with heatpipes), so that when it shows an increase in temperature, you act on it immediately; whereas for ambient temperature changes, you react more slowly.  I know of only empirical tuning of this, i.e. adjusting the rates at which the controller reacts to increased thermal loads, and recording the ensuing temperature curves; and changing until it satisfies your needs.

This is also exactly why I suggest USB-based DIY fan controller: it won't crash even if your CPU does, but you can still easily create a small UI widget to switch between "be quiet" and other modes like "number crunching" where you want the fans to be more aggressive.  For logging the temperatures, you can use a microSD card socket, either via SPI or via SDMMC if the MCU has that peripheral.

[Berni]

PC cooling doesn't just have one magic solution that fits all, especially if you want to be silent.

Different PC cases and PC components direct airflow in different ways. So the combinations of them can interact in many ways.

But in my personal experience CPU cooling is the easy part. The normal consumer CPUs don't really burn that much power(~ 60 to 120W) and most of the thermal resistance is inside the CPU package itself. So better cooling won't really reduce the actual CPU die temp that much. Just stick on the largest heatsink you can find and run the fan at a speed that keeps the fins at a reasonably slightly warm temperature. Orient the fan in a way that it blows the exhaust towards a case fan that blows it out and that's it. If you want really good CPU temps then get a AIO watercooler and stick the radiator directly on an intake to get as cold air as possible(helps maximize the delta over the thermal bottleneck in the CPU package)

The tricky part are powerful graphics cards. These things all burn more than 200W and go up to 600W for the big ones (or even more if you go into overclocking). These blast a ton of heat into the case and often blow it in all sorts of directions. It takes experimentation to figure out how to move that out of the case without being loud. This was my reason to go into watercooling because you can stick a huge radiator on a large exhaust surface of the case, making that heat instantly leave the case, makes the case interior much cooler and the GPU much cooler.

The last thing is aircooling random low heat generating components such as VRMs, RAM, M.2, chipsets etc.. For that i like to strap a slow spinning 120mm fan on that area(directly facing the motherboard), it doesn't need to move much air to significantly improve temperatures of the components since it just needs to push away the stale air in the area. Smaller 80mm fans are fine for this too as long as they are ran slow.

For a PC fan controller with USB i am surprised nobody has made one as a open source project that i could just build. It is easy to make one, it just takes a fair bit of work to design a PCB and write the software. I like to bootstrap myself off existing open source projects to reduce the amount of work for me (since it is a hobby project and i have a day job)

[BradC]

PC cooling doesn't just have one magic solution that fits all, especially if you want to be silent.

True. My biggest issue was building an 18mm marine ply adapter plate for the 3x150mm intake fans. I drive all the fans using the multiple onboard PWM (6 of which I use 3) in a loose PID control loop, but the 3 intake fans are all in parallel so might suffer from the beat issues Nominal Animal states. Having said that, given they idle at less than 200RPM, the point is mostly moot.

I don't have a GPU though, so it's all pretty low power.