$25 active memory cooler reduces temperatures of 'naked' RAM by up to 41% — Corsair Vengeance Airflow cooler prevents 768GB DDR5-6400 of server RAM from overheating

News
By
published

Server RAM can run hot without proper cooling.

RAM sticks
(Image credit: Shutterstock)

Computer memory is a component predominantly known for operating with passive cooling or no cooling at all. However, Phoronix reports that DDR5 server memory operating at DDR5-6000 or greater can run very hot, depending on the workload level. Therefore, server builders must consider running active coolers to prevent overheating.

Phoronix clarifies that pre-built DDR5-equipped servers already run the necessary cooling components to keep their memory under proper operating temperatures. This overheating problem is targeted at DIY server builders who might overlook RAM cooling, thinking passive cooling will be enough, as with mainstream desktop DDR5 memory.

The Linux-focused outlet discovered during testing with a Supermicro H13SSL-N motherboard in a 4U rackmount chassis featuring 12x64GB (768GB) DDR5-6400 ECC Micron RDIMMs that the chassis' triple 120mm intake fans and three to four 80mm exhaust fans were inadequate for keeping the memory at optimal temperatures. Temperatures could run above 90 degrees Celsius under load on the modules, causing all the fans to spike to full load.

Phoronix DDR5 RDIMM cooler memory benchmark

(Image credit: Phoronix)

Phoronix bought a pair of cheap Corsair Vengeance Airflow coolers on Amazon for $25 to remedy the situation. These coolers kept the DDR5 server modules cooled on both sides of the motherboard.

Temperatures dropped significantly with the active cooling modules installed. Phoronix benchmarked the coolers, revealing a 49-degree reduction in average operating temperatures during memory-intensive stress testing. The system RAM got as hot as 119 degrees Celsius without the coolers while running the High-Performance Conjugate Gradient benchmark and slapping the coolers on dropped temps to a maximum temperature of just 70 degrees Celsius.

The performance also increased due to the memory chips no longer thermally throttling. OpenVINO and NAMD 3.0 benchmarks saw massive gains, with the coolers equipped seeing anywhere from 13% to a 36% gain in performance in OpenVINO and a 16% to 51% performance jump in NAMD 3.0.

Any server-grade memory operating at DDR5-6000 or higher can purportedly run into this overheating issue. Server builders should be mindful of cooling for the memory since the new MRDIMMs clock as high as DDR5-8800.

See more RAM News
Aaron Klotz
Aaron Klotz
Contributing Writer

Aaron Klotz is a contributing writer for Tom’s Hardware, covering news related to computer hardware such as CPUs, and graphics cards.

14 Comments Comment from the forums
  • ttquantia
    What is the best way of calculating the "percentage" of reduction of the temperature? By using the Celsius, Kelvin, or Fahrenheit grade?
    Reply
  • hotaru251
    likely C which is used by most of the world.
    Reply
  • thestryker
    Kelvin, but I think it's generally better to just avoid percentages with regards to temperatures.

    As for the article DRAM cooling has been a known issue with DDR5 for a while now due to the increased frequency. It doesn't take a lot of cooling, but some direct airflow will definitely work. Most of the heat spreaders on DRAM don't do much of anything which undoubtedly isn't helping the situation. I'm somewhat surprised that there hasn't been much movement on convenient DRAM cooling solutions, but perhaps as the clocks continue to rise this will shift.
    Reply
  • JarredWaltonGPU
    hotaru251 said:
    likely C which is used by most of the world.
    But he does have a point, in that calculating a percentage drop where the base (minimum) value is -273C would become problematic if you ever had a sub-zero solution. :D
    Reply
  • JarredWaltonGPU
    thestryker said:
    Kelvin, but I think it's generally better to just avoid percentages with regards to temperatures.

    As for the article DRAM cooling has been a known issue with DDR5 for a while now due to the increased frequency. It doesn't take a lot of cooling, but some direct airflow will definitely work. Most of the heat spreaders on DRAM don't do much of anything which undoubtedly isn't helping the situation. I'm somewhat surprised that there hasn't been much movement on convenient DRAM cooling solutions, but perhaps as the clocks continue to rise this will shift.
    Since this is server RAM, I suspect part of the "issue" is that most servers have lots of airflow (and are extremely loud), so it's not really a problem. But like fast SSDs, if you run them in a zero airflow (or close to it) environment and then do a bunch of writes, they'll overheat. Even a little bit of airflow does wonders.
    Reply
  • Kamen Rider Blade
    JarredWaltonGPU said:
    Since this is server RAM, I suspect part of the "issue" is that most servers have lots of airflow (and are extremely loud), so it's not really a problem. But like fast SSDs, if you run them in a zero airflow (or close to it) environment and then do a bunch of writes, they'll overheat. Even a little bit of airflow does wonders.
    Even a basic HeatSpreader over the DIMM's would be nice, it doesn't have to be super fancy, but basic DIMMs w/o any RGB non-sense would also help as well.

    No HeatSpreader with pointless Aesthetic Embellishments should be used.

    Only basic HeatSpreaders that serve to get the job done.
    Reply
  • bit_user
    hotaru251 said:
    likely C which is used by most of the world.
    Celsius has a bias, which messes up ratio computation. In some hypothetical scenarios, you could have an infinite reduction in temperature, if you managed to cool something down to 0 C. Or, even a negative ratio, if it went below C!

    On the other hand, if you used Kelvin, the ratios would probably seem low, since most of us aren't used to dealing in K.

    That's probably why most people tend to temperature deltas, and not temperature ratios.
    Reply
  • bit_user
    JarredWaltonGPU said:
    Since this is server RAM, I suspect part of the "issue" is that most servers have lots of airflow (and are extremely loud), so it's not really a problem. But like fast SSDs, if you run them in a zero airflow (or close to it) environment and then do a bunch of writes, they'll overheat. Even a little bit of airflow does wonders.
    This.

    The Phoronix article points out that he just used a generic 4U server case with 3x 120mm intake fans, 3x 80 mm exhaust fans, and no special ducting. My hands-on experience has been mostly with Dell servers, and they've always had airflow ducts for the RAM and CPU heatsinks, since at least 15 years ago.

    In this particular build, he added a 360mm water cooler, which he put on the intake side of the motherboard, since that's the only place there's room for it. So, that pre-heated the air actually hitting the DRAM. I wonder how much better his baseline DRAM temperatures would've been, if he'd just propped up the AIO so that there was a gap underneath it, instead of above it.
    https://www.phoronix.com/review/silverstone-epyc-sp5
    Keep in mind that we're talking about as much as 500 W being dissipated upstream of the DRAM, since that's how much these server CPUs can burn. Also, he's specifically running stress tests that actually approach those limits, in some cases.
    Reply
  • thestryker
    Kamen Rider Blade said:
    Even a basic HeatSpreader over the DIMM's would be nice, it doesn't have to be super fancy, but basic DIMMs w/o any RGB non-sense would also help as well.

    No HeatSpreader with pointless Aesthetic Embellishments should be used.

    Only basic HeatSpreaders that serve to get the job done.
    Current designs are actually mostly useless RGB/aesthetics or not though there are some good aftermarket heat spreaders. The only modules with good heat spreaders that I'm aware of on the market right now are Team Group Xtreem. I would love to see DRAM manufacturers get back to having useful heat spreader options since not only does the memory IC need cooling, but DDR5 added a PMIC (and CUDIMMs have a clock driver).

    I don't think I'll ever forget Roman's (Der8auer) video featuring an old (2008) Thermaltake RamOrb DRAM cooler which had a heatpipe and fan installed on DDR5. It reminded me of when we had all sorts of random useful cooling products hitting the market (whether necessary or not) and over time we seem to have lost the useful aspect for shiny.
    Reply
  • Thunder64
    Percentages should only be used on an absolutes scale. Example, if it's 0 degrees today and will be twice warm tomorrow, what temperature would it be (silly example but makes the point). On a scale that starts at 0 would make sense. Since 0 K is not attainable you can always get an answer.
    Reply