iMac 2020 Memory Testing

[RLight]

Working on a much larger article regarding fine tuning the iMac 2020, posting the memory results portion, thus far.

What's above? Micron 32GB x 2 config (Apple OEM) vs Impact Fury (fastest 2666 DDR4 1.2V SODIMMs) 16GB sticks in x2 and x4 configs with correct channels vs Mushin 32GB x 2 config (fastest 2666 when going with denser 32GB DDR4 SODIMM 1.2V sticks).

Haven't decided which is "better" yet... Mushkin 32GB x2, or x4 (thinking about ordering more) or the Impact Fury in x4 config.

Impact Fury 16GB sticks have slightly lower latency than Mushkin 32GB sticks, but there appears to be a hit on Threaded Memory going from x2 to x4 configs even though I'm using correct channels (2, 4, when doing half-populated).

https://www.bhphotovideo.com/c/prod..._kf426s15ib1_16_fury_impact_16gb_2666mhz.html

https://www.newegg.com/mushkin-64gb/p/0RM-001Z-000S7?Item=9SIA1K6JJJ1495

I don't need more than 64GB, btw. But cost isn't a concern. How fast my post-processing loads though? Yup. That’s a Bandwidth x latency fight.

These are all dual rank sticks btw, in case anyone’s curious.

The late Rob's handiwork inspired this write-up btw.

https://barefeats.com/2020-iMac-5K-8core-memory-tester.html

Rob never broke out low-latency DDR4 1.2V SO-DIMMs that are rated for 2666 and pitted them against each other and Apple stock RAM. Rather, he focused on the importance of correctly installing the memory in channels, and pairings, so the timings don't drop to lowest demoninator. Arguably the most important consideration is how you install the RAM and ensuring it's correctly paired (no mix and match as Apple will down-clock mix and match detections, which I've personally encountered in the PowerMac G5 days I should add).

What the reader is witnessing here, the iMAC 2020, does honor the Intel XMP ratings of these sticks, when using 1.2V DDR4-SODIMMs rated for 2666. Both the Impact Fury's and Mushkin Redlines are "outgunning" that OEM Micron RAM. The Impact Fury's are a touch faster in latency, as demonstrated in both the database score, and latency measured, but, the higher density Mushkin Redline 32GB's, are able to kick out more raw bandwidth. Also, the memory controller on the iMac does take advantage of more sticks as the uncached read and writes demonstrate, but, notably the latency ratchets up a bit, and the threaded memory drops, when dialing in 4 sticks (vs 2), which introduces some complexity for folks like me that don't need 128GB (4 x 32GB) where either a 4 x 16GB or 2 x 32GB kit would do.

I'm already thinking this side of my final draft the Mushkin 32 x 2 Redline kit makes more sense in both dollars per GB, not that I care, but moreover, gives me top speed, with, ability to upgrade later (I've got 2 free slots). I'll continue to stew on this one as I continue my quest.

Up next? Plan to do some more adventurous things to this iMac... Wish me luck, hopefully I don't break it.

Update: I think I’m dealing with memory interference and / or cache coherence overhead with the cached memory and threaded memory numbers dipping with 4 sticks instead of 2. This is opposite thinking from my perspective as a professional that’s deals with ensuring you “feed” say a brand new sapphire rapids Xeon with 8 channels of RAM in my line of professional work. Not so with comet lake in my personal use… All to say think twice about 128GB upgrades for this system unless you NEED it (you’re exhausting 64GB). I can now see why Apple ships these in a 64GB config with 2 sticks now. That’s not a mistake, it’s optimal in mixed use scenarios, at least for the i9 comet lake (and presumably i7/i5 configurations as well).

 

[JohnMoyer]

I have 2019 iMac with Intel(R) Core(TM) i5-9600K CPU @ 3.70GHz. "iMac (Retina 5K, 27-inch, 2019)"

I bought memory upgrades from Crucial.

I purchased it with 8GB and now have 96 GB, two 32GB sticks and two 16GB sticks.

It did not seem fast enough to me with 32GB. Even though Apple said 64GB was maximum, the chipset supports 128GB.

 

[RLight]

I have 2019 iMac with Intel(R) Core(TM) i5-9600K CPU @ 3.70GHz. "iMac (Retina 5K, 27-inch, 2019)"

I bought memory upgrades from Crucial.

I purchased it with 8GB and now have 96 GB, two 32GB sticks and two 16GB sticks.

It did not seem fast enough to me with 32GB. Even though Apple said 64GB was maximum, the chipset supports 128GB.

You can test your own memory performance with PerformanceTest by Passmark...

https://apps.apple.com/us/app/performancetest/id1560051043?mt=12

.

Best practice is not to run anything else in the background while the system is under test. Rebooting prior is recommended (free's memory being used as cache/hasn't been released), but not required. Give your system a few minutes after a reboot before running the benchmark to ensure your system completes any CRON jobs or post-boot tasks so they aren't "competing". Re-run the memory suite a couple times to ensure you're getting consistent results as there is going to be some testing variance and aggregate the results if you're not getting a consistent set of metrics.

Rob's results https://barefeats.com/2020-iMac-5K-8core-memory-tester.html from testing the 2020 iMac with mixed modules reflect Apples guidance https://support.apple.com/en-us/HT201191 of:

"For optimal memory performance, DIMMs should be the same capacity, speed, and vendor. Use Small Outline Dual Inline Memory Modules (SO-DIMM) that meet all of these criteria:

• PC4-21333
• Unbuffered
• Nonparity
• 260-pin
• 2666MHz DDR4 SDRAM"

Your 2019 iMac has the exact same set of supported RAM and requirements as mine (2020), btw, so the testing I just did, applies to your model to a tee.

.

Notably, Apple's motherboards are neither "gaming" boards (custom voltage capability to the RAM, IE toggling between 1.2 and 1.35V to the RAM channel/s), so target best RAM capable of 1.2V. The fastest RAM out there, requires higher voltages, and in my testing on gaming boards, lowest latency slower RAM vs "best" RAM at a higher voltage is problematic on non-gaming boards as you need the RAM to "present" the correct best timings to the non-gaming board. All to say, pay close attention XMP timings support (for 2666 @ 1.2V) in vendor documentation for RAM, and hence why I tested the chips I did.

I'm working on a 2020 optimization article much like the one I did for turning a Hades Canyon NUC into a "beast" that can be used if you choose for either MacOS, Ubuntu or Widows https://www.dpreview.com/forums/post/66490056/

Populate slots 2 and 4 with identical RAM, slots 1 and 3 with identical RAM for best results in the 2019 and 2020 iMacs.

Pay attention to the activity monitor's RAM stats if you're concerned you don't have enough RAM.

[ATTACH alt="From my 2019 MacBook Pro i9 w/32GB (soldered), note the Memory used, Cache Files used and Swap used. If the Swap used exceeds 1GB, your Memory used exceeds your Physical Memory? You need more RAM. Typically most people don't need more RAM if you have 32GB or more, unless you're running "Pro" apps with large workloads."]3561654[/ATTACH]
From my 2019 MacBook Pro i9 w/32GB (soldered), note the Memory used, Cache Files used and Swap used. If the Swap used exceeds 1GB, your Memory used exceeds your Physical Memory? You need more RAM. Typically most people don't need more RAM if you have 32GB or more, unless you're running "Pro" apps with large workloads.

I'm working on a larger article as I mentioned which is more Mac-centric which is going to (planned) go over default software tweaks, not just hardware, to include some things of note that will apply across the board to iMac 2020 (and iMac 2019) users, but thought I'd drop this sub-thread since the RAM is "first", I put that in quote as I've already overridden some of my software out the gate.

If your 2019 iMac is slow with 96GB of RAM, it's probably not the RAM btw. You may be interested in my other tweaks on the hardware and software which is in progress... Also, if you have a spinner (the 2019 model supports Fusion drives), that'll easily bottleneck things if you've got data on the spinner tier that can't fit on the SSD side.

 

[JohnMoyer]

Thanks. It is fast enough for now.

I bought RAM in pairs. I read the CPU specs at: https://ark.intel.com/content/www/u...-9600k-processor-9m-cache-up-to-4-60-ghz.html

I increased RAM until no swap is ever used.

When I need more speed and have money, I plan to replace fusion drive with SSD.

My Debian desktop has single socket Xeon with 56GB RAM. Most of my storage is on Debian machine shared via SMB.

--
https://www.rsok.com/~jrm/
John Moyer

 

[RLight]

Side note, I did pick up another Mushkin kit and will test with both. However, I also successfully "delidded" my iMac 2020 i9. It does result in cooler temps, however interestingly enough, the performance benefits are power-constrained as it appears there is a 150W CPU limit on the i9-10910/iMac 2020. Could be either the CPU, or the PowerSupply footprint. Both aren't exactly beefy. In theory, a delidded i9-10900K could be added to an iMac 2020, however, it may or may not take full advantage of it as Apple's known for utilizing "undervolting"; the i9-10900K is NOT, an Apple-supplied option so running it means "vanilla" power curves which ironically, probably means slower multi-core speeds vs the "slower" i9-10910 in a iMac 2020. I do suspect the iMac 2020 can honor the 5.3GHz single core requirement of the 10900K though as 150W comes under the 125W required for single core 5.3 but I suspect particularly any combination requiring more than 150W, not to be honored (power throttled down) on the i9-10900K should it be swapped. Anyhow, more to follow with a writeup and pictures, time permitting.

 

[RLight]

As suspected, the Mushkin 32GB x 2 is "king" so long as you don't "need" 128GB. Fastest Cached Read/Write, and Memory threaded is in the simplest, densest and fastest timing, configuration. Going from 2 DIMMS to 4 DIMMS introduces regression in Cached and threaded performance, so don't add more DIMMS unless you are physically out of RAM "just in case" as the numbers say you're slowing your system down.

 

[RLight]

The Atlantis "mega thread" is up here https://www.dpreview.com/forums/post/67725586 which I tested the memory on, and the memory testing was for.