Benchmarks: https://openbenchmarking.org/result/1810059-FO-MEMORY29908
This is a companion discussion topic for the original entry at https://level1techs.com/video/memory-unleashed-threadripper-128gb-2933-ecc-tested
This is a companion discussion topic for the original entry at https://level1techs.com/video/memory-unleashed-threadripper-128gb-2933-ecc-tested
6 Likes
Just watched the video and took a peek at the benchmarks. Definitely interesting and somewhat surprising results. I’ll have to keep this in mind when I see unexpected results in the future. I have a few memory intensive workloads where I had strange behaviour regarding performance. I always blamed some oddities in python (and my shoddy coding skills)…
I’ll have to do some testing when I have the time.
Thanks a lot for your work testing!
Slightly tangental question. Do you think that the sys mem --> gpu mem --> sys mem in GPGPU computing would be faster in Unbuntu? It looks like there was a very slight advantage in memory copy (except for the last test).
I liked that nod to memcpy at the end.
I’m running into something similar with UVC video devices where it doesn’t work the same across operating systems due to a design flaw with UVC HDMI capture devices related to framerate.
Enjoyed the video.
Do you have any plans to try the different RAM setups on the same CPU and motherboard? That would be a much more definitive way to demonstrate differences due to clock speed and number of ranks.
?
Both systems were 2990s, and which system had what ram was swapped. The 2950 was additional info only because it has Uma mode. Plus faster clocked but less ram.
The temptation, I couldn’t resist.
https://openbenchmarking.org/result/1810084-SK-1810059FO87
I come close to the “2990WX + GSKill 2933 128gb” machine in the “Stream” tests, which isn’t surprising as it matches my machine spec almost exactly (I have the non RGB version of the G.Skill RAM).
In the other tests, I kinda STOMP. 
I am running relaxed TDP restrictions, but by the nature of the test, it’s restricted to memory bandwidth and the fans didn’t spin up too hard during the run. It will be interesting to see how much it has to do with things (certainly no advantage in the “stream” tests).
I’ll rerun the test with default TDP restrictions and post an updated result. It would be rather sad if we can’t sustain a memory copy without hitting default TDP limits.
EDIT:
It’s as I thought, after clearing the CMOS and setting the XMP profile to 2933, I still STOMP. I guess it’s all down to OS optimizations, quite crazy really.
https://openbenchmarking.org/result/1810082-SK-1810059FO68
2 Likes
Wendell, you mentioned in the video that you had problems getting 128GB to run, but that the newer mainboards helped, because they have better memory interface circuitry.
You tested one of the oldest boards, the Fatality, and one of the gen2 Gigabyte boards with the same setup. Did you notice any significant difference? Is this why in the Linus build you used the Aorus xtreme although it’s so costly vs the Fatality or Taichi, which you previously seemed to prefer?
Ok. I didn’t understand that from the presentation. Thanks for clarifying.
Zenith extreme had problems with 128gb but I finally did get it working , designare came out later and is gen 1.5 imho.
Fatality was a higher end board but perhaps a bit of a roll of the dice from board to board.
1 Like
The way you put it all boards have issues 
But with as little choice there is on x399, I’m glad you got it working on all of them.
I have the chance to buy a 128GB kit from Corsair, for cheap. It’s on the compatibility list from Corsair for TR, but it’s 2400 CL14. Would I lose much vs the 2666 kits? It’s not ECC, can’t afford that. But I can get this one for the price of a 64GB kit of 2666.
It seems the direct line of communication is infested with crickets! No matter, here’s just fine
The “2990WX + GSKill 2933 128gb” benchmark scores from the video look very wrong to me. I’ve just installed Fedora 28 and run a number of benchmarks at different ram speeds to be sure, and, yeah, something’s not right.
I’ve removed the non Fedora 28 and non 2990WX scores for clarity.
RAM speeds tested:
2133 - Clear CMOS, load defaults.
2666 - Clear CMOS, load defaults, set memory frequency to 2667.
2933 - Clear CMOS, load defaults, set highest (2nd) XMP profile (2933).
There’s also a relaxed TDP/CnQ disabled profile for comparison.
https://openbenchmarking.org/result/1810131-SK-1810135SK04
The video loses a bit of meaning in light of the results, and it would be a shame if anyone made any purchasing decisions based on something that isn’t quite right.
Wendell, what say you?
Timings on ram? Cpu cooler? CPU OC? Your numbers seem pretty far off tho
Your drives in raid?? He is running optaine in that rig as well
The cooler is a 360mm Enermax TR4 II, so I’m in good company with suspect machine.
Timings on RAM:
2133 - 2-15-15-15-15-36
2666 - 1-14-14-14-14-34
2933 - 1-14-14-14-14-34
No OC except for the test run with my “relaxed TDP” profile, which isn’t important for the comparison. The other test runs are a case of clear the CMOS, and set the memory speed (the timings above are what the motherboard chose for the given speed, I didn’t manually touch them).
Storage setup i see 5 drives but not on how its configured
One has my Gentoo install (my normal OS).
One has a fresh Fedora 28 install (used “Network Installer image” to install, automatically bypassing AGESA hang issues with an updated kernel)
Three are empty, I had been playing with four in RAID, but broke down the array to install Fedora 28 on a single drive.
I’m not trying to outperform Wendell’s benchmarks with trickery. I’m trying to point to something that looks wrong, so am trying to match the configuration as close as possible.
just trying to get all the info for when he wakes up and responds
we shall see
That’s fine, is there anything not covered yet?
not that i can see / think of your kernal is higher but thats visible on the results
Yep, the network install image pulls in current packages. It pulls in a kernel that’s configured to work around the AGESA bug/boot hang, which was my reasoning for that install route.
Notice that the suspect machine gets slower and slower as the size of memory transfer increase, it should be going up, not down.
Edit:
My memory is 128GB G.SKILL Flare X F4-2933C14Q2-128GFX
Not running any Optane here. (These tests don’t really do anything with the drives though anyway)
Edit: Edit:
Notice the close scores between Wendell’s ECC 2666 machine and my G.SKILL clocked at 2666, many tests are within margin of error, as I’m sure they should be.