I purchased a new AMD 3960x and TRX40 Auros Extreme motherboard a couple weeks back. I havent had a whole lot of time to spend messing with it, but I did so this weekend and have gotten some interesting results and was looking for others to compare to. Sadly, the availability of the CPU’s is still low, so there isnt a whole lot of comparison data out there so thought I would join up here and see if I had better luck.
First off, I should note a few things. I am coming from two 1950x systems both with Zenith Extreme motherboards. I was pretty shocked to see how different the Auros is not to mention how differently the 3rd Gen Ryzen behaves vs the Gen1 I was on before. I do have an Intel system with an Auros motherboard, so I was at least used to their crap UI. The other interesting thing to note is I am using all 8 memory slots with a matched set of 128GB Samsung B-die memory. The kit is rated to run at 2933mhz which is obviously higher than “supported” 2666mhz but I was able to get the kit to work perfectly stable at 2933mhz on my prioir Gen 1 1950x systems. The crazy thing is while I was not able to get the memory kit to work at any other speed on the Gen 1 Threadrippers, I have been able to successfully overclock them to 3400MHz and use 1:1 ratio with my Infinity Fabric @ 1700 on the 3960x! Lastly, I am custom watercooled and have the CPU loop set up to a chiller that outputs a constant 20C outlet temp. Its actually a medical grade chiller so its extremely precise…
Here are some results (originally had more pics, but I am “new” so I can only put in 2 pics apparently):
Nice! Yeah, I was surprised at the clocks I was able to get on my 3600 kit when I couldn’t get my 3200 kit (both 128) past 2933 on the 1950X or 2950X.
I have the same motherboard and cpu. I have my Samsung Samsung M391A2K43BB1-CTD ECC memory(4x) overclocked to 3600mhz and 1800 mhz infinity fabric speed. Aka 1:1 mode.
This is my result in a VM(3733 mhz, but due to ECC memory correction error down to 3600mhz should be in the same ball park):
Those are great results with ECC! WOW!
Would you mind sharing what voltages you are using for your SOC? What did you have to do to get your infinity fabric stable at that speed? I still have a lot to learn on this platform. I can get up to 1733MHz infinity fabric, and while it does finish benchmarks its not be any means stable. I am trying to figure out an easy way to determine whether the memory controller on the CPU is holding things back or if its the memory. Any idea what odds are of attaining 1800MHz infinity fabric on the CPU’s? Maybe having 8 modules is holding me back?
Well, off top my head I think its either 1.05v or 1.1v. Not sure if I can copy 1:1 from 2nd gen ryzen for SOC voltage.
As for DRAM voltage, I defaulted at 1.35V. Due to my experience again on previous platform.
As for timings, I think I stick with the default 19-19-19-36 and just crank the voltage and memory speed.
So to answer your questions, I don’t there is any benefit to go higher. Whether is limitation of your CPU or RAM. Hard to say, I am no expert on either ryzen or threadripper, nor DRAM overclocking. I assume its combination of both?
As for why there is no benefit well… here is buildzoid video with his crazy memory timings and voltages!
I think you and I are at the sweet spot.
Here are my results. Memory is 64GB (4x16GB) G. Skill FC-3200C14-64GTZ (samsung b die) running at 3400mhz 14-15-14-14-30-44 at 1.39v. Fabric is 1700mhz, SOC is stock voltage. I tried for 3466 and 1733, but it was not stable at 1.39v and I don’t really want to push more voltage for 24/7.
Cinebench R20 only gained a couple hundred points when going from XMP profile (3200-14-14-14-34-48) to the current settings.
just curious, how much power is being consumed by this setup with the chiller and all?
I will see if I can find my kill-a-watt and if/when I do I will do some power draw tests on the chiller and whole system.
Thanks for the info. From my experience, there is quite a bit of performance to be gained from higher infinity fabric speeds as long as your memory stays in a 1:1 ratio. The only other caveat with this is it will really only be measurable if you have a load that spans across more than one CCX module.
Oddly seeing 5% to 10% higher R20 scores when running G.Skill Neo 16 x 4 (64GB) 3600 16,19,19,39 at stock speeds than when running it at XMP.
R20 of over 14,000 at stock, while only around 13,500 at 3600 with fabric set at 1800, 1:1
Just posting my 3960x results on a fully loaded WIN10 Pro, with all kinds of services running at startup. This is not optimized for benching whatsoever. This has a full Adobe CC suite installed, 3 game clients, hwinfo with all sensors, a stream deck etc.
I do have some special sauce in this build, being a Liqid LQD3000 AIC. I use it as a scratch disk / steam library. I believe Wendell posted a video about it.
Wanted to get back to you as I finally found my kill-a-watt.
The chiller itself runs through various modes when its being used. The pump is continuously running, but the compressor only turns on when the coolant needs to be cooled so it runs as often as it needs to. For example, if I am using the system at the desktop and not doing any major background tasks, the compressor only turns on for about 3 mins every 15 mins or so. The compressor also has a couple of stages; it starts in a lower power mode but it can kick into a more aggressive cooling state,. The lower power state runs at 400W while the more aggressive runs between 450W and 480W. I am also running ~5 gal of coolant, and you can configure the flow rate too, so there are a lot of variables in play. If you crank the flow rate to max, the “idle” consumption is ~50W.
I mainly use this for determining if I have good contact with the CPU cooler, etc since I can control the output temp so accurately. I have used it non=stop for a few days at a time just for the heck of it, but there really isnt a lot of reason to do that.
So while this isnt exactly what you asked for, I did want to at least give you some additional info.
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.