Ryzen 9 3900x the undervolting champ?

I can totally see why people are so frustrated by this post. Dude buys an 8 core and runs it at a quarter performance for the sake of power savings. You might as well buy a 6-core and run it normally. It'll give you more per-core performance which is usually more efficient than multithreading. You'll save more energy cuz ur not powering two extra cores to begin with. You'll save money on buying the actual CPU.

The only use case where this MIGHT make sense is if he only puts it in regular power mode for intense tasks, and low power otherwise. But OP is making a point about his undervolted benchmarks under high load, which is irrelevant to that specific use case. And the power/heat savings at idle or low load is almost negligible because heat is only significantly generated during heavy load. You can understand OP's point and intended use case, but the conclusion doesn't correlate to his benchmarks and it doesn't make much practical difference in terms of power savings.
 
I am keen to get the 2 core version of Ryzen for my server as the load is very low, and the power is the one thing I need to sort. Thing is, I'm running ddr3 memory in it, so it's not a simple change
 
Thread is ironic. Buy a super smart cpu that regulates itself then disable all the smart stuff about the chip and turn it into an Intel.
 
I can totally see why people are so frustrated by this post. Dude buys an 8 core and runs it at a quarter performance for the sake of power savings. You might as well buy a 6-core and run it normally. It'll give you more per-core performance which is usually more efficient than multithreading. You'll save more energy cuz ur not powering two extra cores to begin with. You'll save money on buying the actual CPU.

The only use case where this MIGHT make sense is if he only puts it in regular power mode for intense tasks, and low power otherwise. But OP is making a point about his undervolted benchmarks under high load, which is irrelevant to that specific use case. And the power/heat savings at idle or low load is almost negligible because heat is only significantly generated during heavy load. You can understand OP's point and intended use case, but the conclusion doesn't correlate to his benchmarks and it doesn't make much practical difference in terms of power savings.

The 3900X is a 12 C 24T.

Switch for 12 and 8, and yes. I don't get it either - guy buys a $500 3900X 12c cpu only to undervolt and downclock it to the same performance and power consumption as a $320 3700X, but with reduced single core/lightly threaded performance. Why not just get the 3700X and get the best of both worlds while saving almost $200?
 
I've been working on my binning 2 3900x chips in another thread and have some somewhat interesting testing I completed on one of the chips so far. I clocked it at 4.25 all core and lowered voltage 1 bin at a time and ran 5 minutes of cinebench r20, watching scores and temps. I don't have it all together yet, but I was able to keep essentially the same score from 1.3v all the way down to below 1.2v without clock stretching. I got my load temps in the low 60s. When I have time, I'll put together a graph. The same chip can't do the same runs at 4.3 with voltage at 1.31875. 4.25 is definitely the sweet spot.
 
hot 40+ deg day here in QLD AU

started under-clocking mine @ 45w
still runs doom 2016 over 170 FPS on ultra with an old R9 290 also under clocked 850:800 900ma

all cores 1.2ghz @ 100% load ... aiming for 800 mhz and still does 4ghz single thread average load runs at 2.4ghz

and shot below 800 to the 545 floor no lower than 5.5x

PPTXX are the watts
https://openbenchmarking.org/result/1912165-AS-1802125FO69 45watts
doom 2016 ran fine no issues using ultra ~100 FPS

https://openbenchmarking.org/result/1912162-AS-1802125FO47 30 watts
doom 2016 took 10x longer to load using ultra @20 FPS while AI were alive in area and 60 FPS when the area was cleared

and I have booted with 10w !!! will bench while i sleep

65w is 3.4ghz all cores 4.2 ghz single core

oh and we get charged 55 cents per kw , so every watt saved when it can be ..... my i7 4790k lasted me years underclocked , just a shame AMD is not as easly to adjust in linux

intel is was a matter of turning turbo on and off and changing max cpu % avail

/sys/devices/system/cpu/intel_pstate/no_turbo
sys/devices/system/cpu/cpu*/cpufreq/scaling_governor
/sys/devices/system/cpu/intel_pstate/max_perf_pct set % cpu to use here

still sussing out how do to this with AMD



/sys/devices/system/cpu/power/control ???
/sys/devices/system/cpu/cpufreq/boost
( on off for boost but that does not help much too hot to boost anyhow )

actually if I get 800mhz I will call that a record because that is with 12 cores vs the 6x the 3600x
546.7 mhz record has ( ps I beat it 545mhz 12 cores 30watts :p )

I am using
phoronix-test-suite benchmark 1802125-FO-LINUX415N23
for validation

some interesting results some of the 45w benchmarks have beat stock ?


I would not go below 45w you can notice the difference below that
65w seems to be the sweet spot watt/performance
after 90w its just packing in the heat
 

Attachments

  • 3900x@45w.png
    [email protected]
    1.5 MB · Views: 0
  • 3900x545mhz30watts.png
    3900x545mhz30watts.png
    1.5 MB · Views: 0
Last edited:
I lowered my voltage on all core .1375???

Ran at stock and lowered my performance. Undervolted and now it runs faster. Whatever guess because it runs a lot cooler. If you guys are curious I can run some tests and show you the results. Fastest core is 4625mhz. 2nd Chiplet is a bum @ around 4.4ghz. What a surprise. I did upgrade to a 360mm AIO and it definitely makes a difference in temps. I was at 85c with an H100i Pro RGB, now with a Liqtech 360mm AIO, max temp is 73c.
 
Calavaro, do you by any chance able to map out performance versus power consumption to create a power efficiency curve and see if there is any inflection point (sweet spot) on whatever tasks that you can run for some time and benchmark it? I did a study on GPU here back in 2016. I'm not sure how you want to define "performance" in your case but in distributed computing perspective we call this PPD (points per day). I do run my rigs 24/7, hence power efficiency is really important to me. Other than the server chips which is designed for performance and efficiency, I'm interested in consumer chips for my next possible upgrade. The 16 core ryzen is tempting.

BTW, I've running old xeon chips (from sandy bridge to haswell). The big epyc server motherboard is quite expensive, hence I'm interested to know this, if it is available.
 
Calavaro, do you by any chance able to map out performance versus power consumption to create a power efficiency curve and see if there is any inflection point (sweet spot) on whatever tasks that you can run for some time and benchmark it? I did a study on GPU here back in 2016. I'm not sure how you want to define "performance" in your case but in distributed computing perspective we call this PPD (points per day). I do run my rigs 24/7, hence power efficiency is really important to me. Other than the server chips which is designed for performance and efficiency, I'm interested in consumer chips for my next possible upgrade. The 16 core ryzen is tempting.

BTW, I've running old xeon chips (from sandy bridge to haswell). The big epyc server motherboard is quite expensive, hence I'm interested to know this, if it is available.
I've noticed the sweet spot to be about 15% lower in both clockspeed and vCore on any Ryzen CPU (except maybe the R5 1600AF). To me that simply means AMD has overclocked the chips out the door. Same with their GPU's, obscenely overclocked as stock.
However, my workload is not folding, but more rendering and encoding.
 
Wonder what the mobo is under that chip?
For longevity, is it a x570?
FWIW I'm in the market of upgrading my 1700x to either a 3700x or 3900x on a x370 taichi mobo. It amazes me that I am able to simply plop in a 3rd gen chip on this mobo and keep the system running an additional 4-5 years. 65 vs 105 TDP is a concern so undervolting for SILENCE is right up my thermal port.
 
Crazy how long this post has been running. A whole lot of life coaching was going on an enthusiast computer hardware site. Weird. Anyway thanks for the whole thread I actually really like the stock cooler for the looks and light effects (matched the rest of my build so I'm hanging onto it as long as I can stand the heat) so i am looking at undervolting somewhat and others who jumped in to this post had exactly what I was looking for to start with. Also I'm wondering how undervolting + increasing Infinity Fabric effects temperatures. I'm hoping to maintain like 95% or more performance while keeping temps down. If I have time I might start with PPT to see how that effects things and then try undervolting next to push down temps as low as I possibly can.
 
Tried it but the additional power requirement caused hitching while gaming and it was very annoying. Maybe it was clock stretching. Never bothered to investigate. Don't be fooled by my username I've been ocing amd cpus since the thunderbird procs were new.
 
Tried it but the additional power requirement caused hitching while gaming and it was very annoying. Maybe it was clock stretching. Never bothered to investigate. Don't be fooled by my username I've been ocing amd cpus since the thunderbird procs were new.

That's doubtful about ecomode. And fyi, Zen 2 has completely different and non-similar properties to anything AMD ever made previously even to Intel, so suggesting that you've done this so long equals you know, that doesn't hold water. When ppl talk about -.1v or turning off SMT, that's it right there...
 
Bit of topic. But some of my friends drive "performance" cars. You know, like Porche's and such. I tell them I too drive a performance car, it's a Honda Civic Hybrid.
 
Bit of topic. But some of my friends drive "performance" cars. You know, like Porche's and such. I tell them I too drive a performance car, it's a Honda Civic Hybrid.
I think I see what your getting at but with AMD doesn’t speed = more cores as far as their binning process goes? So if you want to get an AMD part for gaming and you end up getting the fastest you also have the most cores even if the cores aren’t for your particular uses?
 
I think I see what your getting at but with AMD doesn’t speed = more cores as far as their binning process goes? So if you want to get an AMD part for gaming and you end up getting the fastest you also have the most cores even if the cores aren’t for your particular uses?

I was really just trying to point out the different perspectives that people might have.
 
Depends on your application. In numerous games SMT reduces performance

That's the byproduct of the whole point, reducing powerdraw to reduce temps. And, fact the difference in disabling SMT varies in games by not a lot. :facepalm:
 
FWIW the Ryzen 9 3900 pro (not X) is a 65tdp chip. Same tdp as the 3700x. Only down side to this is that it is unpossible to obtain a 3900pro.
 
Back
Top