Upgrade to 5900X

[kapone32]

I'm not sure what sorts of "intensive" games you play where the 5900X always draws less than 50W, seems a little exaggerated. With a 2060 Super and 3070 Ti I've never seen less than 50W package power in any game. 50W is honestly pretty ambitious for demanding games even on a single-CCD (3700X) or APU (5700G).

~60W is about the norm for the lightest games, most settle around 90-100W, an increasing number of games will regularly run up to the 125-142W area. DCS, MW2019, BFV, just off the top of my head......

I have a 6800XT so that could be part of it. When I am playing a 4x20 x2 unit (Ultra) battle with TWWH3 Immortal Empires I can see up to 100W CPU power draw the same can be see with High Risk Areas in Everspace 2. When I am playing Pacman 256 or Raiden Legacy or even Torchlight I don't see past 40 W on the CPU. I do agree with you though that new Games do push the CPU (And GPU) to run as high as possible.

 

[Zach_01]

Was just testing FarCry NewDawn. With GPU capped at 60FPS the 5900X was by average at 100W. Without GPU cap around 110W.
Boost both times up to ~4.95GHz for best cores and 4.85~4.9GHz for worst.
CPU thermal limit at 75C, ambient 27~28C. Boost override +50MHz
System power by avg 100W difference. (+10W CPU, +80W GPU, +10W the rest I guess)





With same PBO settings (see specs) and CO (negative 7~15) on MT max temp is below 70C.

If you want to control temp, its different for MT and ST/middle-theaded loads.
For MT you limit PPT/EDC to the desired level.
For ST you cant really do anything else than limit temp.

 

[tabascosauz]

Was just testing FarCry NewDawn. With GPU capped at 60FPS the 5900X was by average at 100W. Without GPU cap around 110W.
Boost both times up to ~4.95GHz for best cores and 4.85~4.9GHz for worst.
CPU thermal limit at 75C, ambient 27~28C. Boost override +50MHz
System power by avg 100W difference. (+10W CPU, +80W GPU, +10W the rest I guess)

View attachment 261949

View attachment 261950

With same PBO settings (see specs) and CO (negative 7~15) on MT max temp is below 70C.

If you want to control temp, its different for MT and ST/middle-theaded loads.
For MT you limit PPT/EDC to the desired level.
For ST you cant really do anything else than limit temp.

There is always the option of setting negative boost clock override to reduce ST clock, but I fully agree - no reason to do so on a Dark Rock Pro 4. ST temp spikes won't be anywhere near concerning, MT temps will be fine at 142W.

 

[Zach_01]

There is always the option of setting negative boost clock override to reduce ST clock, but I fully agree - no reason to do so on a Dark Rock Pro 4. ST temp spikes won't be anywhere near concerning, MT temps will be fine at 142W.

Yeah there's that too, to cut down ST boost but I find it less optimal performance wise.
I believe that by positive boost override + thermal limit you're getting a more sustained ST boost that can potentially improve on lower ambient - better cooling.

 

[freeagent]

5900X @ 142w is really not that bad at all. To put it in perspective, my 5600X under full power limits, and an aggressive curve running Linpack can do ~135w PPT @ 4600MHz. !42w on 5900X should be about 55-60c.

 

[tabascosauz]

5900X @ 142w is really not that bad at all. To put it in perspective, my 5600X under full power limits, and an aggressive curve running Linpack can do ~135w PPT @ 4600MHz. !42w on 5900X should be about 55-60c.

Yeah, but thats just because you leave your windows open in the winter FC140 with fast fans is a bit more capable than DRP4

 

[Zach_01]

5900X @ 142w is really not that bad at all. To put it in perspective, my 5600X under full power limits, and an aggressive curve running Linpack can do ~135w PPT @ 4600MHz. !42w on 5900X should be about 55-60c.

135W with what PowerReportingDeviation?

-----------------------------------------------------------------

CB R23 run (10min). Readings steady after 4~5min
Ambient 28C
Cooler maxed (5+year old H110i 280mm)

CPU PPT ~142W (true avg)
CPU temp 67C avg,
~4.4GHz

 

[freeagent]

135W with what PowerReportingDeviation?

Have a look

 

[SpittinFax]

Efficiency falls off a cliff at higher clocks, in other words a CPU will use twice the power to run 20% faster.

If you want multicore performance without silly power consumption, buy more cores and tell them they can only have 142W. You can absolutely run a 5950X with a 120mm air cooler and get great performance, it's just that you won't get 4.4GHz all-core at the base TDP of 105W (142W boost)

I have a bunch of 5950X rendering nodes that are cooled with dual-fan NH-U9S 92mm coolers and they render animation frames in VRay at 3.5GHz or so all day without breaking a sweat. Just because you can run them at 280W with insane cooling to get another 900MHz doesn't mean you have to. Modern CPUs and boards let you choose how much power you want to use, and a 5950X is remarkably impressive in eco-mode (65W) and will probably still get close to a 5900X drawing 4x the power.

That's nuts. And that's what I like about the older Xeons chips, they're normally downclocked so that they can churn through productivity work all day. The E5 2678-V3 runs so cold that even the cheapest tower cooler is enough. Tuning a 5900X in a similar way would limit performance a lot but I'm not too worried about leaving performance on the table if it means better efficiency.

Dark Rock Pro 4? You'll be fine. C14S is a step down from DRP4, was like 2-5C hotter on my 3700X, and I've been running fine all this time. U9S should still be okay even, maybe at the edge at 142W in hotter workloads.

At stock, a 5900X in MT behaves similarly to a 3700X stock. Just with higher and more unpredictable ST spikes from higher ST boost. 4950 is the ST cap out of the box, if you're lucky PBO can take you up to 5150 but that's strictly dependent on your CPU sample. Have seen 5900X that max out at 5150 out multiple cores, have seen 5900X that struggle to reach 4900.

L12S and lower is where the 5900X starts to fall apart (without a cool ambient and case airflow that is overwhelmingly geared towards helping it out).

Better multicore always comes from having more cores, OCing a 5800X to beat a 5900X or a OC 5900X to beat a 5950X is extremely tough and pointless. The 5700X may look efficient with its 76W PPT but once you start pushing it there will be little difference to the 5800X. Get the single CCD parts over 100W and you start to feel the heat.



I'm not sure what sorts of "intensive" games you play where the 5900X always draws less than 50W, seems a little exaggerated. With a 2060 Super and 3070 Ti I've never seen less than 50W package power in any game. 50W is honestly pretty ambitious for demanding games even on a single-CCD (3700X) or APU (5700G).

~60W is about the norm for the lightest games, most settle around 90-100W, an increasing number of games will regularly run up to the 125-142W area. DCS, MW2019, BFV, just off the top of my head......

ST on the 5600X is pretty good so I could just downclock the 5900X to 4.6GHz. For comparison my 5600X hovers around 40W PPT while gaming so it has very good efficiency under those loads.

Too bad it's not possible to have the single-core efficiency of a 5600X with the multi-threaded performance of a 5900X, but I guess there has to be a compromise at some point.

 

[tabascosauz]

ST on the 5600X is pretty good so I could just downclock the 5900X to 4.6GHz. For comparison my 5600X hovers around 40W PPT while gaming so it has very good efficiency under those loads.

Too bad it's not possible to have the single-core efficiency of a 5600X with the multi-threaded performance of a 5900X, but I guess there has to be a compromise at some point.

The cores are the same, maybe better binned if lucky.

All round the 2CCD parts will just draw more power, more silicon to power and more losses. As usual, SOC power scales with VSOC so especially around 3600 you can probably drop VSOC quite a bit (~1.02V for me vs. 1.11V at 3800), but you just won't match 1CCD SOC power draw. In turn, 1CCD doesn't have a chance in hell of matching APU SOC power draw. Just chiplet things.

SOC doesn't tell the whole story either; on 2CCD there's also a fair bit of power always lost to minor rails. Minor rails are less accurately measured/estimated iirc so there's always something like 10W overhead when subtracting CPU+SOC from Package power. 1CCD is better by a few watts (whereas overhead on APU is negligible) but again, just chiplet life. It makes Zen 3's overall efficiency look even worse than it really is.

With a decent cooler, SOC power doesn't contribute much to overall thermals so it doesn't matter that much. All of the significant heat and spikes come from CCD. Towards the high end of Fabric speeds you'll be looking at 15-20W regularly, maybe peaking slightly above 20W in extreme UMC load (ie. TM5).

There is not always correlation between core clock and per-core power. If you try a variety of different games/workloads you will sometimes see extreme power draw and blistering heat out of surprisingly low ST clocks (ie. 18-21W), in other workloads you may see crazy high 5.0GHz+ at average temps and barely over 10W per-core.

 

[Zach_01]

Have a look

View attachment 261954

I cant really look as you have it hidden... lol
Power Reporting Deviation must be observed during the run. After the test finishes I would never know what it was just by looking a screenshot like the above.
Also you dont have Snapshot CPU Polling enabled like you should and It would be nice for avg values to be visible too.

 

[SLObinger]

I went from an R9 3900X to a 5900X and it was night and day difference especially in single threaded workloads. PBO works really well with Zen 3. I would definitely go for that upgrade

 

[freeagent]

Power Reporting Deviation must be observed during the run. After the test finishes I would never know what it was just by looking a screenshot like the above.
Also you dont have Snapshot CPU Polling enabled like you should and It would be nice for avg values to be visible too.

I didn't think PRD was important that is why it's not visible. I don't have the average visible because it drops as soon as the load is off, and you have to be right there to snapshot it or else it is just wasted time.

 

[Zach_01]

I didn't think PRD was important that is why it's not visible. I don't have the average visible because it drops as soon as the load is off, and you have to be right there to snapshot it or else it is just wasted time.

Yes that is what I was looking for. Because the 135W can be something else completely when PRD is taken account. It is important on 100% MT.
It can be truly 135W or 110W or 150W…

If anyone wants to learn something more about their system they should start the bench/stress test, reset HWiNFO values and take the screen shot right before test is over. Anything else is just wrong. You can’t conclude anything by seeing lows/highs

 

[freeagent]

I cant really look as you have it hidden... lol
Power Reporting Deviation must be observed during the run. After the test finishes I would never know what it was just by looking a screenshot like the above.
Also you dont have Snapshot CPU Polling enabled like you should and It would be nice for avg values to be visible too.

Ok, this is what you mean, right? Snapshot polling, PRD, and the averages... I let TM5 run for a couple of hours

 

[tabascosauz]

Ok, this is what you mean, right? Snapshot polling, PRD, and the averages... I let TM5 run for a couple of hours

View attachment 262033

TM5 load and clocks vary a lot, maybe a dedicated CPU test works better (dunno about Linpack, also kinda variable). TM5 doesn't even run true all-core at all times I think

As long as you're there observing when the CPU is relatively consistently at 100% load, the number you see should be reliable

 

[Zach_01]

Ok, this is what you mean, right? Snapshot polling, PRD, and the averages... I let TM5 run for a couple of hours

View attachment 262033

Is TM5 loading this much the CPU? Haven’t run it for some time now and I’m not close to my system now.

Nevertheless…
At first sight someone could say that you are beating the crap out of that 5600X. With power up to 170W and 180A. For a single CCD CPU this is killer readings. Not in the good way…
Taking account the PRD though at 130~135% things are changing drastically.

135W PPT with 132% PRD means that the actual power of the CPU (PPT) is 135/1.32 = 102~103W.
I find it ok for a 6core CPU as my R5 3600 run about the same but with lower current(A).
Applying the same to the 180A…
180/1.32 = ~136A
I’m not sure this is legit though to calculate true current(A).

I wouldn’t like my CPU to sustain this kind of current through it, even a 140A on a single CCD.
Personally I don’t care for 100% load MT boost so I keep it under the default limit at 125A (from 140A) on the dual CCD 5900X.

Also I’m not sure that any Ryzen CPU needs 140A let alone more, for “just” 100~105W.
This looks to me like it’s on the inefficient side.
High current(A) with not so high power could mean “unnecessary” voltage (?).
Are you using PBO Scalar beyond the default (X1) to override health management?

I assume also (by PBO limit readings) that you’re using motherboard PBO limits?
If yes, then boards are stupid enough to increase all PBO limits to the max available.
Another insufficiency…

Maybe you should try to limit current(A) more. If you care about CPU longevity.
I know that temp is not all that high but this is not really a CPU test, right?

Curve Optimizer settings? Max (30)?

I’d like to see it with something a little more CPU oriented test like CB R23. And it doesn’t have to be for 2+hours.
10~15min suffice.

 

[freeagent]

Sure I’ll run it. My first pic was 5600X and the second was my 5900X

 

[tabascosauz]

Sure I’ll run it. My first pic was 5600X and the second was my 5900X

Sometimes cinebench doesn't max out PPT (hence why it runs so cool at stock) but CPU-Z Stress option is pretty stable and should be reasonably accurate

 

[freeagent]

Sometimes cinebench doesn't max out PPT (hence why it runs so cool at stock) but CPU-Z Stress option is pretty stable and should be reasonably accurate

Oh yeah it does not come close to maxing PPT, I just use Linpack Xtreme because it will max out PPT pretty easily. A half hour of R23 and all cores just run at 4600 the entire time with temps in the mid to upper 70s with a 20c ambient, with Linpack they run at 4475-4525 with loads at 80-85ish depending on my ambient. One thing I like about 5900X is that it can be as tame as a kitty cat, or be a fire breathing dragon depending on how much you want/can feed it

I can set 240 PPT but will not get more than 235.5w. That must be a coded limitation within the depths of the board or CPU somewhere. But still, 235w PPT is pretty intense..

 

[tabascosauz]

Oh yeah it does not come close to maxing PPT, I just use Linpack Xtreme because it will max out PPT pretty easily. A half hour of R23 and all cores just run at 4600 the entire time with temps in the mid to upper 70s with a 20c ambient, with Linpack they run at 4475-4525 with loads at 80-85ish depending on my ambient. One thing I like about 5900X is that it can be as tame as a kitty cat, or be a fire breathing dragon depending on how much you want/can feed it

I can set 240 PPT but will not get more than 235.5w. That must be a coded limitation within the depths of the board or CPU somewhere. But still, 235w PPT is pretty intense..

You don't have to run high PPT. For the purposes of determining power deviation you can (and maybe should since 200W+ PBO is so wonky nowadays) just run completely stock and look at the number during all-core

 

[lexluthermiester]

Sure I’ll run it. My first pic was 5600X and the second was my 5900X

Sometimes cinebench doesn't max out PPT (hence why it runs so cool at stock) but CPU-Z Stress option is pretty stable and should be reasonably accurate

Prime95 64bit.

 

[tabascosauz]

Prime95 64bit.

Small FFT doesn't max out PPT anymore on stock Ryzen. Boost algorithm for Vermeer and newer CPUs basically recognizes Prime95 as a power virus and automatically throttles down to exactly base clock (3.7GHz), only draws about 130W.

That said, power deviation metric still appears to be accurate in Small FFT.

 

[DebenPoison]

Talk to me friends ~ what should I be enabling, disabling or manually adjusting in my BIOS to achieve better PBO results on this 5900X?

Currently all I have is everything default other than DOCP @ 3200mhz for memory ; and 3 x PBO enabled settings in BIOS. Enabled seems to grab an extra ~100mhz vs Auto. I did see an increase max cpu mhz PBO related setting with a limit of up to 200mhz but that is set to auto at present.

Running the Asus TUF X570 Wifi motherboard with a fairly current BIOS (2021), but there is newer available. Also running a Cooler Master ML240L V2 AIO which I'll be replacing soon with a 360mm Artic Liquid Freezer II A-RGB in the next couple weeks. Current thermal paste is MX-2.

I have set Thermal limit in bios to auto perhaps this should be set to 80'C or more? Also I note the voltage looks lower than it could be 1.288V CPU Core VID and SoC Voltage seems low compared to the above screenshots. Are those being set manually or do we typically leave them auto ?

 

[Mussels]

"and 3 x PBO enabled settings in BIOS"
What does this mean?


The Tuf x570 is one of their weaker boards, performance is fine but they cut corners. Watch your chipset temps because if it's like my x570-F, the thermal pad they used will have dried up and gone to shit long ago


CPU-Z is not a good benchmark or stress test on AMD, cinebench is heavier but a more reliable test - if you're getting 4.55GHz all core, you're doing better than most of us.



HWinfo results dont tell us much if it's been running for under 2 minutes btw, you need to start it before the benchmarks for the idle results, then run the test programs so we have something to compare to and see how high temps get over time