Hello all, as you see i am new to the forums and new to level 1 techs youtube channel. Ive been doing a lot of research about 7950x and 9950x. I also watched vids from level 1 about both. I am still a little confused about the core parking fixes so looking for direction on what one would do with so fixes. First i want either 7950x or 9950x for a media creation/ maybe streaming pc. Would one buy a 7950x and not worry about core parking or should one get the 9950x and hope for fixes to come besides just using a game bar or others as such. I know this had been talked about before but if one were considering them for the type of build i need what would one choose? Or recommend. Thanks all. If u could just point out like well you had to deal with those issues and can apply a fix its okay to get still type of thing for 9950x. Or naw go 7950x donât worry about core parking for now with that process and do 7950x. I just am planning this build and want to make sure i get best direction. I know that fixes could and should come but just need a little more lets say outlook between the 2. Thanks all if ya understand.
Welcome to Level1Techs!
There are lots of reasons to buy new gear. From Intel, from AMD, or any other manufacturer.
â[M]edia, virtual machines learning stuffâ doesnât require remotely that much horse power (itâs nice to have, tho).
All Iâm saying is that youâre in a rabbit hole in your decision process and should come out to breathe some fresh air
I am not unhappy with running virtualization and learning stuff on a core i3-6100 (dual core, quad threads), running media on a i5-6500 (quad core) from eons ago - all of these sipping power. I have much newer and nicer hw, for same and other purposes, just making a point.
I realized that hw over the last few years has power consumption and performance stongly correlated in a linear relationship and have adjusted my always-on hw to run on the lowest power consumption platfrom that can do the job. The other hw gets turned on when I need it.
I assume that the âmediaâ use case is a 24x7 use case, the âlearning stuffâ is a whenever needed use case, the âvirtualizationâ possibly a mix of both.
As you now realize - I donât have much to add in regards to your actual question. I just write to point out that
- AMD âŚ950x CPUs are well suited for mid-size server workloads (AMD even sells these rebranded in the EPYC 4004 line)
- as a result, platforms based on these typically consume correlating amounts of power. In 2025 you should probably take power consumption into account when designing a new system
- You could probably benefit from clarifying your use case for the new hw - at least in regards to separating your needs in always-on vs. occasionally-on categories.
I assume youâll get lots of answers from the fine folks here on your question.
Good luck and have fun!
Havenât seen anything about core parking and game bar since Zen 5 launch. Havenât ever parked cores on the 9900X I use. Windows 11âs scheduler remains poor compared to 10 and Linux but itâs arguably adequate most of the time with the September 2024 combination of 11 patches and AGESA 1.2.0.2.
In general, 11 has varying degrees of difficulty with anything thatâs not single thread or all core. So 7950X, 9950X, 285K, 7900X, 9900X, 265K, etc. etc. etc. doesnât matter. What does make a difference is not using 11 or configuring workloads to avoid 11âs most incompetent points.
Core parkingâs maybe a partial workaround for non-configurable workloads on an OS known for dropping 20-30% perf in certain cases. Personally Iâve had 11 drop up to 70% and when it gets that bad parkingâs unlikely to help.
Thank you both on info above, yes my question is a bit opened sorry about. Was trying to ask all at once. Leema, i am still on windows 10 and plan running that for now till end of support however was trying to future proof or hoping a bit so that is what i was hoping for answers to the future outside of the videos i had reviewed. So if i go 7950x and windows 10 it would work great until 11 later and by then i would have more know how from learning and get my work flow down?
Yes yes power consumption important and 7950x will draw more. Windows 10 does not have the core parking issues with 9900x 9950x is that correct?
And are you using 10 then with your 9900x build or saw no core parking on 11 with your build? Thx.
You only need core parking if you are gaming. If the main purpose of the PC is going to be media content creation or streaming, you are going to use all of your cores all the time, so no need for core parking.
If you update to the latest BIOSâ, then you donât use Windows broken core parking and you simply enable much better core parking in the BIOS. Again, you only turn off the second CCD when you are gaming, BIOS controls give you much finer control over how you use the cpu for various workloads.
You shouldnât have power concerns since your use case requires as many cores as you can get. So you should accept that the cpu will use what the power budget allows for your cpu. AMD two CCD cpus will use up to 280W if you use all of the 32 threads for highest level SIMD instructions like AVX-512. But will sip power if just idling on the desktop doing typical low level browsing and such.
If AMD didnât hamstring the multi-CCD cpus with one good die and one mediocre die, then you wouldnât need core parking in any use case at all. But that is probably never going to happen as most of the âgoodâ dies are being utilized in the server cpus and the desktop cpus get short changed. Such is life.
Forced to 11 23H2 by IT, though at least 24H2âs still blocked. I think you might be confusing, or at least partially conflating, core parking and sleep. Sleep happens whenever a coreâs unused.
How much youâll find out about how workloads run depends on how much monitoring you do. If threading aligns with one or both CCDs 11 mostly doesnât fvck it up too badly in my experience. So far rebootingâs worked when itâs gotten really delusional. While whatever messed up state 11 keeps persists across process restarts, bonking the kernel seems to clear it out. Maybe thereâs some bugged cache which can be flushed but, if so, I havenât found it.
To address some of the misinfo thatâs been posted, dual CCD default PPTs are 88, 145, 162, or 230 W depending on the part and a lot of them wonât pull 280 with unrestricted PBO. Dual CCD processors are binned with a higher boosting compute die (whose cores usually all come first in the performance order) and a compute die binned more for efficiency. Itâs an elegant way to allocate socket power across different active core counts.
11âs problem is basically itâs shit at keeping threads affinitized to cores and is particularly prone to struggling with whether to hyperthread on Granite Ridge CCD0 versus putting threads to CCD1. Cache refill penalties are particularly high when 11 thrashes threads across CCDs, though at least Microsoftâs mitigated their bug where branch prediction state wasnât getting transferred.
Thanks all very much with my questions and helping with a better understanding of all this! You all have been a great help and i have much better outlook! Thanks for the info about the default PPTâs as well. That is noted!
I find that consumer grade cpus like 7950 or 9950 is waste of time due to limited number of PCIe lanes. You throw GPU, 25gbe NIC, few NVMes and you are out of those.
If the idea just to learn, I guess it doesnât matter then take any. But otherwise I would buy something like Epyc 7402P
Beg to differ, this is a screenshot of Zen3 monitor as root to be able to read the MSR registers for power usage. PC is only running per CCX, not any PBO or anything else not default other than memory EXPO and the perCCX at locked 52.25/51.00 multiplier at 1.27V.
Package power at the top of the list shows 280W. Running my normal BOINC workload at 92% core utilization and also driving a 3090 and 3080 Ti besides my normal cpu tasks.
No changes to default power settings.
PC normally pulls 1075 watts from the UPS with the 3090 and 3080 Ti power limited to 265W because allowing any more exceeds the APC SMT-1500C UPS 1000 watt overload limiter.
Given the multiple obvious reporting bugs there Iâd suggest using a reliable tool or measuring at EPS. Even if you have actually juiced your 9950X to 286 W sustained the prevailing body of dual CCD data remains mostly ~250 W and under.
Couple other cautions Iâd make are 1) in the 9950X benches Iâm aware of itâs not uncommon overclocks run slower than stock and 2) 30-35+ W at idleâs hardly sipping power compared to Arrow (~14 W) or APUs (~5 W).
The only âobviousâ reporting bugs for zenmonitor3 is that was never supposed to handle any newer architecture other than Zen 3. Development stopped there.
However it runs correctly and reports correctly for Zen 4 as the defines in the code have the ZEN 4 family IDâs handled. Power reporting is derived from the system MSR registers directly.
Zenmonitor 3 is only reading the system registers.
When I attempted to use it on my new ZEN 5 cpus, it simply reported âNo ZEN cpu foundâ
But I discovered that if I simply replaced the Zen 3 Family 19H ID with the new ZEN 5 Family 1Ah family ID for the Zen 3 define, the program was able to read the MSR registers for power and the clock frequencies correctly for my 9950X cpus.
But the program tries to read the MSR registers with the Family 19H offsets which have changed I guess over what the offsets were defined for Zen 3 and Zen 4. So the frequencies in root mode are wonky. But in normal userland mode, the frequencies are correct.
I donât have the information about how to probe what the Zen 5 MSR register offsets are. It was necessary for the program users to experimentally probe around the registers in Linux based on some hints derived for what the Windows LibreHardwareMonitor app users found.
I did figure out what registers to read experimentally for my Asus ProArt X670E Creator Wifi board used for the EC sensors. I provided that information to the asus-ec-sensor module developer so that this board was added to the compatible hardware list. Proud of that.
That is how the Zenmonitor3 developer figured things out initially. I donât have any Windows system to run the LibreHardwareMonitor app on to try and figure out what registers that program pokes.
I believe what the system shows for cpu power. It tracks with what the UPS front panel shows when the system is booted without starting up BOINC and the gpus are idle.
Subtracting about 10-15W with the monitor off and the network gateway and switches are pulling which are also plugged into the UPS, the display shows about 102W being pulled as a idle system which falls in line with what the 9950X reads in idle in Windows monitoring apps from reviews.
That covers the motherboard power for the SoC, RAM, disk drives and the fans in the PC.
I am not âjuicingâ my cpus in any way, Iâm in fact downclocking/downvolting them on purpose to what they will stably run on full BOINC loads. I donât need gaming full boost clocks but the highest clocks that will remain within the stock cpu/motherboard PPT limits.
If you try and run Auto on consumer motherboards, your clocks will fall way down under stress loads because Auto tries to keep a few cores boosted all the way to max. No point in that with full core loading like BOINC which uses mostly AVX2 or AVX-512 apps.
So purposely locking the two CCDâs to the highest clocks the power limits allow via reduced perCCX multipliers and reduced VIDâs is the smart decision. I get better clocks overall on all cores that way compared to what Auto settings accomplish.
Crunching BOINC loads is not about power saving, itâs about producing the most science results possible on a host.
You are effectively overclocking though⌠AVX2/512 will pull very high currents which combined with your static clocks and Vcore put you OVER the the stock PPT of 230W. If it runs stable, thatâs great, but for this particular workload you are over spec.
Running stock/auto the CPU will run as fast as it can within temperature, PPT, TDC, EDC, voltage and frequency limits. In your case you fixed voltage and frequency and let the rest run to mobo limits.