Question Zen 4 builders thread

[Markfw]

I know its early, but as someone pointed out, I better get my ram now before the scalpers are all over it.

To go with a 7950x, what do you think of this ram:

https://www.newegg.com/g-skill-32gb-288-pin-ddr5-sdram/p/N82E16820374369?Item=N82E16820374369

(edited for my best recommendation, 12/5/2023)

 

[H433x0n]

But do you mean the AMD chiplet design has a disadvantage w.r.t. idle power consumption even when a single chiplet/CCD is used? What would be the reason for that?

You have to power 2 individual dies (as well as the communication between them through the substrate). When you break down power consumption per component a single CCD itself will usually be sub 5W (at least it was on my 7950X), while the IODie can be up to 30W+ (depending on memory & motherboard). It's a dual edged sword because that same concept / architecture makes scaling Epyc trivial and is clearly the winning strategy for datacenter and for desktop it works fine because idle power consumption isn't really a big deal for that market.

When AMD needs a proper mobile SKU, they'll design a monolithic processor like the phoenix line (7940HS) to get that superior low power characteristics / battery life.

 

[Abwx]

But do you mean the AMD chiplet design has a disadvantage w.r.t. idle power consumption even when a single chiplet/CCD is used? What would be the reason for that?

10W difference in idle power at the CPU level is meaningless given the huge difference in perf/Watt when loading occur.

For one doing Handbrake the 7950x is 11% faster and use 60W less watts, wich increase the effective power difference per task to 66.6W.
If encoding during one hour it would require almost 7 hours subsequent idling power for the 13900K to compensate the difference,
if encoding is running 2 hours then that s 13 hours idling that are necessary to compensate for the lower efficency, and for 3 hours of encoding it s 20 hours subsequent idling to match the overall task + idling efficency...

But you can also set the 142W eco mode, at wich point the 7950X will be roughly as fast as the 13900K and use 110W less power, so in this configuration 11 hours idling are necessary to compensate for an hour of encoding, and if encoding is two hours then 22 hours idling etc etc..

At some point the debate is moot unless the CPU is idling all the time, but then what would be the use of a 7950X or a 13900K.?.

 

[Det0x]

Did a little 1:1 mode vs 2:1 mode memory comparison in clamchowder/MicrobenchmarksGui with maxed out memory timings on a other forum, but think i can share results here also
My goal was manly to take a closer look the performance numbers for single CCD cpus and show why 2:1 mode are wasted on these cpus since they are purely limited by FCLK. (everything above ~5800MT/s can fully saturate the maximum ~70k GB/s bandwidth @ 2200FCLK)
But ofcourse i also did comparison with all 16 cores enabled we can look at first:

Settings used for all tests:
7950x3d
Asus x670e GENE
32gigs teamgroup 7800 a-die (all memory settings have been tested stable to over 10k in Karhu)

12-16 core dual CCD cpu:
8000/2200 ~ 6600/2200


2:1 mode have 3.24GB/s higher read bandwidth but 1.1ns worse latency compared to 1:1 mode on my system.
Everything pretty much as expected, dual CCD cpus is the only place where 2:1 makes sense from a performance standpoint. Both setups can be pretty even performancewise depending on what you benchmark them in, but overall i would maybe say the 2:1 profile is faster, something which will be especially true when 82 and 84 memory multiplier get unlocked in the future.

Next up we disable the frequency CCD and simulate a 7800x3d which was my main interest for this comparison

6-8 core single CCD CPU:
6600/2200 vs 6400/2133 vs 8000/2200 vs 8000/2000


Big screenshot with lots of information to digest above, so have condest it some to make it easier to only look at the dram memory performance alone.

Think this paints a pretty clear picture, single CCD cpus are missing the GMI links to take advantage of the extra bandwidth afforded by 2:1 mode, which means they are solely capped by FCLK speed. Its only dual CCD cpus that can take advantage 2:1 mode since the IO-die is serving two CCDs instead if one.
The numbers are also showing that we are paying a latency tax for running the memory controller (UCLK) at half speed, something we cant make up for with todays max memory-limit at 8000MT/s (higher mem multipliers are bugged in ageisa, only way higher is by baseclock)
Performance ranking end up at 6600/2200 > 6400/2133 > 8000/2200 > 8000/2000

For single CCD cpus i would always say a proper 1:1 6400MT/s 2133mhz FCLK tune should be the goal

 

[Det0x]

Oh and this is how the numbers look in Aida with all cores enabled:


Full stability with ~25K% in Karhu and ~1.5hours in Ycruncher 8.1 all tests enabled

 

[Fjodor2001]

You have to power 2 individual dies (as well as the communication between them through the substrate). When you break down power consumption per component a single CCD itself will usually be sub 5W (at least it was on my 7950X), while the IODie can be up to 30W+ (depending on memory & motherboard). It's a dual edged sword because that same concept / architecture makes scaling Epyc trivial and is clearly the winning strategy for datacenter and for desktop it works fine because idle power consumption isn't really a big deal for that market.

When AMD needs a proper mobile SKU, they'll design a monolithic processor like the phoenix line (7940HS) to get that superior low power characteristics / battery life.

But the separate IO die will be needed for single CCD AMD Zen 4 CPUs too.

And if there was no separate IO die, its functionality would have to be included on some other die (the one that contains the actual CPU cores, like on the monolitic AMD G-series CPUs). So the IP blocks it contains would still be needed and consume power.

 

[H433x0n]

But the separate IO die will be needed for single CCD AMD Zen 4 CPUs.

And if there was no separate IO die, its functionality would have to be included on some other die (the one that contains the actual CPU cores, like on the monolitic AMD G-series CPUs). So the IP blocks it contains would still be needed and consume power.

It’s same on single ccd or dual ccd, just more pronounced on a dual ccd SKU since the IO die has to do more work for a dual CCD processor.

I really wouldn’t worry about it though. It’s negligible in the grand scheme of things.

 

[Fjodor2001]

It’s same on single ccd or dual ccd, just more pronounced on a dual ccd SKU since the IO die has to do more work for a dual CCD processor.

I really wouldn’t worry about it though. It’s negligible in the grand scheme of things.

Well, I somehow got the impression that you previously meant that the separate IO die would result in substantially higher idle power consumption compared to a monolitic CPU (where the functionality that the IO die provides will be on the monolitic die instead... most of it at least, but functionality for cross die communication will not be needed). But maybe I misinterpreted you?

 

[Fjodor2001]

10W difference in idle power at the CPU level is meaningless given the huge difference in perf/Watt when loading occur.

For one doing Handbrake the 7950x is 11% faster and use 60W less watts, wich increase the effective power difference per task to 66.6W.
If encoding during one hour it would require almost 7 hours subsequent idling power for the 13900K to compensate the difference,
if encoding is running 2 hours then that s 13 hours idling that are necessary to compensate for the lower efficency, and for 3 hours of encoding it s 20 hours subsequent idling to match the overall task + idling efficency...

But you can also set the 142W eco mode, at wich point the 7950X will be roughly as fast as the 13900K and use 110W less power, so in this configuration 11 hours idling are necessary to compensate for an hour of encoding, and if encoding is two hours then 22 hours idling etc etc..

At some point the debate is moot unless the CPU is idling all the time, but then what would be the use of a 7950X or a 13900K.?.

Well, I think there is a valid case where the PC normally is turned on for quite a large portion of the day, but is mostly idling or doing low performance "office type work".

Then occasionally (perhaps a few hours per week) the PC will be used for really high performance tasks, such as compiling large amounts of source code, transcoding movies, etc. And you want those tasks to be completed as quickly as possible so you do not have to wait for the results longer than necessary.

Then you'd like both really low idle power consumption and also very high performance (preferably also at as low power consumption as possible given the circumstances).

 

[exquisitechar]

Well, I somehow got the impression that you previously meant that the separate IO die would result in substantially higher idle power consumption compared to a monolitic CPU (where the functionality that the IO die provides will be on the monolitic die instead... most of it at least, but functionality for cross die communication will not be needed). But maybe I misinterpreted you?

No, you understood correctly. A monolithic or more sophisticated chiplet CPU would be better in that regard (and others), but AMD doesn’t really want to put much dedicated effort into their desktop CPUs since they’ll sell anyway and we get gimped server design hand-me-downs instead.

Well, I think there is a valid case where the PC normally is turned on for quite a large portion of the day, but is mostly idling or doing low performance "office type work".

Then occasionally (perhaps a few hours per week) the PC will be used for really high performance tasks, such as compiling large amounts of source code, transcoding movies, etc. And you want those tasks to be completed as quickly as possible so you do not have to wait for the results longer than necessary.

Then you'd like both really low idle power consumption and also very high performance (preferably also at as low power consumption as possible given the circumstances).

Yes. There’s no ideal in the current CPU market since AMD is generally less efficient in idle/low load scenarios and Intel is generally less efficient in high load scenarios.

 

[In2Photos]

Wow, 6 pages of discussion on idle power consumption. If you are that concerned about power usage maybe you don't need a high performance system.

 

[Fjodor2001]

Yes. There’s no ideal in the current CPU market since AMD is generally less efficient in idle/low load scenarios and Intel is generally less efficient in high load scenarios.

I guess the AMD G series could be an option, if no more than 8 cores is needed. IIRC, those have lower idle power consumption, and should still have fairly good characteristics in high performance loads too.

Problem there is no AMD G series CPU available based on Zen 4.... yet. But hopefully it will come soon.

 

[Fjodor2001]

Wow, 6 pages of discussion on idle power consumption. If you are that concerned about power usage maybe you don't need a high performance system.

Seems like you haven't actually followed the discussion. It's been covering lots of different aspects, e.g. max multi-core performance, power consumption and performance at different power limit levels, idle power consumption, suitable work types for different types of CPUs, technical reasons for the differences, etc. Also, various sources of information sometimes provide conflicting information, so there's a bit of back-and-forth discussion to sort out what is correct.

At least I for sure have found it very helpful and interesting.

 

[Abwx]

Wow, 6 pages of discussion on idle power consumption. If you are that concerned about power usage maybe you don't need a high performance system.

Or buy something like this, 5-7W idle power for the full system, no build can beat such a solution when it comes to idle power :

https://www.notebookcheck.net/AMD-Ryzen-7-7840HS-performance-debut-Beelink-GTR7-mini-PC-review.730507.0.html

 

[In2Photos]

Seems like you haven't actually followed the discussion. It's been covering lots of different aspects, e.g. max multi-core performance, power consumption and performance at different power limit levels, idle power consumption, suitable work types for different types of CPUs, technical reasons for the differences, etc. Also, various sources of information sometimes provide conflicting information, so there's a bit of back-and-forth discussion to sort out what is correct.

At least I for sure have found it very helpful and interesting.

No I followed the discussion. But let's look at the worst case scenario for the 7950X vs 13900K comparison. In that case the peak numbers for both CPUs have the 13900K consuming more than the 7950X, correct? Let's say that it consumes 100w more. If you ran the two systems at full load for 24/7 you would consume 72kWh more per month when using the 13900K. If you pay $0.12 per kWH this equates to $8.64 more per month. At $0.30 per kWh this is $21.60 per month. If you are considering paying $500 for a CPU are you really that concerned about $9-$22 per month more on your electric bill? And that's if you use if at full load 24/7! Since that is not your plan, in fact you mention that it may only be a few hours out of the month, the operating costs between the two systems are likely negligible. Which is the conclusion that has been reached regarding system power consumption under real world scenarios.

In regards to performance, does it again matter that much for something you do a few hours a month?

None of the discussion taking place is ground breaking. It has all been discussed before, but I'm glad you are finding it helpful and interesting.

 

[moinmoin]

But is it known what the actual technical root causes are for the lower idle power consumption on the Intel desktop CPUs? In what way are they designed differently (e.g. more quickly ramps down frequency when idle, has better power gating for CPU blocks, or something else)?

AMD cores are very efficient. What's the issue when you get high idle power consumption is all the uncore, I/O and peripherals. AMD's non-mobile chips are MCMs, so consist of several dies that need to communicate which also consumes some power. Intel up to now only uses monolithic dies on desktop that furthermore usually include less I/O support than comparable AMD chips. This can explain some of the difference between Intel systems idling at ~7W and AMD systems idling at ~22W.

Now the issue that there are boards for both using much more, like AMD systems idling at ~60W, appears to be down to overclocking which more and more manufacturers try to include as standard configuration of "premium" boards as they result in better performance numbers at the cost of power efficiency. Higher memory frequency will use more power both for the memory itself as well as for the integrated memory controller. Other areas are likely similar decisions for more performance in exchange for more power consumption/less efficiency.

 

[mv2devnull]

Asrock X670E Steel Legend, 7800X3D, 2*16GB G.Skill Flare X5 (6000 CL32).
In the BIOS set: Temp limit 75C, PPT 88W, TDC 75A, EDC 150A, and memory to EXPO profile.

What else is there to do to keep the CPU somewhat cool and quiet (under load)?

 

[Markfw]

Seems like you haven't actually followed the discussion. It's been covering lots of different aspects, e.g. max multi-core performance, power consumption and performance at different power limit levels, idle power consumption, suitable work types for different types of CPUs, technical reasons for the differences, etc. Also, various sources of information sometimes provide conflicting information, so there's a bit of back-and-forth discussion to sort out what is correct.

At least I for sure have found it very helpful and interesting.

I will reply, especially since its my thread. First, it seems like the 7950x or the 7950x3d is your choice. As stated @142 watts, its 95% of its performance, and both would have the same performance EXCEPT the rare case where the software can use the extra cache. Personally, I would not bet on that since you are not a gamer. As to the low power@idle, when they calculated that the 100 watts extra even at 24/7 full load, then the power usage at idle, be it 20 watts or 60 watts should make virtually no difference in your power bill. Here is one thing that WILL make a difference to you, the extra heat put out by the 13900k will be annoying, even of you limit the power a little. Remember, at the same wattage as the 7950x, it will be slower. Therefore there is no reason to get a 13900k.

I would recommend the ASRock motherboards as the best choice for many reason. My favorite is the TaiChi. Go for 6000 cl30 EXPO memory, it does not require much "OCing" and the power hit is minimal, but most applications see a big jump with it.

 

[A///]

I will reply, especially since its my thread. First, it seems like the 7950x or the 7950x3d is your choice. As stated @142 watts, its 95% of its performance, and both would have the same performance EXCEPT the rare case where the software can use the extra cache. Personally, I would not bet on that since you are not a gamer. As to the low power@idle, when they calculated that the 100 watts extra even at 24/7 full load, then the power usage at idle, be it 20 watts or 60 watts should make virtually no difference in your power bill. Here is one thing that WILL make a difference to you, the extra heat put out by the 13900k will be annoying, even of you limit the power a little. Remember, at the same wattage as the 7950x, it will be slower. Therefore there is no reason to get a 13900k.

I would recommend the ASRock motherboards as the best choice for many reason. My favorite is the TaiChi. Go for 6000 cl30 EXPO memory, it does not require much "OCing" and the power hit is minimal, but most applications see a big jump with it.

My only disagreement with you here is the cache on the x3d may be useful in some software outside gaming. It is however there if he wishes to game. shortly said if he were to upgrade to the zen 5 or zen 5 x3d processors next year the resale value of a x3d 7950 would have an edge over the non x3d variant, IMO. If no further down the line the x3d processor may net him more in resale.

Igor had some data on the x3d outside gaming which he offered to share with me when I said I was dumping intel for AMD with my plans which have been placed on hold due to a clusterfbombfbomb with everything going on right now. If he can get the 3d price for not a whole lot more than the non 3d it's a no brainer purchase imho. It does boost lower than the non 3d but it doesn't matter. The Intel 13900K will be that fat overweight kid you knew in summer camp who talked a mean game but was a huffing puffing piece of fat lard trying to hike up a hill while the 7950x/x3d was a spritely cross distance runner like you or anyone with a brain and a pulse to know intel is crap in this higher segment.

 

[igor_kavinski]

What else is there to do to keep the CPU somewhat cool and quiet (under load)?

V-cache CPUs run cool by default. I don't think there is any need to restrict the CPU. Get a bigger heatsink or use better thermal compound for better heat dissipation and lower fan speed.

 

[Fjodor2001]

I will reply, especially since its my thread. First, it seems like the 7950x or the 7950x3d is your choice. As stated @142 watts, its 95% of its performance, and both would have the same performance EXCEPT the rare case where the software can use the extra cache. Personally, I would not bet on that since you are not a gamer. As to the low power@idle, when they calculated that the 100 watts extra even at 24/7 full load, then the power usage at idle, be it 20 watts or 60 watts should make virtually no difference in your power bill. Here is one thing that WILL make a difference to you, the extra heat put out by the 13900k will be annoying, even of you limit the power a little. Remember, at the same wattage as the 7950x, it will be slower. Therefore there is no reason to get a 13900k.

I would recommend the ASRock motherboards as the best choice for many reason. My favorite is the TaiChi. Go for 6000 cl30 EXPO memory, it does not require much "OCing" and the power hit is minimal, but most applications see a big jump with it.

Been on away on vacation, so sorry for the late reply.

I think what you write makes sense. My takeaway so far from the discussion in this thread is:

* AMD 7950X (or optionally 7950X3D if I can get it at a not much higher price, mainly because it's has a bit better perf/watt in some use cases)
* B650 chipset
* Set PPT power limit to 142 Watts
* Keep VSoc low
* ASRock motherboard, or optionally Asus which had low idle power consumption
according to this test. Will have to check in more detail to find one with the features and form factor that I want too.

With regards to high power consumption at full load, I agree about trying keep that low to avoid heat output and fan noise.

With regards to low idle power consumption, it does add up over time. Example if idle power consumption is 30W extra compared to the most frugal alternative:

30W * 12h per day = 0,36 kWh/day => 365 * 0,36 = 131 kWh/year
Assume $0,3/kWh:
131 kWh/year * $0,3/kWh = $39/year
If you keep the PC for 5 years, that's $39/year*5 years = $195

Also, I've kinda made it a hobby to keep the power consumption low both at load and at idle, while still impacting performance as little as possible. We've all got our PC build related hobbies/hangups or we would not be hanging around this forum, but just buy Dell prebuilts instead.

 

[Hotrod2go]

Has anyone had success with 6600MHz RAM + FCLK=MCLK, stable in windows yet?

Edit: typo, mean't UCLK=MCLK.

 

[SteinFG]

Has anyone had success with 6600MHz RAM + FCLK=MCLK, stable in windows yet?

I think you mean UCLK=MCLK.
FCLK maxes out at around 2100.

 

[Det0x]

Has anyone had success with 6600MHz RAM + FCLK=MCLK, stable in windows yet?

Edit: typo, mean't UCLK=MCLK.

Yeah since ~1 year ago, how come ?

 

[Markfw]

After upgrading bios to the latest for my ASUS PRIME X670-P WIFI (1636) My machine MIGHT now be stable. I have had problems as of late with rebooting several times a day with various errors, that I believe were caused by memory errors. I was, and still am running the memory at 6000 cl 30. with this new bios I am now running 80c max, and CO of -25. Below is the wattages, and I think its now running 135 watt. Its at 4.5 ghz 100% full load. I am quite impressed with this bios.

 

[AdamK47]

After upgrading bios to the latest for my ASUS PRIME X670-P WIFI (1636) My machine MIGHT now be stable. I have had problems as of late with rebooting several times a day with various errors, that I believe were caused by memory errors. I was, and still am running the memory at 6000 cl 30. with this new bios I am now running 80c max, and CO of -25. Below is the wattages, and I think its now running 135 watt. Its at 4.5 ghz 100% full load. I am quite impressed with this bios.

View attachment 84583

I wouldn't rule out the negative CO as the cause.