AMD Ryzen 7000X3D Power Consumption Spiking Beyond 100 W in Idle Mode

[T0@st]

According to investigations undertaken by Igor's Lab and Hardware Busters this week it seems that AMD's problematic lineup of Ryzen 7000 & Ryzen 7000X3D CPUs are consuming unexpected levels of power in short bursts when running in idle mode. In conducting more in-depth tests over the past few days, Igor Wallossek and (outgoing TPU PSU expert) Aristeidis Bitziopoulos have both found that that the aforementioned AMD processors are producing (to the testers' slight concern) power spikes in situations involving minimal computing activity. It is not currently known whether the sharp climbs in power consumption are in any way related to the burnout issues experienced by unlucky overclockers this week.

Aris/crmaris (at Hardware Busters) says that he has tested many of the affected processors in the past, but was not privy to any major problems relating to burnout or power consumption spikes. By running new tests this week, using his own Powenetics v2 board, Aris has found out that: "There are some interesting facts here, which I didn't pay much attention to during the reviews because I only look at the average values and not the peak ones in idle. In the 7950X3D, there is a high spike during idle at 130 W, which is unjustified because the peak CPU load is only 3.53%. Even with the Curve Optimized enabled and a -15 setting, the idle power spike is close to 125 W, so something is happening there. On the 7800X3D, the spike during idle stays low, but this is not the case for the 7900X, which has an idle power spike at 109 W, while the peak CPU load at idle was at 5.12%, so these 109 W are not justified, either."



Video coverage of the matter has been provided by Hardware Busters in their "AMD Ryzen 7000 Series burn issues? Our Findings!" upload to YouTube:

Igor does not get dramatic about the latest findings, and suggests that AMD should not be judged too harshly: "Aris (crmaris) and I are of the opinion that these load peaks (which I could reproduce 1:1 in the meantime, by the way), primarily have nothing to do with the described CPU dying - we put a lot of emphasis on that to avoid a false panic. But on the other hand, such transients are nothing you need for a really stable system. Aris has also already spoken to some of his industry contacts, who have told him that the overall RMA rates of the 7000 series are actually lower than those of the 5000 series. This should avoid panic and also not discourage anyone from buying a new AMD CPU."

View at TechPowerUp Main Site | Source

 

[AusWolf]

How did they test this? I haven't seen any such spikes on my 7700X with HWinfo. Not that a momentary spike of 130 W (which is well within TDP limit) should burn your CPU anyway.

 

[R0H1T]

Did they check for any rogue processes/services with high CPU usage during the spikes? If they haven't tested that then their findings are at least incomplete/inconclusive if not invalid as well, Windows literally has thousands of threads/hundreds of services & even at default the vast majority of them will be enabled. Then there's "memory compression" ~

Spoiler

Just in the evening when I resumed windows from sleep I saw memory compression using lot of CPU for a substantial amount of time, this was fairly odd because it runs infrequently & only when the system is kinda running out of memory! Also observer effect anyone?

 

[evernessince]

How did they test this? I haven't seen any such spikes on my 7700X with HWinfo. Not that a momentary spike of 130 W (which is well within TDP limit) should burn your CPU anyway.

It's likely motherboard related. I had a 7700X and now have a 7800X3D and didn't see any idle spikes when keeping HWInfo open for hours. I'd get between 28w - 32w on idle.

The title of this article contradicts the content and is downright misleading. The CPUs aren't consuming a ton on idle, only that Arris found that his one sample happened to have a few spikes that were a bit high.

Did they check for any rogue processes/services with high CPU usage during the spikes? If they haven't tested that then their findings are at least incomplete/inconclusive if not invalid as well, Windows literally has thousands of threads/hundreds of services & even at default the vast majority of them will be enabled. Then there's "memory compression" ~

Spoiler

View attachment 293570

Just in the evening when I resumed windows from sleep I saw memory compression using lot of CPU for a substantial amount of time, this was fairly odd because it runs infrequently & only when the system is kinda running out of memory! Also observer effect anyone?

The source of this article was not intended to show high idle power consumption, it was intended as an investigative video as to the source of AMD CPU burnouts. The author is taking a small comment Arris made and trying to make it seem like it's something much bigger. I don't know how we go from a few small spikes on a single CPU to All 7000 series CPUs are always consuming a high amount at idle.

 

[Scircura]

Not just the peaks. The average power consumption readings are surprisingly high.

Igor/Aris:
7950X3D Idle AVG - 50.531 W
7950X3D Idle AVG, CO -15 - 45.952 W

W1zzard (from review):
7950X3D 1T - 37W
7950X3D 1T PBO+UV - 39W

Why are the idle readings higher than when loaded by an MP3 encoder?

It's even more out of whack for the 7900x. Aris reports idle is 63.8W, W1zzard reports 1T is 36-42W (depending on the review). Unfortunately W1zzard no longer records idle power for his reviews so I can't compare more directly.

OT: I would be pissed if I upgraded to a 7900, nominally 65W TDP, to later learn that it idled at 47W.

 

[mechtech]

So
Bios related
Or
Windows power settings related
Or
Windows scheduler
Or
Other
Or
Combinations
Or
Unknown??

 

[T0@st]

The title of this article contradicts the content and is downright misleading. The CPUs aren't consuming a ton on idle, only that Arris found that his one sample happened to have a few spikes that were a bit high.

Cheers for the feedback - adjusted the title and couple of details.

 

[Xuper]

Power Spike for 20ms? Just like GPU ?

 

[evernessince]

Not just the peaks. The average power consumption readings are surprisingly high.

Igor/Aris:
7950X3D Idle AVG - 50.531 W
7950X3D Idle AVG, CO -15 - 45.952 W

W1zzard (from review):
7950X3D 1T - 37W
7950X3D 1T PBO+UV - 39W

Why are the idle readings higher than when loaded by an MP3 encoder?

It's even more out of whack for the 7900x. Aris reports idle is 63.8W, W1zzard reports 1T is 36-42W (depending on the review). Unfortunately W1zzard no longer records idle power for his reviews so I can't compare more directly.

OT: I would be pissed if I upgraded to a 7900, nominally 65W TDP, to later learn that it idled at 47W.

Likely down to motherboard.

Here is a snapshot of my 7800X3D power breakdown:

The CPU cores are only taking 7.153 watts here, the rest is taken by overhead. Total power consumption is about 32 watts at the time I took the screenshot. Wizard's review shows 16w single threaded.

 

[TumbleGeorge]

Chiplets for logic... lnfinity fabric, I/O chiplet with tons of controllers. This is a penalty from the CPU layout architectural choice.

 

[Keivz]

I notice my cpu fans randomly (maybe every 2-3 minutes) rev up on idle on my 7700x (+gigabyte mobo) which I’ve never seen from a processor before. The cpu load doesn’t seem to change iirc. Gonna pay more attention to it now.

 

[ymdhis]

This is completely normal for the Ryzen line (at least starting from gen 2 aka 3000 line). They use extremely aggressive clock gating and auto voltage overclock, and they can shoot up to max wattage limits for microseconds when there is ANY load. Even my 3600 used to shoot up to 60-70W when there's only a few percentage of load. Also note that the load percentage means nothing without context, the 7900x at 5% load could mean that all CPU cores are at 5% load or that one core is at 100% load, a second core is at 25% load, and 22 other ones are idling. If all the cores would be at 5%, the cpu wouldn't shoot up too much, but if you have one core at 100%, the cpu will shoot that one core to the maximum turbo clocks possible.

This is literally how the entire product line works. It is how they were designed. It is normal.

What is not normal is that they tried beating Intel at the maximum watts used metric too, so the chips overclock themselves to 5+GHz so they can get +1 fps out of CS:GO. And this requires a stupid amount of power, burning down the chips and/or requiring 360 rads to keep them cool.

Just turn PBO off, problem solved, and the chip won't go higher than 70C even with just a hyper 212.

Also observer effect anyone?

And this too. AMD already warned about that a long time ago. The ryzen chips change clocks faster than applications can measure them, and polling the chip already gives it some work, increasing the clocks.

Just having Ryzen Master open made both my 3600 and 5600G idle at +10W higher, despite there being 0 actual workload. That could explain the difference in watts between Igor and W1zzard. But that difference could also be a ton of other things like board type, BIOS/AGESA version, whatever.

 

[TheDeeGee]

 

[tabascosauz]

How did they test this? I haven't seen any such spikes on my 7700X with HWinfo. Not that a momentary spike of 130 W (which is well within TDP limit) should burn your CPU anyway.

They're using hardware to take readings at a very high polling rate. Software will never pick up on spikes as fast as these, just like HWInfo is far too slow to pick up GPU transients.

That said, the average idle readings alone are still basically nothing surprising, and if they truly want to measure average idle free of disruptions, they should have tested in safemode.

VRM efficiency is also low at idle, so seeing higher readings on EPS than SVI3 software also seems par for the course; the gap should narrow under load.

Did they check for any rogue processes/services with high CPU usage during the spikes? If they haven't tested that then their findings are at least incomplete/inconclusive if not invalid as well, Windows literally has thousands of threads/hundreds of services & even at default the vast majority of them will be enabled. Then there's "memory compression" ~
Just in the evening when I resumed windows from sleep I saw memory compression using lot of CPU for a substantial amount of time, this was fairly odd because it runs infrequently & only when the system is kinda running out of memory! Also observer effect anyone?

What aris and igor observed is not your typical Ryzen idle spikes, Aris uses a physical power measurement board to grab EPS readings from the cable.

130W peaks aren't idle or Windows related process spikes anymore, lol. Windows services and processes that are significant enough to disrupt idle also apply loads that pretty much always are sustained for a few seconds and long enough for even software to record the load and wattage increase.

 

[Max(IT)]

I could people still trust AMD nowadays ? Gone are the Zen2 and Zen3 days... AMD is on a downward slope.

Likely down to motherboard.

Here is a snapshot of my 7800X3D power breakdown:

View attachment 293576

The CPU cores are only taking 7.153 watts here, the rest is taken by overhead. Total power consumption is about 32 watts at the time I took the screenshot. Wizard's review shows 16w single threaded.

no way HWinfo64 polling can see those spikes.

 

[R0H1T]

What aris and igor observed is not your typical Ryzen idle spikes, Aris uses a physical power measurement board to grab EPS readings from the cable.

130W peaks aren't idle or Windows related process spikes anymore, lol. Windows services and processes that are significant enough to disrupt idle also apply loads that pretty much always are sustained for a few seconds and long enough for even software to record the load and wattage increase.

It can be another process not just some random windows service or something, also 3.53% CPU load is quite substantial on 16c/32t halo x86 chip.

This is the CPU load with 2 separate browsers & 20+ tabs combined.

This when I started hwinfo.

This when I fired up the sensors on hwinfo with 200ms polling.
The higher CPU usage at the end is when I changed the polling interval to 20ms about a minute or two after launching hwinfo.
So don't assume you know something when we don't have the full data on this. It could well be a defective chip, windows or any other "rogue" process!

 

[pavle]

Looks like the affected CPU's were rushed to the market for whatever reason and now these are the results, or Microsoft just prefers Intel?
Does the same thing happen on Linux (near-idle power spikes)?

 

[tabascosauz]

It can be another process not just some random windows service or something, also 3.53% CPU load is quite substantial on 16c/32t halo x86 chip.

This is the CPU load with 2 separate browsers & 20+ tabs combined.

So don't assume you know something when we don't have the full data on this. It could well be a defective chip, windows or any other "rogue" process!

So? CPU load doesn't necessarily correlate to a certain wattage power draw at all times. I didn't come to any conclusion, just that bringing up the "observer effect" makes no sense in this instance, and that a fast spike to 130W from idle is a lot, even after taking into account that 2CCD Ryzen regularly fluctuates up to the 50-100W area at idle. Sounds like you're the only one jumping to conclusions about it being a Windows process.

The "spikes" you see in your monitoring are sustained far longer.

This when I fired up the sensors on hwinfo with 200ms polling.

You do realize that 200ms is laughably slow compared to how they're testing, right? lol

Kinda the whole point of monitoring with hardware from EPS, like I said, picks up a lot more data than HWInfo ever could.

 

[R0H1T]

The observer effect was in regards to hwinfo, which I assumed was also running on their system, But it "applies everywhere" surely you know that?

Observer effect (physics) - Wikipedia

en.wikipedia.org

You still haven't explained how without knowing what's running in the background we can come to any conclusion? Like a I said 3.53% load on a 16c/32t CPU isn't negligible, unless they were talking about load just on the one core?

 

[mechtech]

How did they test this? I haven't seen any such spikes on my 7700X with HWinfo. Not that a momentary spike of 130 W (which is well within TDP limit) should burn your CPU anyway.

Hmmmm I wonder if it was done with integrated graphics and no dedicated gpu??

 

[freeagent]

PBO or not, it doesn't seem to matter. X3D or not, doesn't seem to matter.. a problem with Zen 4?

Maybe..

 

[Why_Me]

I could people still trust AMD nowadays ? Gone are the Zen2 and Zen3 days... AMD is on a downward slope.


no way HWinfo64 polling can see those spikes.

I live for these threads ^^ Such a refreshing change from iNTeL and Ngreeeeeeeeediaaaaaaa are capitalist peegs who steal souls.

 

[Darmok N Jalad]

This is completely normal for the Ryzen line (at least starting from gen 2 aka 3000 line). They use extremely aggressive clock gating and auto voltage overclock, and they can shoot up to max wattage limits for microseconds when there is ANY load. Even my 3600 used to shoot up to 60-70W when there's only a few percentage of load. Also note that the load percentage means nothing without context, the 7900x at 5% load could mean that all CPU cores are at 5% load or that one core is at 100% load, a second core is at 25% load, and 22 other ones are idling. If all the cores would be at 5%, the cpu wouldn't shoot up too much, but if you have one core at 100%, the cpu will shoot that one core to the maximum turbo clocks possible.

This is literally how the entire product line works. It is how they were designed. It is normal.

What is not normal is that they tried beating Intel at the maximum watts used metric too, so the chips overclock themselves to 5+GHz so they can get +1 fps out of CS:GO. And this requires a stupid amount of power, burning down the chips and/or requiring 360 rads to keep them cool.

Just turn PBO off, problem solved, and the chip won't go higher than 70C even with just a hyper 212.


And this too. AMD already warned about that a long time ago. The ryzen chips change clocks faster than applications can measure them, and polling the chip already gives it some work, increasing the clocks.

Just having Ryzen Master open made both my 3600 and 5600G idle at +10W higher, despite there being 0 actual workload. That could explain the difference in watts between Igor and W1zzard. But that difference could also be a ton of other things like board type, BIOS/AGESA version, whatever.

Yeah, there has been an approach in cpu design for some time now where the conventional wisdom is to ramp up sharply to retire a process quickly, as opposed to staying low and letting it drag out. You get some responsiveness benefits as well, since the CPU wastes little time getting your demand met. If the spike is very brief, there won’t be much a temp change. I guess we’ll see what comes of this next.

 

[evernessince]

I could people still trust AMD nowadays ? Gone are the Zen2 and Zen3 days... AMD is on a downward slope.


no way HWinfo64 polling can see those spikes.

Arris also reported a higher than average idle power consumption along with those spikes, something that HWInfo undoubtedly would be able to record.

130W peaks aren't idle or Windows related process spikes anymore, lol. Windows services and processes that are significant enough to disrupt idle also apply loads that pretty much always are sustained for a few seconds and long enough for even software to record the load and wattage increase.

The video states that his average load was 3.53% on his 7950X3D during the 15-20 minute idle testing, which is approximately equivalent to having steam, discord, Microsoft Anti-maleware, and one other process actively taking resources. I get about 3% load on my 7800X3D with those first three running. So not complete idle (which is not really possible without heavily tinkering with the OS) and IMO not long enough or a large enough sample size. Then again the video was never intended to say it's a problem, only point out a curiosity. A potential followup would be nice.

Yeah, there has been an approach in cpu design for some time now where the conventional wisdom is to ramp up sharply to retire a process quickly, as opposed to staying low and letting it drag out. You get some responsiveness benefits as well, since the CPU wastes little time getting your demand met. If the spike is very brief, there won’t be much a temp change. I guess we’ll see what comes of this next.

That's how they should work but then again if that were the case power consumption on average should still match that of reviews. That he's seeing a higher average power consumption hints that something else is going on or perhaps it's simply not a large enough sample size. After all the point of his video wasn't the power spikes and he only measured for 15 - 20 minutes on a single sample.

 

[R-T-B]

How did they test this? I haven't seen any such spikes on my 7700X with HWinfo. Not that a momentary spike of 130 W (which is well within TDP limit) should burn your CPU anyway.

Likely with socket level voltage probes. The reviewer in question was TPUs PSU reviewer, and does not mess around.