Currently seeing ~100W full system idle power with a ryzen 7950x in eco mode, and 2 gpus. GPUs are reporting ~10W power each at idle.
lm_sensors is reporting ~50W package power under the CPU’s iGPU which seems high.
Is this expected? Anyone else running Linux on these able to share their power stats?
analyzing CPU 0:
driver: amd-pstate
CPUs which run at the same hardware frequency: 0
CPUs which need to have their frequency coordinated by software: 0
maximum transition latency: 131 us
hardware limits: 400 MHz - 5.88 GHz
available cpufreq governors: performance schedutil
current policy: frequency should be within 400 MHz and 5.88 GHz.
The governor "schedutil" may decide which speed to use
within this range.
current CPU frequency: Unable to call hardware
current CPU frequency: 2.74 GHz (asserted by call to kernel)
boost state support:
Supported: yes
Active: yes
AMD PSTATE Highest Performance: 166. Maximum Frequency: 5.88 GHz.
AMD PSTATE Nominal Performance: 127. Nominal Frequency: 4.50 GHz.
AMD PSTATE Lowest Non-linear Performance: 85. Lowest Non-linear Frequency: 3.01 GHz.
AMD PSTATE Lowest Performance: 12. Lowest Frequency: 400 MHz.
# sensors output for CPU's iGPU
amdgpu-pci-6e00
Adapter: PCI adapter
vddgfx: 1.29 V
vddnb: 1.00 V
edge: +46.0°C
PPT: 58.08 W
$ uname -a
Linux lun-hisame-nixos 6.0.2 #1-NixOS SMP PREEMPT_DYNAMIC Sat Oct 15 06:02:59 UTC 2022 x86_64 GNU/Linux
I do not think that there is anything wrong with the sensoring, high temperatures and high power consumption is the sad reality with the new Ryzen 7K platform - even @ idle. There are quite a lot of rumors that this is caused by a very high PPT and VCore, and that this may be fixed by lowering the core voltage without have a significant performance hit.
Looks like this video has it at 42w cpu idle. then 2x 10w for your graphics cards, 5w per stick of ram and then the rest of the motherboard + psu inefficiency; 100w idle from the wall doesn’t really sound totally broken just not great
I appreciate high power consumption on load. Especially, if I can tune this with BIOS settings (PBO/CBO, ECO mode). But in reality most computer systems run idle most of the time. I have a hard requirement for that to be low - lower than last gen at least (AM4 isn’t that great in comparison to Intel since gen 6).
I hope that over the next few months we’ll see idle power consumption on AM5 come down.
The only real benefit from AM5 is AFAIK PCIe 5 which may yield some respectable increments in data transfer rates for NVMe drives, up to three times faster than PCIe 4 (15GB/sec vs. 5GB/sec). For professionals working with video and photo processing, this may be appealing.
With the current platform prices and the insane power consumption combined with high electricity cost (currently €1/kWh in Denmark for some customers), the AM5 platform seems less appealing for the average user.
… and even this is not a real benefit - yet. As far as I know there are no PCIe Gen5 expansion cards, and PCIe Gen 5 m.2 (or other format) SSDs still have to arrive.
The big curve ball could be RDNA3 / NAVI31 since that has support for PCIe 5 16x. But since NVIDIA’s 4000 series is not able to saturate PCIe 4, RDNA3’s support for PCIe 5 may be a fuzzy nothing burger.
SSD’s for PCIe 5 will eventually arrive, I’m quite sure of that. Samsung has flashed a 990 Pro with transfer rates of up to 13GB/sec read and 6,6GB/sec write.
@Danese, I can see your point about PCIe 5, but no one is talking about the limits of physical PCIe slots on the new motherboards. What I mean is that this time around, the motherboard manufacturers are trying to force more people to Threadripper Pro boards. On the new AM5 motherboards, I can’t see how you can have two graphics cards, an enterprise-level ethernet card, some other equipment you want to add for pass-through, and NVMe on the same physical system. My point is motherboard manufacturers are forcing people who want to mess with virtual machine pass-through to the Threadripper Pro system because that is the only current system that has enough physical PCIe slots.
Running Fedora 36, Radeon VII, and a 7950x with the “Balanced” Kernel setting:
analyzing CPU 0:
driver: amd-pstate
CPUs which run at the same hardware frequency: 0
CPUs which need to have their frequency coordinated by software: 0
maximum transition latency: 131 us
hardware limits: 400 MHz - 5.88 GHz
available cpufreq governors: conservative ondemand userspace powersave performance schedutil
current policy: frequency should be within 400 MHz and 5.88 GHz.
The governor "schedutil" may decide which speed to use
within this range.
current CPU frequency: Unable to call hardware
current CPU frequency: 400 MHz (asserted by call to kernel)
boost state support:
Supported: yes
Active: yes
AMD PSTATE Highest Performance: 166. Maximum Frequency: 5.88 GHz.
AMD PSTATE Nominal Performance: 127. Nominal Frequency: 4.50 GHz.
AMD PSTATE Lowest Non-linear Performance: 85. Lowest Non-linear Frequency: 3.01 GHz.
AMD PSTATE Lowest Performance: 12. Lowest Frequency: 400 MHz.
and
amdgpu-pci-1a00
Adapter: PCI adapter
vddgfx: 729.00 mV
vddnb: 1.24 V
edge: +41.0°C
PPT: 31.14 W
I followed the advice Wendell gave in one of the videos and use a negative PBO offset of -20. The video card is for compute only (and these numbers reflect no compute workload), using onboard gpu.
Under Linux:
From my experience, using boot parameter amd_pstate=passive, with “schedutil” scaling governor results lower temperature. My 7950x idles at 40C (air cooled). Room temperature is about 22C.
Also, there is a rumor that AGESA 1.0.0.6 lowers power consumption. I am still on 1.0.0.5, so I cannot confirm that.
afaik, high idle power is a result of the CCD+IO chiplet design, and there is no fix. It’s just one of the limitations of chiplet based architectures, and is why mobile designs are still monolithic.
I also found these couple articles that go into much more detail on this (for AM4, but likely a lot of it still applies to AM5):
TLDR; is main sources seem to be 1) CCDs and 2) Memory controller, but other parts add up too, and AMD doesn’t seem to be focused that much on low-end power consumption for the desktop platform… It also severely under-reports the actual power consumption in software.