Do Ryzen 7000 / 9000 chips still not idle efficiently?

I’m looking to build smaller Kubernetes / Proxmox nodes or storage servers and wanted to find a platform with ECC support and low idle power draw since I expect it to be spending most of the time at low load. Does anyone have experience with Ryzen 7000 / 9000 for such a use case? I’m looking at Ryzens mainly for cost reasons, suspect Siena would be much slower and more expensive and other Epycs would idle less efficiently.

I know the pro 4000G or 5000G CPUs idle very efficiently since the monolithic designs are from laptop chips. They’re pretty old though, so was looking for something more recent.

I’ve seen reports of the IO die making Ryzen 5000 chips very inefficient at idle, including how the package power reported from software doesn’t capture the actual power drawn by the CPU.

  1. Does package power draw on Ryzen 7000 or 9000 still not accurately reflect the actual power consumed by the CPU? From this review of the ASRock Rack 1U4LW-B650/2L2T with a Ryzen 7900:

    At idle, we saw only a 14W package power figure and so 20-35W depending on how the system is configured is reasonable at idle.

    If the actual system power draw was 20 - 35W, that’s pretty reasonable, implying that the package power measurements are accurate of CPU power draw. However, I see much higher idle values for the entire system in most other reviews for Ryzen 7000 / 9000 (50+ W) which contradicts this figure.

  2. That article recommends the Ryzen 7900 because of the 12 cores and 65W TDP, so that is one option I’m considering. Are there meaningful differences for idle power consumption if I went with a one-CCD chip (9700X) vs a 2-CCD chip (7900 or 9900X)? I know that the latency for inter-CCD communication is an issue for games, is it an issue for server type workloads?

  3. Is there any reason to get the Epyc 4004 series branded chips? They seem to be rebadged Ryzen 7000s.

If 7000 / 9000 still aren’t efficient at idle, what would you recommend for this use case?

Vermeer, Raphael/EPYC 4004, Granite Ridge, and Siena all use IFOPs and thus have similar idle characteristics. If AMD documents Siena’s internal power management details I haven’t found it but, since 8004’s positioned for low power, might also be one IFOP per CCD with ability to park unused IO.

Main AMD options would be V3000 or 8000G Pro with Krakan maybe releasing this quarter. Arrow Lake seems to be ~14 W idle, I suspect mostly on Foveros chiplet links being shorter range than SP5 capable ones.

50+ W wall draw’s routine in my experience. When comparing idle power data wall side versus load side, PSU light load efficiency, CCD count, DDR amount and clock, GPU and display configuration, and drive loadout all need to be controlled for. But usually they’re not.

Depends on workload and OS. Main issue’s 11’s incompetency with thread affinity can blow off 20-40% of hardware potential, hence Linux’s higher geomeans.

Did you mean the 8000G? AMD skipped 7000G and used the 8000 naming. ECC is only supported in the pro versions which seem to be unobtanium, and the 8000g APUs have terrible PCIe bifurcation

I don’t know much about the V3000, will take a look.

Edit: the V3000 are also unobtanium as well.