One thing i repeat often since people are unaware: AGESA and EXPO do not set the SoC voltage. That is 100% upto the motherboards BIOS, AMD are simply putting in a limit of 1.30v - and the boards hardware can still overvolt beyond that setting (like asus do)
Asus had a bios offset of +0.05v and a hardware +0.05v too (possibly LLC controlled) giving them a solid +0.1v over what the user thought was being used at all times, now it's been reduced to half that - but it's still over volting.
With asus defaulting to 1.35 in the BIOS, that meant it was outputting 1.45v - no other brand did this
Running 32GB of RAM at 6000, seems to require around 1.25v SoC on average. Seeing boards default to 1.30-1.45v to achieve that is absurdly overkill.
A summary from anandtech:
- AMD Ryzen 7000X3D CPUs are shutting down too late to mitigate physical damage.
- ASRock, GIGABYTE, and MSI have a 116°C thermal trip point, and ASUS has 106°C, but sometimes didn't work as intended.
- The thermal cut-off for Ryzen 7000X3D is supposed to be 106°C and 116°C for Ryzen 7000.
- AMD EXPO enabled on ASUS is 1.35V on SoC voltage up until BIOS 1202 (AGESA 1.0.0.6).
- ASUS's SoC Voltage settings were/are too high.
- The AGESA firmware rollout has been nothing short of chaos at this point.
- AMD is offering RMA (paying shipping both ways) on killed CPUs, even if EXPO has been used (at least in the US)
- No word on if motherboard vendors will honor the warranty (at the time of writing)
106c and 116c is already too high for a cutoff, but then they fail to work - since they're relying on the CPU to do this and not the motherboard, when the CPU is crashed for any reason or hasnt finished posting, that meant no safeties (at least on the asus crosshair)
Crosshair specifically:
1.35v was really 1.42v, 1.4v at the CPU
a 1.35 V SoC setting within the ASUS firmware (and with EXPO enabled) resulted in an observed 1.398 V from an SoC pad. This was typically even higher when probed at the choke, at an eye-watering 1.42 V
That gives away roughly 0.02v of droop - which actually matches what i see on my 5800x3D where 1.15v becomes ~1.13v
On asrock Taichi
on the latest AGESA 1.0.0.7 (BETA) firmware, our G.Skill DDR5-6000 kit of DDR5 memory (2 x 16 GB) on the ASRock X670E Taichi would automatically preset 1.30 V on the SoC when applying the EXPO memory profile. To elaborate, unfortunately, we tried 1.15 V, which was a no-go, and even 1.20 V was a no-go. We eventually settled on 1.25 V on the SoC for this kit and our Ryzen 9 7950X3D
Defaulting to 1.30v (sigh) - but unstable at 1.20v with 2x16GB.
Most setups would be perfectly fine with 1.20-1.22v, with high speed/high capacity at 1.25v. These guys are going super aggressive on their overvolting for artificial compatibility.
For all we know, they could raise the IO die voltage and run less SoC - this is all a panicked approach focused on one value, and ignoring other alternatives.
Look at this weirdness, some of it is from AMD but the board makers are also changing their own settings (SoC voltage is NOT dictated by EXPO, or AMD)
The latest AGESA seems to boost temperatures a fair bit on the non 3D cache - it's possible it's better optimised and boosting higher, but it's a large change.
The other values didn't really change, but peak amperage did. Only about 5 watts more, but clearly 5 watts can make a big difference if its all dumped into one core.
As to why, the agesa changelogs seem to show that they disabled the EDC throttle for "PHX2" - the non 3D cores now have the full EDC limit of a regular chip, not the 3D chips limit
Hopefully this comes to AM4 to fix the EDC bug.
As to EXPO enabled, heres a great example: Undervolting the SoC from 1.30v to 1.25v, freed up more power for the CPU to run faster and hotter.
Overall temps are quite a bit higher, but performance should be too, on the non 3D cores.
Notice how the undervolted SoC didnt change the PPT wattage?
Yet somehow, clocks and temps are up. That implies it's not all-core that changed, but single core boost.
Specfic to the SoC, this is a great comparison of how things could go bad, on one of the better boards facing this issue
The asus boards going from 1.30 to 1.45v, would be blasting a LOT more
You'll notice that 6 has the worst peak, but 5c has the worst average -
but only with EXPO, which is entirely controlled by Asrock and not AMD
LOL
