Prompted by the news of a Ryzen 3900 (without -X) in early October, somebody tried different package power targets on a 3900X and reported Cinebench 20 results along with at-the-wall power consumption
in a German forum. The test system had a lower-end HD 7750 GPU, a 650 W titanium-rated PSU, and was of course supplied from 230 V. Firmware was AGESA 1.0.0.3 ABBA.
The user calculated CB-points-per-Watt for package power in his post. I am showing CB-points-per-Watt for at-the-wall power instead.
comment | TDP (W) | PPT (W) | TDC (A) | EDC (A) | CPU package power (W) | Cinebench 20 points | relative performance | system power (W) | relative system power | points/W of system power | relative efficiency |
stock | 105 | 142 | 95 | 140 | 127.5 | 7360 | 1.00 | 189 | 1.00 | 39 | 1.00 |
hard limit | 105 | 105 | 95 | 140 | 107.6 | 7119 | 0.97 | 161 | 0.85 | 44 | 1.14 |
eco mode | 65 | 88 | 60 | 90 | 89.1 | 6942 | 0.94 | 137 | 0.72 | 51 | 1.30 |
hard limit | 65 | 65 | 60 | 90 | 66.6 | 6426 | 0.87 | 107 | 0.57 | 60 | 1.54 |
sweetspot | | 58 | 60 | 90 | 59.4 | 6125 | 0.83 | 98 | 0.52 | 62 | 1.60 |
The performance degradation at lower power targets is rather moderate. Consequently, the lowest PPT tested is the optimum for performane-per-Watt not only when looking solely at CPU power consumption, but also when taking all of the platform power consumption overhead into account. I wonder how representative this image rendering workload may be for other computing scenarios.