Trying to evaluate the power consumption to see what frequency the CPU actually runs at in rendering. Here's where I noticed something interesting. The 4.0GHz 9900k actually beats 3700x(4.05G) by more than 20w in rendering. The voltage is auto(default p-state frequency-voltage pair).
Here's the list I collected. Already seen a lot of interesting results, like how taxing high-frequency is.Currently still looking for more data, to compare both designs at a fair level.
Here comes the questions:
1. Is 7nm+Zen2 really better than 14nm+SKL at power consumption, at the same frequency&corecnt?
2. Is it really possible for something like 3990x to run at 3.5GHz all core while doing something as intensive as rendering, as some reviews suggest?
Test method: Capture both the per-core frequency and the package power during the second half of Cinebench run, when they're stablized.
Reason for this method:
1. both values are stable at that time.
2. almost perfect cpu-bound(dram bound 0.1%-0.8%).
3. other rendering scenarioes(vray, keyshot, corona) yield almost the same frequency under the same TDP limit.
4. The "package power" is known to be reasonably accurate compared to direct measurement on the +12v input lanes subtract the losses on power rails. It's also the deciding factor of what frequency CPU runs at. For example, 8280L will be throttled to a steady-state frequency where "package power" limit was capped at 205W as shown above.
And yes I/O power is considered. They're quite static in the same design. Also the efficiency of F-IVR should be considered as well(this is the main reason why Xeon fall behind in high-load non-simd situations).
Here's the list I collected. Already seen a lot of interesting results, like how taxing high-frequency is.Currently still looking for more data, to compare both designs at a fair level.
Here comes the questions:
1. Is 7nm+Zen2 really better than 14nm+SKL at power consumption, at the same frequency&corecnt?
2. Is it really possible for something like 3990x to run at 3.5GHz all core while doing something as intensive as rendering, as some reviews suggest?
Test method: Capture both the per-core frequency and the package power during the second half of Cinebench run, when they're stablized.
Reason for this method:
1. both values are stable at that time.
2. almost perfect cpu-bound(dram bound 0.1%-0.8%).
3. other rendering scenarioes(vray, keyshot, corona) yield almost the same frequency under the same TDP limit.
4. The "package power" is known to be reasonably accurate compared to direct measurement on the +12v input lanes subtract the losses on power rails. It's also the deciding factor of what frequency CPU runs at. For example, 8280L will be throttled to a steady-state frequency where "package power" limit was capped at 205W as shown above.
And yes I/O power is considered. They're quite static in the same design. Also the efficiency of F-IVR should be considered as well(this is the main reason why Xeon fall behind in high-load non-simd situations).
Attachments
-
385.2 KB Views: 12
-
307.5 KB Views: 11
Last edited: