1)
The TDP equals the power consumed by the CPU. It is a consequence of the first law of thermo that all the power dissipated as heat is power that is being consumed by the CPU. Energy is neither created nor destroyed. So all the energy leaving the CPU as heat had to enter through the socket first. Measuring power consumed at the socket level is how we measure TDPs. This is standard and well-known methodology.
2)
TDP is related to sustained power. This is also well-known. If the CPU is dissipating 150W when doing a given task as transcoding a video during thirteen minutes, then the cooler has to be able to dissipate 150W during those thirteen minutes, otherwise the CPU overheats and would have to throttle to reduce the heat generated. Sustained power is not peak power.
3)
HFR is not measuring peak powers. They are measuring average power. They are measuring sustained power. The average power consumed by the CPUs during the full x264 workload. HFR is measuring the sustained power during the workload.
4)
Peak power can be higher than TDP. The 1.5 factor mentioned by Hennessy & Patterson is an empirical value based in older hardware. It could be 1.7 or 1.3 with modern hardware. But it is irrelevant which is the peak, because no one is measuring peak powers to find TDPs, what we measure are average power consumed during the workload.
In fact all of you that are mentioning this book as proof that "Thermal Design Point is not power consumption" just didn't read the book, because Hennessy & Patterson just state the contrary than you pretend. The authors claim that the TDP is the [sustained] power consumption
TDP = sustained power consumption
Second, what is the sustained power consumption? This metric is widely called the thermal design power (TDP), since it determines the cooling requirement.
Hennessy & Patterson know the first law of the thermo.
5)
109.4W and 153.6W are not TDPs of i7-7700K and i7-5960X.
LOL No. Those are power consumptions measured at the 12V rail. To obtain the TDP of the CPU one has to subtract the measurement/VRM loses to get the power consumed by the CPU alone. Using the same 85% factor that HFR uses for AMD hardware we find the values of 92.99W and 130.56W for those Intel chips. Both values are within the 95W and 140W official TDPs. So both chips the i7-7700K and the i7-5960X are within the rated TDP. On the other hand several RyZen models violate the official TDP. The measured 128.9W for the 1800X gives, using the same 85% factor, a total of 109.57W at the CPU level. This measured value agrees with the value reported by the internal monitor of the RyZen processor which indicates 112W. Evidently 110--112W are a violation of the official 95W TDP. The 1800X is violating the official TDP. The i7-7700K and i7-5960X aren't.
6)
AMD admitted the real TDPs for the violating chips and communicated those values to HFR. 128W for the R7 1800X/1700X and 90W for the R7 1700. CanardPC confirmed with measurements that the real TDP of the 1700 is 90W.
The 1700 has 90W TDP in reality: AMD bullshit son TDP.
7)
This is not AMD versus Intel.
The RyZen R5/R3 models satisfy the official TDPs. Bulldozer, Piledriver and Phenom models also satisfy the official TDPs. Both ThreadRipper models also satisfy the official TDP. Examples: FX-8350 does 125W. Phenom II X6 110T does 121W. TR 1950X does 171W. RyZen 5 1600X does 87W... Those are power tired at the 12V rail. It is evident that the CPUs are working within the official TDP. 125W TDP for Piledriver and Phenom. 180W TDP for ThreadRipper. 95W TDP for the R5 model...
All those CPUs from AMD satisfy the official TDP. The only CPUs that violate the TDP are the 1800X/1700X and 1700. So, it is not AMD vs Intel. It is not about AMD and Intel measuring things differently. It is AMD vs AMD. It is about many AMD CPUs satisfying the official TDP, whereas some few AMD CPUs violate the official TDP.
8)
We know
why those R7 RyZen models violate the official TDP. We know that engineers planned higher clocks for those models. Recall what I said then. I said those target clocks were impossible on a 95W envelope for 8-core Zen. I said then that engineers only could get around 3GHz on a 95W envelope for 8-cores. And that is just what happened. Engineers forced the clocks on the three R7 models to the desired target and this increased the TDP above the marketing values. It is exactly the same that happened with the infamous PCIe issue on AMD cards. The hardware didn't behave as engineers expected and then increased clocks on the cards, which forced a violation of the PCIe spec, as anyone knows now.
9)
Since the violation of the TDP is happening in the top part of the frequency range of the 14LPP node. One has to ask why the 1800X/1700X violate the marketing TDP but the ThreadRipper models with similar clocks don't violate the TDP. The answer is very funny. The original engineering samples of ThreadRipper also violated the marketing TDP. The marketing TDP was 180W but
the actual ThreadRipper samples had a real TDP >200W.
So how does magically the final ThreadRipper CPUs stay within the official TDP? AMD engineers use a trick, one old. The trick consists on the TR4 socket having a power-limit mechanism that reduces the clocks
under the base clocks to force consumption to remain under the 180W limit defined for the platform:
If we look at the consumption ATX12V, we see that at rest, consumption is particularly reduced. In charge of several cores we are also quite far from the TDP of 180W, which makes us wonder whether these processors would not draw some of their power from the 24-pin ATX connector. Because when we look at the consumption of the platform, the idle consumption is significantly higher than for Ryzen platforms which is not justified by anything in the case of the Gigabyte we used. While the use of a 360mm watercooling and its three fans can generate a surcharge, but it does not seem to us alone to offset the difference noted.
In all cases, we note that we are far from consuming twice what uses a 1800X! And whatever happens this time, we hold in the TDP ... with almost identical consumption? By what miracle? Under x264, we could see that in practice, we were not exactly at the frequency that we could expect on the 1950X:
It is actually under 3.7 GHz, and even below the base frequency of [3.4] GHz. We could see such a throttling under several applications on the 1950X although it is far from systematic, the applications using AVX seem the most affected, logically. The limitation in our case, after many checks was not thermal, but related to consumption. AMD has confirmed to us that such a throttling was theoretically possible without completely confirming it. The 1920X did not exhibit such behavior in our tests, logically enough.
Somehow, it's a shame to see such a throttling even if in absolute terms, it is rather happy to see the Threadripper hold their TDP (a passing at the 1800X multiplied by two would have been complicated to pass). Going under the basic frequency is however something embarrassing, even if it is not the first time that one sees this behavior at AMD.
and here is the proof of the clocks on the 1950X running under the advertised base frequency
As HFR mentions, this is not the first time that AMD uses the trick of advertising false base clocks to maintain TDPs under control. The same happened with former Kaveri/Godavari APUs. Next link we have people complaining in AMD support forums about how the advertised chip and the real chip are different because the base clocks aren't real base clocks and the chip underclocks to maintain TDP within the official value
https://community.amd.com/thread/191127
At least Intel, in the link/image given above in this thread confirms that Intel warrants that its CPUs can maintain the base clocks at the rated TDP, unlike AMD, which violates this elementary requirement, sometimes.
10)
GamerNexus is not measuring TDPs. First, they are measuring power consumption at the 12.3V cable. One has to eliminate the measurement loses and the power consumed by the VRM to get the real power consumed by the CPUs. They don't give any of those values.
Second, it is well-known now that some initial X299 boards had problems with overheating of the VRM, which increased the total power cosnumed by the plattform. Many initial reviews also used beta BIOS with not working turbo/power policies what increased power consumption. It is also well-known all those power/temperature problems were solved with newest mobo iterations and final BIOS. After the Blender graph given above, GamerNexus studied the imnpact of different mobos/BIOS on power consumption. One can see a bad BIOS/mobo combination can vary the power consumed by so much as 60W, which is a lot of.
We also know that GamerNexus doesn't know how to measure stock power consumption. Burke got wrong power measurements for 7900X and lately wrong measurements for CoffeeLake. Burke didn't know he was testing with autoverclock settings. :-D
Now let us check power measurements of SKL-X CPUs using the same methodology used for RyZen, ThreadRipper, Piledriver, Bulldozer, Phenom, Broadwell, and Kabylake.
It is easy to check that i9-7900X, 7960X, 7980XE all them verify their TDP.
i9-7900X: 150W * 85% = 127.5W, which is witinh the 140W TDP.
i9-7960X: 175W * 85% = 148.75W, which is witinh the 165W TDP.
i9-7980XE: 168W * 85% = 142.8W, which is witinh the 165W TDP.
Conclusion)
So Phenom, Bulldozer, and Piledriver CPUs work under the advertised TDP. Intel CPUs work under the advertised TPD. Lower R5/R3 RyZen models also work under the advertised TDP. Top R7 RyZen models violate the advertised TDP by a large margin. And the top ThreadRipper models 12C and 16C also work under the advertised TDP, but the 16C model does by violating the advertised base clock.
Facebook
Twitter
Instagram