Can you help me make a comprehensive efficiency graph for CPUs?

[Skeptyka]

I'm working on creating a graph showing how different CPUs perform across their entire range of supported wattages. Just like this:

---

If you're willing to help, please download CPU-Z
Run the very brief benchmark at different wattages, from the lowest to the highest.

Then simply write down the Single Thread and Multi Thread scores next to the wattage that has been used for the test. Thank you so much! You don't necessarily have to fill out all the wattage increments, any effort is appreciated!

I'll be adding the results here: https://www.canva.com/design/DAGQAy...share&utm_medium=link2&utm_source=sharebutton

---

Template:

CPU Model:

5W Single Core:
5W Multi-core:
10W Single Core:
10W Multi-core:
15W Single Core:
15W Multi-core:
20W Single Core:
20W Multi-core:
25W Single Core:
25W Multi-core:
30W Single Core:
30W Multi-core:
35W Single Core:
35W Multi-core:
40W Single Core:
40W Multi-core:
45W Single Core:
45W Multi-core:
50W Single Core:
50W Multi-core:
55W Single Core:
55W Multi-core:
60W Single Core:
60W Multi-core:
65W Single Core:
65W Multi-core:
70W Single Core:
70W Multi-core:
75W Single Core:
75W Multi-core:
80W Single Core:
80W Multi-core:
85W Single Core:
85W Multi-core:
90W Single Core:
90W Multi-core:
95W Single Core:
95W Multi-core:
100W Single Core:
100W Multi-core:
105W Single Core:
105W Multi-core:
110W Single Core:
110W Multi-core:
115W Single Core:
115W Multi-core:
120W Single Core:
120W Multi-core:
125W Single Core:
125W Multi-core:
130W Single Core:
130W Multi-core:
135W Single Core:
135W Multi-core:
140W Single Core:
140W Multi-core:
145W Single Core:
145W Multi-core:
150W Single Core:
150W Multi-core:
155W Single Core:
155W Multi-core:
160W Single Core:
160W Multi-core:
165W Single Core:
165W Multi-core:
170W Single Core:
170W Multi-core:
175W Single Core:
175W Multi-core:
180W Single Core:
180W Multi-core:
185W Single Core:
185W Multi-core:
190W Single Core:
190W Multi-core:
195W Single Core:
195W Multi-core:
200W Single Core:
200W Multi-core:
205W Single Core:
205W Multi-core:
210W Single Core:
210W Multi-core:
215W Single Core:
215W Multi-core:
220W Single Core:
220W Multi-core:
225W Single Core:
225W Multi-core:
230W Single Core:
230W Multi-core:
235W Single Core:
235W Multi-core:
240W Single Core:
240W Multi-core:
245W Single Core:
245W Multi-core:
250W Single Core:
250W Multi-core:
255W Single Core:
255W Multi-core:
260W Single Core:
260W Multi-core:
265W Single Core:
265W Multi-core:
270W Single Core:
270W Multi-core:
275W Single Core:
275W Multi-core:
280W Single Core:
280W Multi-core:
285W Single Core:
285W Multi-core:
290W Single Core:
290W Multi-core:
295W Single Core:
295W Multi-core:
300W Single Core:
300W Multi-core:

 

[AusWolf]

Top class CPUs already have such charts on TPU, under separate reviews from the main one.

 

[ShrimpBrime]

Cpu-z is too light of a work load. I recommend people use CBr23 or something heavy instead.

14700K with my 4000mhz 20 cores profile which is 150w (with some swing depending on app) Cpu-z reported only 135w.

Second Pic (recently run as well) is Only 8p 16t (no e-cores) pulling 320w running 6145mhz - The Cpu-z wattage would probably be lower than 300w. Unfortunately, this one doesn't fit on the list

 

[W1zzard]

How do you measure power? Hint: the software sensor is wrong, motherboards cheat it all the time

 

[ShrimpBrime]

How do you measure power? Hint: the software sensor is wrong, motherboards cheat it all the time

Well, I use MDTech Plus purchased through my Matco Tools dealer guy with the truck. It's pretty accurate, I use it for Dice and LN2 pot temps also. Have for many years now.

HWInfo 64 is actually pretty close. edit: which is what benchmate uses for recording wattage, voltage, temps ect for that "verification" screenshot.

But if you don't believe that pulled 320w, you should come over and watch the light dim and hear the A/C window unit bog down. It's quite impressive.

 

[chrcoluk]

Cpu-z is too light of a work load. I recommend people use CBr23 or something heavy instead.

14700K with my 4000mhz 20 cores profile which is 150w (with some swing depending on app) Cpu-z reported only 135w.

Second Pic (recently run as well) is Only 8p 16t (no e-cores) pulling 320w running 6145mhz - The Cpu-z wattage would probably be lower than 300w. Unfortunately, this one doesn't fit on the list

View attachment 362726

View attachment 362727

I think there is no right or wrong way, whats important the same tool is used every time.

CB e.g. may be heavier but its not representative of typical consumer workloads, I know CPUz isnt either, but as CPUz runs much faster its way more of a convenient test.

 

[ShrimpBrime]

I think there is no right or wrong way, whats important the same tool is used every time.

CB e.g. may be heavier but its not representative of typical consumer workloads, I know CPUz isnt either, but as CPUz runs much faster its way more of a convenient test.

Well... I dunno. It's a lot of doing something I'd rather not be doing lol. I could argue this and that, but I'm yet to figure out the purpose of 10 people with the same hardware popped in all of a sudden, what we'd do with this information.

Do you guys even care about my badass 150w configuration at all??? Eh? EH? EEEHHH???

No? well, that's my point.

 

[claylomax]

CPUz benchmark is very unreliable. Any Cinebench except the last one is better.

 

[Zach_01]

How do you measure power? Hint: the software sensor is wrong, motherboards cheat it all the time

True
Although it’s not the sensor “wrong”, but the board and you can’t always call it cheating and I will explain.

I have observed many AM4 systems since this is known.
On many Zen2/3 systems, boards overstate current/power and others even under state it with the latter being more common maybe:
Maybe half of them are providing close to reality feedback to the CPU.
Unfortunately this can be confirmed only on 100% CPU load and not on 1-99%.
Look for the Power Reporting Deviation sensor on HWiNFO.

On my board when I had the R5 3600 installed PRD was hovering around 90% under 100% load which means that the board was telling the CPU that it draws 90W when in reality was drawing 100W (you can call that cheating).

I’ve seen a lot of boards with PRD up to 125% and in this opposite case the CPU is under performing from specs.

With my current CPU the PRD sensor reports around 100-101% so PPT is pretty much accurate reading… but again this is only for 100% load. Anything between is unreliable.

EDIT:
And also PRD under 100% CPU load can be different with different type of load.
It’s a mess really.

 

[chrcoluk]

Well... I dunno. It's a lot of doing something I'd rather not be doing lol. I could argue this and that, but I'm yet to figure out the purpose of 10 people with the same hardware popped in all of a sudden, what we'd do with this information.

Do you guys even care about my badass 150w configuration at all??? Eh? EH? EEEHHH???

No? well, that's my point.

Oh I just had a look at your post, yeah not bad 29k with 150w. I think I am managing 29k on 170w or so.

CPUz benchmark is very unreliable. Any Cinebench except the last one is better.

I get scores within 1% or so every time I run it.

 

[ShrimpBrime]

Oh I just had a look at your post, yeah not bad 29k with 150w. I think I am managing 29k on 170w or so.


I get scores within 1% or so every time I run it.

The e-cores at 4000mhz sometimes boot at 1.24v at load and sometimes runs 1.18v at load. I think CBR23 made the VID jump from the load. May had been the difference between cpu-z and that bench.

 

[Bill_Bright]

CPUz benchmark is very unreliable.

It is not CPUz that's to blame. Software based monitors must rely on the sensors built into the hardware. And those sensors probably cost a $1 for a dozen. The only way to properly measure power consumption is with actual test equipment, such as a power meter or multimeter in series measuring current.

 

[ShrimpBrime]

It is not CPUz that's to blame. Software based monitors must rely on the sensors built into the hardware. And those sensors probably cost a $1 for a dozen. The only way to properly measure power consumption is with actual test equipment, such as a power meter or multimeter in series measuring current.

I am Quoting Martin from HWInfo64 here. This is not my statement and using it for informational purposes only!

Retrieving of such information is very complex and very specific - there are several different kinds of sensors, which very different access methods and interfaces. There's no unified and universal standard to specify how to read information from all sensors (well, perhaps IPMI, but that's only rarely used on some server machines).
To describe the full variety of sensors available and how they are read would be very complex. Just to name one example - mainboard sensors. These are usually implemented via dedicated LPC/Super-IO chips or dedicated sensor monitoring chips on mainboard. So the tool needs to know how to detect and access such chips, they can be accessible via SMBus or ISA/LPC space. Then it's required to know how to retrieve sensor information from them. Such chips usually offer several temperature, voltage and fan speed inputs, but these are exposed as general values (e.g. temperature 0, 1, etc.) only. Then it depends on particular mainboard how/where it connects the particular sensor inputs. For example mainboard A can have temperature #0 connected to CPU socket diode, temperature #1 to a local mainboard sensor. Other mainboard model can have this different. And again, there's no standard that would tell us where such general sensor is connected, so we need to examine every mainboard model to understand the connection and implement such adjusting into the monitoring tool.
Because of the lack of unified standards and the large variety and complexity of sensors, there is no dedicated support in operating systems like Windows, one needs to rely on specific tools like HWiNFO.

How does HWiNFO app gather info from computers

Hi everybody, Might be a stupid question but I was hoping to get a better idea as to how the HWiNFO is gathering it's data for individual counters. Specifically am curious about the temperature sensors for CPU cores, Ambient, DIMM and I've got something that is called Temp3 with a reading...

www.hwinfo.com

 

[Vayra86]

CPU efficiency = type of workload.

The harsh reality is that the overwhelming majority of typical workloads a CPU gets to swallow is of highly varied nature, not parallelized (like Cinebench) and generally the CPU is running numerous threads of different services, applications and stuff all at the same time.

How do you measure efficiency. Look at an X3D. They game at 4800-5000mhz, but they take just 40-60W (not even full TDP) for that. Just clockspeed isn't going to be fine grained enough. Clocks & cores then? Hmm. Is SMT on? Etc.

You need a highly varied, combined benchmark suite result number to say anything meaningful about CPU efficiency. As for the graph, that's just plotting numbers on an X/Y. The real information is how those numbers are obtained. And then, by running that exact same set of benchmarks to calculate the same kind of number on various CPUs, you can get an idea of relative efficiency. You also need to measure power at the wall and deduct system power.

 

[phanbuey]

when there are this many variables it's almost better to go top down, measure at the wall, lock it in by measuring idle as a baseline, then blast a workload and use the difference. Its not perfect but it will give pretty accurate real world results..

 

[Bill_Bright]

there are several different kinds of sensors, which very different access methods and interfaces. There's no unified and universal standard to specify how to read information from all sensors

And I totally agree with this too. All those sensors do is put out some hexadecimal number that is "supposed" to represent some value. Because these are very inexpensive, relatively low-tech devices, odds are, 5 identical sensors will output a different number. Yes, unless defective, they will be close, but still might vary several percentage points.

Then the major CPU makers put those sensors wherever they want. This is why we cannot directly compare an AMD to Intel in terms of heat for example.

And then the software developers (including BIOS developers), because there is no industry standard, must come up with their own method to interpret those values.

So I am saying when it comes to hardware monitors like HWiNFO, HWMonitor, CPUz, Speccy, whatever, they all have no choice but to rely on a sensor that cost pennies to make, cannot be calibrated, and is never tested for accuracy before leaving the factory.

So the best you can really hope for with these software based monitors is to be somewhere in the ball park.

So again, when it comes to precision accuracy you can rely on, it must be done with precision measuring test equipment.

 

[ShrimpBrime]

CPU efficiency = type of workload.

The harsh reality is that the overwhelming majority of typical workloads a CPU gets to swallow is of highly varied nature, not parallelized (like Cinebench) and generally the CPU is running numerous threads of different services, applications and stuff all at the same time.

How do you measure efficiency. Look at an X3D. They game at 4800-5000mhz, but they take just 40-60W (not even full TDP) for that. Just clockspeed isn't going to be fine grained enough. Clocks & cores then? Hmm. Is SMT on? Etc.

You need a highly varied, combined benchmark suite result number to say anything meaningful about CPU efficiency. As for the graph, that's just plotting numbers on an X/Y. The real information is how those numbers are obtained. And then, by running that exact same set of benchmarks to calculate the same kind of number on various CPUs, you can get an idea of relative efficiency. You also need to measure power at the wall and deduct system power.

I'd say that 4000mhz profile I have is really efficient compared to the stock boosting configuration. I imagine the "Base" profile could be improved. It's 3400mhz P-cores and 2500mhz e-cores. That's a very low base profile in my opinion. (for 14700K). I'd have to run it, but I think VID is like .980v or just on 1v with a touch of droop there. I'd have to double check that though.

So I am saying when it comes to hardware monitors like HWiNFO, HWMonitor, CPUz, Speccy, whatever, they all have no choice but to rely on a sensor that cost pennies to make, cannot be calibrated, and is never tested for accuracy before leaving the factory.

Just this part. AMD used algorythms in the past for cpu monitoring. I believe most sensors are done in this fashion simply because the board doesn't have an external temp sensor like a multi-meter. That means the internal sensors have nothing to compare to. This is the main reason for skew. (for temperature)

There has always been an on die temp sensor and off die temps sensor. In the past in the common place you just didn't know if it's a temp sensor on the board or on the cpu pcb. Still don't know. Have to read cpu white papers to find out exactly what sensor and where the location of them are.

Here's an example of a functional data sheet. You or someone else can take the liberty of finding the specifics of the temp sensors with this 1100+ page data sheet. I don't care to go through it again for this platform in particular, I just have it saved still...

https://www.amd.com/content/dam/amd/en/documents/archived-tech-docs/programmer-references/55072_AMD_Family_15h_Models_70h-7Fh_BKDG.pdf

I find HWInfo64 to be pretty accurate to the advertised wattage and settings to the CPU advertisement. Not sure why this would "Need" to be followed up with a multimeter unless you plan to actually care enough to do so. Therefor the temp and voltage monitoring apps are all you have.

Since Martin, the Author of HWinfo 64 uses the same approach as W1zard, and updates his program to accommodate all the different motherboards when released or brought to his attention, I feel are accurate ways of monitoring the system. This has been common place for many years. And tested with a multimeter plenty. Of course there's variances. Heat causes resistance. So many variables yes??

So HWinfo64 is within .050v accurate to my multimeter at the CPU EPS connector. Could take the time and measure all the FETs and so forth, but why. My rig is working fine and I trust HWInfo64 just like I trust GPU-Z

 

[Zach_01]

What is really skewed (my words), aka AMDs proprietary and undisclosed calculation method (HWiNFO Martin words) is almost everything you see on live values in Ryzen Master reports (temp, power, speed and so) …I hate that thing

HWiNFO sensors data is as real as it can get with software that polls every 1-2sec. By Martin, at least the “CPU enhanced” section of (most) sensors is accessing straight the CPU with no board middle stage if we can say this.

That’s why you can see differences between 2-3 sensors reporting the same thing.
For example:
CPU Vcore (board sensor)
VRVOUT (board VRM sensor)
CPU Core Voltage SVI2/3 TFN (Straight from CPU)

And with “Snapshot CPU Polling” method for Ryzen it also reduces to a minimum the “observer effect” (access = disturbance)

This software is many levels above anything else really.
Huge amount of work and development of many many years.
I encourage every regular user of HWiNFO to at least donate a small amount. It’s worth it.

 

[A&P211]

This is plain stupid thing to measure. Waste of time.

 

[Skeptyka]

How do you measure power? Hint: the software sensor is wrong, motherboards cheat it all the time

Interesting, thanks for sharing. Can you recommend evidence for me to look into? I'm wondering if it's innacurate enough across different types of software and hardware such that it makes zero sense to even attempt to compare efficiency like this. Or is it within the margin of error?

For me, I set a maximum wattage in ThrottleStop.

 

[W1zzard]

MDTech Plus

How do you measure power with that?

Since Martin, the Author of HWinfo 64 uses the same approach as W1zard

I don't think that's possible. I hook up test equipment to measure ATX12V voltage + current, both at teh same time, to derive power from that (in my CPU reviews). Using any kind of software sensor cannot provide the data needed, because there is no reliable/calibrated circuit to measure current (measuring voltage is fairly easy and within a few percent). Also look into TrueRMS, "voltage" is not "voltage"

Can you recommend evidence for me to look into? I'm wondering if it's innacurate enough across different types of software and hardware such that it makes zero sense to even attempt to compare efficiency like this. Or is it within the margin of error?

you can change the outcome of the measurement using BIOS settings, and mobo mfgrs will ship their boards with whatever settings they please, because it gives them more boost, because the CPU will boost higher, because it hits the power limit a bit later

Just to clarify, to make a measurement in the context of a review, where a purchase recommendation must be made, software sensors are not good enough. If you ask yourself "roughly how much more/less power is used, in my current setting, without changing any other variable" this could be usable.

 

[Bill_Bright]

Using any kind of software sensor cannot provide the data needed, because there is no reliable/calibrated circuit to measure current (measuring voltage is fairly easy and within a few percent).

^^^This^^^. Voltage is measured in parallel so is easy. In the simplest way, you just put your two meter probes on the end points and measure. But current is measured in series. This means you must break the circuit and insert the meter in circuit to complete the circuit again, causing all the current to flow through the probes, probe wires and meter.

It is no big deal with voltage to get an accurate reading but it is very easy to change the characteristics of the entire circuit when inserting another device (with long wires) in there to measure current. Capacitance, inductance and resistance is introduced in the circuit.

And remember, the basic power formula is P = I x E (power = current x voltage) so you need both values. This is why I said above, to get an accurate measure, you need to use precision measuring test equipment.

aka AMDs proprietary and undisclosed calculation method

Well, with no industry standards, every one uses proprietary methods. Disclosing them is another issue but even if AMD and Intel were totally transparent and fully disclosed the wheres and hows, they still could not be used for comparisons because each still does it differently.

The best you can do is take measurements to establish a baseline, then, if you suspect a problem, take measurements again to see what has changed.

 

[Zach_01]

^^^This^^^. Voltage is measured in parallel so is easy. In the simplest way, you just put your two meter probes on the end points and measure. But current is measured in series. This means you must break the circuit and insert the meter in circuit to complete the circuit again, causing all the current to flow through the probes, probe wires and meter.

It is no big deal with voltage to get an accurate reading but it is very easy to change the characteristics of the entire circuit when inserting another device (with long wires) in there to measure current. Capacitance, inductance and resistance is introduced in the circuit.

And remember, the basic power formula is P = I x E (power = current x voltage) so you need both values. This is why I said above, to get an accurate measure, you need to use precision measuring test equipment.

This can be an interesting read for some

IMPORTANT - Explaining the AMD Ryzen "Power Reporting Deviation" -metric in HWiNFO

Ryzen CPUs for AM4 platform rely on external, motherboard sourced telemetry to determine their power consumption. The voltage, current and power telemetry is provided to the processor by the motherboard VRM controller through the AMD SVI2 interface. This information is consumed by the processors...

www.hwinfo.com

After reading the above...
At least for AM4 and when the CPU is under 100% CPU load I believe that power (PPT) with accounting the Power Reporting Deviation metric can be very close to the real thing.
Anything else (PPT under 1-99% CPU load) is just a number comparable only to itself under different operating conditions (load variation). Its just to give you an idea of the CPU is consuming more or less. But the exact number is highly questionable.

I cant remember exactly how PRD is determining the true current (A) under 100% CPU load, I have to re-read that thread. Its been a while...

Well, with no industry standards, every one uses proprietary methods. Disclosing them is another issue but even if AMD and Intel were totally transparent and fully disclosed the wheres and hows, they still could not be used for comparisons because each still does it differently.

The best you can do is take measurements to establish a baseline, then, if you suspect a problem, take measurements again to see what has changed.

I cant argue.
Hardly anything (to operating parameters) is directly comparable between different designs of different manufacturers. Sometimes even within the same "house" you cannot compare them. Latest example is Zen5 with AMD claiming to "improve" temp sensor placement.

 

[ShrimpBrime]

How do you measure power with that?


I don't think that's possible. I hook up test equipment to measure ATX12V voltage + current, both at teh same time, to derive power from that (in my CPU reviews). Using any kind of software sensor cannot provide the data needed, because there is no reliable/calibrated circuit to measure current (measuring voltage is fairly easy and within a few percent). Also look into TrueRMS, "voltage" is not "voltage"


you can change the outcome of the measurement using BIOS settings, and mobo mfgrs will ship their boards with whatever settings they please, because it gives them more boost, because the CPU will boost higher, because it hits the power limit a bit later

Just to clarify, to make a measurement in the context of a review, where a purchase recommendation must be made, software sensors are not good enough. If you ask yourself "roughly how much more/less power is used, in my current setting, without changing any other variable" this could be usable.

Measure power? Power is wattage, so it's volts times amps. You only need to know those two facts to come up with power. The multimeter comes with an amp clamp if I wanted to get real in jiggy with it.

What I'm reading from you and Bill is not to use GPU-Z or trust it rather because the "sensor cannot provide data needed" because there's no calibrated circuit.

Thermistors work by resistance. Your car coolant temp sensor is a great example of a cheap sensor that gives accurate valid data. But wait there's 2 temp sensors!! Yeah. one for the dummy gauge and the other for ECM fuel trim.

I'll still use GPU-Z for my power readings. I trust your work and I trust Martins work. If there's actually a valid reason not to use your software for power readings, then I won't use it. not that big of a deal. But GPU-Z reported my 4070 super to pull 220w. I hope that's accurate without needing a machine or multimeter. I just want to use the card and game.

 

[W1zzard]

Measure power? Power is wattage, so it's volts times amps. You only need to know those two facts to come up with power.

Yup, what I don't understand is how you measure both at the same time, so you can multiply them?

What I'm reading from you and Bill is not to use GPU-Z or trust it rather because the "sensor cannot provide data needed" because there's no calibrated circuit.

I specifically talked about power reporting for CPUs from ANY software. For example, power reporting on NVIDIA graphics cards is quite accurate, because they have dedicated circuitry on the PCB that really measures current using a shunt resistor + voltage