I live in Canary Islands.
Nice weather to go to beach everyday along the year.
But it is not so nice to high power electronic devices.
I´ve set a particular power boundary at 120 Watts for graphics cards I purchase, in order them to process most of the time without risk of spontaneous ignition.
No plans for air conditioning so far.

Several days ago, Toni released a limited amount of ACEMD3 test WUs.
Thank you very much, Toni. Linux users were waiting for this!
I realized that ACEMD3 test WUs might be an objective mean of testing and compare my currently working GPUs between them.
I´ve been lucky to catch at least one TONI_TESTDHFR206b WU for every GPU I have currently in production under 64 bits Linux OS, SWAN_SYNC enabled.
And here is a table comparing their performance.



Where:
GPU: Graphics card type. All of them are factory overclocked models.
Series: Graphics card micro-architecture.
Power: Card maximum rated power, as indicated by nvidia-smi command, in Watts.
Temp.: Maximum temperature reached while processing WU, as indicated by Psensor application, in ºC, at a room temperature of 26 ºC.
Time: Execution time for each WU, in seconds.
R.P. GTX: Relative performance, comparing execution time between each graphics card and the others.

ServicEnginIC, thanks for that! Good work :-)

Nice work!
Very informative.
I guess you don't spend much time at the beach with all this GPUgrid testing!!

I guess you don't spend much time at the beach with all this GPUgrid testing!!

GPUGrid is the most power exigent Project I currently process.

here are Psensor curves for the same card

There may be a problem with the WINDOWS Cuda 10.1 app - it's slower than it should be. We are working on it.

SWAN_SYNC is ignored in acemd3.

- CPU running cool at 51 degrees (100% utilization reported)

- GPU running warm at 80 degrees.

For comparison:
Here is the same GTX1660TI card mentioned below, now running Einstein@Home WUs due to Toni's current requirement for not to run acemd3 test WUs in Linux systems for better testing in Windows ones.





Spikes in temperature graph are the transition between two consecutive E@H WUs.
Comparing data, GPU power usage drops from 120W to 83W, and accordingly temperature from peak 80ºC to 69ºC.
Toni, this speaks very well about your acemd3 Linux code optimization for getting from GPU its maximum.
Now I'm crossing fingers for your success with Windows one!

E@H is most mining like BOINC project I've seen in that it responds better from memory OC than core OC. Try a math project that can easily run parallel calculations to push GPU temps.

SWAN_SYNC is ignored in acemd3.

SWAN_SYNC is ignored in acemd3.

Well, the CPU time equals the Run time, so could you elaborate this?
Could someone without SWAN_SYNC check their CPU time and Run time for ACEMD3 tasks please?

I still can squeeze a bit more the data from my tests:
I'll calculate energy consumed by every of my tested graphics cards, and order them in a new table.



In this table, increasing positions are more energy efficient cards.
As can be seen, GTX 1660 Ti coincides to be the most efficient (the less energy consumed for the same kind of task) and the most powerful (the less execution time)
And I take note for myself: For example, it would be a self-amortize idea to change GTX 950 graphics card by a GTX 1650 ASAP.
GTX 1650 has consumed less than half the energy than GTX 950 to finish the same kind of task!
Something to think about...

Nice stats.
The Turing GPUs are an exciting GPU for efficiency and power. (If only we could get some work for them....)
I will be buying the GTX1650 SUPER (November 22nd release) to replace my Maxwell cards, once GPUgrid starts releasing ACEMD3.

(If only we could get some work for them....)

How do we calculate energy consumed by GPU to process a certain Work Unit?
It is easy.
I entered my GPUGrid results page.
At current date:



Taking the execution time for my last processed Task: 59.720,55 seconds
It was processed in a GTX 1050 Ti graphics card.
Taking rated TDP (Thermal Design Power) for this GPU: 75 Watts
GPUGrid WUs are highly power demanding, and usually are able to lead GPU to its rated TDP consistently (in fact, even more if GPU is overclocked).
Therefore, aproximate energy consumed to process a WU can be calculated using this formula:
Energy (Watts x hour, Wh) = Rated GPU TDP (Watts) x WU execution time (seconds) / 3.600 (Seconds / hour)

In my example WU: 75 x 59.720,55 / 3.600 = 1.244,18 Wh (/ 1.000 = 1,24418 kWh)

the developement within the past weeks unfortunately was a steadily decreasing number of available tasks :-)

Hello ServicEnginIC;

As you are interested in lower end graphics cards, here is a post which I earlier posted in another post.
I wanted to post them here but I accidentally posted them in another post about more powerful GPUs.

DO NOT buy the GPUs mentioned below anymore for GPUGRID:
As the work units of GPUGRID will become bigger in the future, these low-end cards will not be able to meet the 5-day deadline anymore.
So I give you this information just as information, NOT as buying advice.

Here is a copy/paste.
You may use these data to add them into your list if you like:

If you like, I can give you some more data about other "low end" cards for the NEW ACEMD3 application:

These are all recent work units of last week. All were completed without errors.
You may add these to your list if you like.


1) NVIDIA GeForce GTX 745 Maxwell (2048MB 900Mhz DDR3 128-bit) - driver: 430.86:
CPU = AMD FX4320 quad core
Run TIme: 166,385.50 sec = +- 46 hours
See
http://www.gpugrid.net/show_host_detail.php?hostid=512409

2) NVIDIA GeForce GT 710 Kepler (2048MB DDR5 64-bit) - driver: 432.0
CPU = Core 2 Duo E8200 Dual core
Run Time: 375,108.98 sec = +- 104 hours
See
http://www.gpugrid.net/show_host_detail.php?hostid=502519

3) NVIDIA GeForce GT 710 Kepler (2048MB 800Mhz DDR3 64-bit) driver: 436.48
CPU = AMD A8-9600 Quad Core
Run Time = 431,583.77 sec = +- 119 hours
See
http://www.gpugrid.net/show_host_detail.php?hostid=514223

4) NVIDIA GeForce GT 1030 Pascal (2048MB DDR5 64-bit) - driver: 436.48
CPU = Celeron E3300 Dual Core
Run Time = 113,797.23 sec = +- 31 hours
See
http://www.gpugrid.net/show_host_detail.php?hostid=500299

5) NVIDIA GeForce GT 730 Kepler (2048MB DDR5 64-bit) - driver: 419.35
CPU = AMD Athlon 200GE Dual Core, Quad thread
Run time = 233,535.98 sec = +- 64 hours
See
http://www.gpugrid.net/show_host_detail.php?hostid=503400



Be aware of the following points which may have influenced the performance:
- I suppose that the ACEMD3 work units are not 100% identical in the amount of work they represent, so compare the results with a big grain of salt.
- All are Windows 10 machines.
- All computers have different processors.
- At the same time, the CPUs were crunching Mapping Cancer Marker work units, as many as they have cores/threads.
- 1 of my GT710s has DDR5 memory, the other one has DDR3 memory. I have stated it above.
- All GPUs and CPUs run at default speeds. No (manual) overclocking was done.
- None of the CPUs and GPUs ever reached 70 degrees Celsius.
In fact, except for the GT1030, all even stayed well below 60°.



Hope you like it. :-)
Greetings;
Carl Philip

And by the way, as you mentioned that you live in the Canary Islands and prefer low-power devices, here is some more info about my GPUs:

- The GT710 has a TDP of only 19 Watts for the DDR3 version, the DDR5 version consumes a little more.
- The GT730 has a TDP of 38 Watts for the DDR5 version, the DDR3 version consumes a little less.
- The GT1030 has a TDP of 30 Watts.
- The GTX745 has a TDP of 55 Watts.

So they are all low-power devices.
My preferred website to find info on GPUs is Techpowerup.
Just Google "GT710 Techpowerup" for example and you will find it.

I would not recommend buying these GPUs anymore unless for a PC with low requirement-applications like an office pc, an htpc or maybe for some light gaming with older games.

For example, the slowest of my cards, the GT710, can perfectly run an old but beautiful game called "Bioshock". I run that game on it in an old PC with a Core2Duo cpu, just 2Gb of RAM and Windows 10. It runs perfectly fluid.

But anyway, I would not recommend buying the GT710 or GT730 anymore unless you need their very low consumption.
The GT1030 is the fastest of them all.
The GTX745 is just a little slower than the GT1030, but it was sold to OEMs only. But like me, you can find it new or second hand on the internet. The Fujitsu model which I have even has a nice blower-style fan which blows the hot air directly outside the pc. There is both a normal model and a low-profile model.
You can find some pictures of it here:
https://picclick.fr/Fujitsu-GPU-NVIDIA-GeForce-GTX-745-2GB-FH-264493447418.html

If low-power is important to you, I can recommend using some of AMDs APUs with integrated video chips.
I have for example a rather old APU called AMD A8-9600.
Believe it or not, I currently play the latest Star Wars game "Star Wars Jedi: Fallen Order" on this old A8-9600. :-) (in 720p resolution)
But as these APUs have Radeon GPUs, you cannot use them for CUDA projects like GPUGRID. For OpenCl projects like Milkyway, you can use them.
And my A8-9600 runs circles around my GT1030 in Milkyway... strange but true. It's about 5 times faster than my GT1030... yes, 5 TIMES faster:
My GT1030 completes a Milkyway work unit in about 15 minutes, the A8-9600 does it in about 3 minutes.
That means that in Milkyway, the A8-9600 is probably even faster than a GTX1050.
That is what I call fantastic performance per Watt or per Dollar. :-)

Cheers from Cold Brussels ;-)
Carl

Thank you very much for your appreciations, Carl.
I've read them twice to the end.
Many points to comment:

My preferred website to find info on GPUs is Techpowerup.

But anyway, I would not recommend buying the GT710 or GT730 anymore unless you need their very low consumption.

The GTX745 is just a little slower than the GT1030, but it was sold to OEMs only

Thanks very much for the info.
I recently installed a couple of new Dell/Alienware GTX 1650 4GB cards and they are very productive. They run 60-65 deg.C and around 1850 MHz with a steady 4001 MHz DDR5 speed. I haven't tried overclocking as they appear to be running faster than NVidia advertises already.
At US$135 apiece (Ebay) and slot-powered, I'm pleased.

I recently installed a couple of new Dell/Alienware GTX 1650 4GB cards and they are very productive...

Thanks very much for the info.
I recently installed a couple of new Dell/Alienware GTX 1650 4GB cards and they are very productive. They run 60-65 deg.C and around 1850 MHz with a steady 4001 MHz DDR5 speed. I haven't tried overclocking as they appear to be running faster than NVidia advertises already.
At US$135 apiece (Ebay) and slot-powered, I'm pleased.

... the Nvidia drivers penalize all their consumer cards when a compute load is detected on the card and pushes the card down to P2 power state and significantly lower memory clocks than the default gaming memory clock and the stated card spec.

... the Nvidia drivers penalize all their consumer cards when a compute load is detected on the card and pushes the card down to P2 power state and significantly lower memory clocks than the default gaming memory clock and the stated card spec.

why so ?

... the Nvidia drivers penalize all their consumer cards when a compute load is detected on the card and pushes the card down to P2 power state and significantly lower memory clocks than the default gaming memory clock and the stated card spec.

why so ?

Because Nvidia doesn't want you to purchase inexpensive consumer cards for compute when they want to sell you expensive compute designed Quadros and Teslas.
No reason other than to maximize profit.

Depends on the card and on the generation family. I can overclock a 1080 or 1080Ti by 2000Mhz because of the GDDR5X memory and run them at essentially 1000Mhz over official "graphics use" spec.

And not generate a single memory error in a task. The Pascal generation really hamstrung the memory in P2 mode. The Pascal cards that used plain GDDR5 memory can not overclock as well as the GDDR5X memory cards. So you can't push the 1070 and such as hard.

But the Turing generation family does not impose such a large deficit/underclock in the P2 state under compute load.

The Turing cards only drop by 400Mhz from "graphics use" memory clocks.

What happens when there is a memory error during a game? Wrong color, screen tearing, odd physics on a frame? The game moves on. So what.

What happens when there is a memory error during compute? Bad calculations that are typically carried through in the task producing invalid results.

Down clocking evidently can reduce some errors so that's what NV does.

What happens when there is a memory error during a game? Wrong color, screen tearing, odd physics on a frame? The game moves on. So what.

What happens when there is a memory error during compute? Bad calculations that are typically carried through in the task producing invalid results.

Down clocking evidently can reduce some errors so that's what NV does.

What happens when there is a memory error during a game? Wrong color, screen tearing, odd physics on a frame? The game moves on. So what.

What happens when there is a memory error during compute? Bad calculations that are typically carried through in the task producing invalid results.

Down clocking evidently can reduce some errors so that's what NV does.

In my opinion Nvidia does this not because of concern for valid results, it's because they don't really want consumer GPUs used for compute. They'd rather sell you a much more expensive Quadro or Tesla card for that. The performance penalty is just an excuse.

What happens when there is a memory error during a game? Wrong color, screen tearing, odd physics on a frame? The game moves on. So what.

What happens when there is a memory error during compute? Bad calculations that are typically carried through in the task producing invalid results.

Down clocking evidently can reduce some errors so that's what NV does.

In my opinion Nvidia does this not because of concern for valid results, it's because they don't really want consumer GPUs used for compute. They'd rather sell you a much more expensive Quadro or Tesla card for that. The performance penalty is just an excuse.

exactly this.

On Apr 5th 2020 | 15:35:28 UTC Keith Myers wrote:

...Nvidia drivers penalize all their consumer cards when a compute load is detected on the card and pushes the card down to P2 power state and significantly lower memory clocks than the default gaming memory clock and the stated card spec.

Keith forgot to mention that the P2 penalty only applies to higher end cards. low end cards in the -50 series or lower do not get this penalty and are allowed to run in the full P0 mode for compute.

you cannot increase the performance level of the GTX 1660Ti from P2 under compute loads. it will only allow P0 if it detects a 3D application like gaming being run. this is a restriction in the driver that nvidia implements on the GTX cards -60 series and higher. there was a way to force the GTX 700 and 900 series cards into P0 for compute using Nvidia Inspector on Windows, but I'm not sure you could do it on Linux.

The only thing you can do is to apply an overclock to the memory while in the P2 state to reflect what you would get in the P0 state. there is no other difference between P2 and P0 besides the memory clocks.
____________

Down clocking evidently can reduce some errors so that's what NV does.

In my opinion Nvidia does this not because of concern for valid results, it's because they don't really want consumer GPUs used for compute. They'd rather sell you a much more expensive Quadro or Tesla card for that. The performance penalty is just an excuse.

They simply want to sell you a Quadro or Tesla and have you use it for its intended purpose which is compute. Those cards are clocked significantly less than the consumer cards so that they will not produce any compute errors when used in typical business compute applications.

Many thanks to all for your input. I Just finished a PABLO WU which took a little over 12 hrs on one of my GTX1650 GPUs.
I guess long runs are back.

I find myself agreeing with Mr Zoltan on this. I am inclined to leave the timing stock on these cards as I have enough problems with errors related to running two dissimilar GPUs on my hosts. When I resolve that issue I might try pushing the memory closer to the 5 GHz advertised DDR5 max speed on my 1650's.
What I'd like to know is what clock speeds other folks have had success with before risking blowing WUs finding out.

...I Just finished a PABLO WU which took a little over 12 hrs on one of my GTX1650 GPUs.
I guess long runs are back.

...I have enough problems with errors related to running two dissimilar GPUs on my hosts.

Can I use it on multi-GPU systems?

In general yes, with one caveat: if you have DIFFERENT types of NVIDIA GPUs in the same PC, suspending a job in one and restarting it in the other will NOT be possible (errors on restart). Consider restricting the client to one GPU type only ("exclude_gpu", see here).

Thanks, ServicEnginIC. I always try to use that method when I reboot after updates. I need to get myself a good UPS backup as I live where the power grid is screwed up by the politicians and activists and momentary interruptions are way too frequent.

As soon as I can afford to upgrade from my 750ti I will put both my 1650's in one host. Am I correct that identical GPUs won't have the problem?

I've noticed that ACEMD tasks which have not yet reached 10% can be restarted without errors when the BOINC client is closed and reopened.

Am I correct that identical GPUs won't have the problem?

I've noticed that ACEMD tasks which have not yet reached 10% can be restarted without errors when the BOINC client is closed and reopened.

Yes you can stop and start a task at any time when all your cards are the same.
I moved cards around so that all three of my 2080's are in the same host.
No problem stopping and restarting. No tasks lost to errors.

I can't do that on the other host because of dissimilar cards.

Yes you can stop and start a task at any time when all your cards are the same.

By the way, muchas gracias to ServicEnginIC for hosting this thread! It's totally synchronous with my current experiences as a cruncher. 🥇👍🥇

An unrelated note: BOINC sees a GTX 1650 as device 0 when paired with my 1060 3GB, even though it is slightly slower than the 1060 (almost undetectable running ACEMD tasks).
I think it might be due to Turing being newer than Pascal. I don't see any GPU benchmarking done by the BOINC manager.

Pop Piasa wrote:

By the way, muchas gracias...

BOINC will order Nvidia cards based on CC (compute capability), and then by how much memory it has.

Thanks for your generous reply Ian & Steve, you guys have a great working knowledge of BOINC's innards!
It pays for me to come here and ask, as a newbie in distrib. computing.

Great that you brought your talents and hardware over to this project when they stopped SETI. (Hmm... does shutting down the search mean that they gave up, or that the search is complete?)👽

Great that you brought your talents and hardware over to this project when they stopped SETI. (Hmm... does shutting down the search mean that they gave up, or that the search is complete?)👽

they feel like 20 years of listening to the northern sky was enough for now.

SETI is written in Distributed Computing history with bold letters.

On Apr 18th 2020 | 22:34:50 UTC Pop Piasa wrote:

...I Just finished a PABLO WU which took a little over 12 hrs on one of my GTX1650 GPUs.
I guess long runs are back.

I also received one of this PABLO WUs... and by any chance it was assigned to my slowest GPU currently in production.
It took 34,16 hours to complete in a GTX750, anyway in time to get half bonus.

If anyone has run a PABLO designed ACEMD task on a 1050 or other low power GPU please share your results.

If anyone has run a PABLO designed ACEMD task on a 1050 or other low power GPU please share your results. I would also like to see how these compare to a 2070-super as I am considering replacing the 750ti with one.

my 2080s do them in about ~13000 seconds = ~3.6 hours.
my 2070s do them in about ~15500 seconds = ~4.3 hours.

so maybe around 4hrs on average for a 2070 super.

I also received one of this PABLO WUs... and by any chance it was assigned to my slowest GPU currently in production.
It took 34,16 hours to complete in a GTX750, anyway in time to get half bonus.

Here you go, ServicEngineIC:

GTX 1650 gpu@1860MHz/ mem@4001MHz
(Dell Optiplex 980/i7-860 @2926MHz/ 12GB PC3-10600 @1200MHz)
45,400.71sec =12.6Hrs
http://httpd://www.gpugrid.net/workunit.php?wuid=19582490

GTX 1650 gpu@1860MHz/ mem@4001MHz
(ASUS Prime Z-270/i7-7700K @4200MHz/ 16GB PC4-2132 @3000MHz)
44,027.23sec =12.2Hrs
https://www.gpugrid.net/workunit.php?wuid=19495626

Here you go, ServicEngineIC for the Pablo tasks.

https://www.gpugrid.net/result.php?resultid=24501891
RTX2080@1805Mhz mem@14400Mhz
14,864 seconds = 4.13hrs
AMD Ryzen 3950X@4100Mhz @3600Mhz CL14

https://www.gpugrid.net/result.php?resultid=24502927
RTX2070@1700Mhz mem@14400Mhz
13,911 seconds = 3.87hrs
AMD Threadripper 2920X@4100Mhz @3466Mhz CL14

Ok!
Thank you!
I'll try to throw a new comparative table on this Sunday April 26th, with all data that I receive until Saturday 23:59 UTC...
Data, Data, Crunch, Crunch, 🔥🔥

Well, I tried to manipulate a PABLO WU over to my gtx 750ti by suspending it at 0.23% on my gtx1650, and when it restarted on the 750ti it crashed, falsifying my theory that ACEMD tasks can all be restarted before 10%. Perhaps it is still true running MDAD tasks.

Dmitriy Otroschenko's gtx 1060 3GB ran it:
39,016.09sec = 10.84Hrs.
AMD Ryzen 7 2700X Eight-Core Processor

I'm currently in the middle of 3 more PABLO tasks (almost as if Grosso knows I want them somehow), one on each of the 3 types of GPUs on my hosts. Not sure about being finished on the 750ti by midnight, ServicEnginIC.

i think the PABLO tasks cannot be restarted at all, even in a system with identical GPUs. I had a GPU crash the other day, and after rebooting a PABLO task tried to restart on a different GPU, and it failed immediately saying that it cannot be restarted on a different device. I haven't had that problem with the MDAD tasks though. they seem to be able to be stopped and restarted pretty much whenever, even on another device if it's the same as the one it started on, 0-90% without issue.
____________

I still am running with the switch timers set at 360 minutes. So don't step away from a running task for six hours. Think that might have covered my Pablo tasks but I never have changed my settings even after moving to an all RTX 2080 host.

Here are more PABLO results for your database, ServicEnginIC-

My GTX 1650 on ASUS Prime:
44,027.23sec = 12.23hrs
https://www.gpugrid.net/workunit.php?wuid=1949562

My GTX 1060 3GB on ASUS Prime
37556.23sec = 10.43hrs
https://www.gpugrid.net/workunit.php?wuid=19495626

1 PABLO task still running.

Here are more PABLO results for your database, ServicEnginIC-

My GTX 1650 on ASUS Prime:
44,027.23sec = 12.23hrs
https://www.gpugrid.net/workunit.php?wuid=1949562

My GTX 1060 3GB on ASUS Prime
37556.23sec = 10.43hrs
https://www.gpugrid.net/workunit.php?wuid=19495626

1 PABLO task still running.

I'll try to throw a new comparative table on this Sunday April 26th...

I'll try to throw a new comparative table on this Sunday April 26th...

I'm waiting to pick up data for better complete my original table.
Today I catched a new PABLO WU, and it is running now in a GT1030.
I still miss data for my GTX950 and GTX1050Ti...
I've got from you data for GTX750Ti and GTX1650, among several other models. Thank you!

I stumbled on this result while tracing a WU error history.

https://www.gpugrid.net/workunit.php?wuid=19660936

RTX 2080ti/Threadripper 3970X 🛸
13,616.88sec = 3.78hrs

For comparison against low power/low heat setups. Would like to see a Titan RTX result also...

I stumbled on this result while tracing a WU error history.

https://www.gpugrid.net/workunit.php?wuid=19660936

RTX 2080ti/Threadripper 3970X 🛸
13,616.88sec = 3.78hrs

For comparison against low power/low heat setups. Would like to see a Titan RTX result also...

e8s30_e2s69p4f100-PABLO_UCB_NMR_KIX_CMYB_5-0-5-RND3175

RTX 2080ti/Threadripper 3970X 🛸
13,616.88sec = 3.78hrs = 3h 46m 56.88s

My latest Pablo on one of my RTX 2080's

https://www.gpugrid.net/workunit.php?wuid=19871635

Run time CPU time
13,218.66 13,202.01
3.67 hours

Helps if you don't over commit your cpu.

For future reference,
The OP requested a GPU TDP of 120W, but if he raised the TDP requirements by only 5 more watts, in Linux, this could include many of the RTX series.

A 2060 does the same 7Tflops at 125W, as stock, while a GTX 1660Ti does only 5.4 Tflops at 120W.
A GTX 1660Ti can be brought lower in power, and increase overclock; resulting in the same performance at (an estimated) 85W. Saves you about $40 a year on electricity.

See post here:
http://www.gpugrid.net/forum_thread.php?id=5113

A GTX 1660Ti can be brought lower in power, and increase overclock; resulting in the same performance at (an estimated) 85W. Saves you about $40 a year on electricity.

A GTX 1660Ti can be brought lower in power, and increase overclock; resulting in the same performance at (an estimated) 85W. Saves you about $40 a year on electricity.

Thank you for your tip.
I add it to my list of "things to test"...

I've read your post, with exhaustive tests, at ProDigit's Guide: RTX (also GTX), increasing efficiency in Linux thread.
Good job!

When last ACEMD3 WUs outage arrived, I thoght it would last longer.
So, I searched for new projects in order to maintain active my GPUs, and I joined PrimeGrid.
Also, as I concluded on a previous post, I had decided to retire of production a Pascal GTX 950 GPU based graphics card, replacing it by a more efficient Turing GTX 1650 one.
I had tested four different graphics cards based on the same GTX 1650 GPU, and one dilemma arose: What would be the most convenient model for me to repeat?
- Model #1: PNY VCG16504SFPPB-O
- Model #2: ASUS TUF-GTX1650-4G-GAMING
- Model #3: ASUS DUAL-GTX1650-O4G
- Model #4: ASUS ROG-STRIX-GTX1650-O4G-GAMING

Finally, I chose a second second card of Model #4:
I've already given my reasons on [AF>Libristes] hermes Question for new gtx / rtx thread.

And as a complementary information (in close contact with this "Low power GPUs performance comparative" thread), I built a comparative table between the four tested models.
The main data for this comparative come from PrimeGrid's "Generalized Fermat Prime Search (n=20)" CUDA tasks execution.
Theese tasks are very similar in the amount of calculations required, thus being their execution times a good base for GPU's performance comparative.



Conclusions:
- Model #1: is the (negative) winner for the hottest card... But the (possitive) winner for being the most power efficient one (the least energy consumed per WU).
- Model #2: is the least factory overclocked, and therefore it is the lowest performance one, but it is rocky stable and easy to maintain cool.
- Model #3: is probably penalized for being mounted in a PCIE rev. 2.0 - DDR3 motherboard. Its boost clock is higher than the Model #1: one, but the performance for this last is higher (PCIE Rev. 3.0 - DDR4 motherboard).
- Model #4: power efficiency is penalized due to a factory TDP increase, in order to achieve higher boost clocks. But because of this, it is the winner for the highest performance (the lowest execution time).

Nice analysis of your gtx1650 GPUs. Always enjoy your technical hardware posts.

An alternate conclusion on the data presented:
For Low Power GPUs, Model #1 can be the winner if you have the time or inclination to adjust the Fan speeds of the GPU, and optimise the air flow in the Host chassis.
Model #4 offers a 2% gain in speed, but at a penalty of 11% gain in power consumption.
Considering you are comparing a "VW Polo" (PNY) to a "Ferrari" (ASUS ROG) of the GPU world, the "VW Polo" did pretty well.

It gets down to personal choice as to whether cost, raw speed, power efficiency or a combination of these factors is the goal.