The Performance tab was always a great source of information on gauging how each GPU was performing in comparison to other GPUs.

I have now resorted to scanning the Volunteers \ Hosts list for this info.

Below is a quick summary taken 2nd November 2020 2:00am UTC.

In this listing the Recent Average Credit (RAC) has been used for the comparison due to the variability of the work unit runtimes. Toni has stated that the credit formula has been uniform for several years.

This list is not definitive, just an indicator at best. There are many factors that could affect this listing.

NOTE:
Hosts with multiple GPUs have been exluded.
Recent Average Credit has been rounded
The best performing GPU from each type has been listed

the 2080ti TDP number is way high. most are 250w, some a little more. but 300 would be atypical.

personally i run my single one at 225W, and my 5x system at 215W each.
____________

I have now resorted to scanning the Volunteers \ Hosts list for this info.
Below is a quick summary taken 2nd November 2020 2:00am UTC.

the 2080ti TDP number is way high. most are 250w, some a little more. but 300 would be atypical.

personally i run my single one at 225W, and my 5x system at 215W each.

I have now resorted to scanning the Volunteers \ Hosts list for this info.
Below is a quick summary taken 2nd November 2020 2:00am UTC.

Great (and laborious) job!
Very interesting, thank you for this effort.

Finished the script and gathered some info on the first 3000 Hosts (Volunteers Tab) which I have listed below.

Hosts with Multiple GPUs: 320 (which are excluded from list)
List compiled: 5th November 2020 0:00UTC.

If you have a keen eye, you will see a RTX3070 and 2 RTX3090 GPUs. (Any RAC they have, would be from a compatible card that was removed)

Very much appreciate your effort, Rod4x4.

You've provided me with a better idea of how they stack up running MDAD tasks than I previously had. Thanks.

One thing that throws a wrench into rod4x4's comparison is the ADRIA-Bandit WU.

Every time one of my machines gets one, that host's RAC takes a beating.

I've also found that if restarted they throw an error, even on a single GPU host. That showed up during my last driver update. Don't know if anybody else had this happen.

I think Ian has mentioned the same.

I know I've seen that before, but don't remember specifically which ones. I think I saw that behavior on the PABLO tasks.
____________

I remember Ian mentioning it about ADRIAs when they first came out and comparing them to PABLO WUs, as they are similar in size but ADRIA WUs give around half the points that PABLOs did.

(I searched the forum for "ADRIA" and "Adria" but couldn't find it.)

Speaking of broken tabs, the donation page and the profile creation page haven't worked for me since I started last year. There seems to be a problem with the "I am not a robot" verification picture is missing.

One thing that throws a wrench into rod4x4's comparison is the ADRIA-Bandit WU.

Every time one of my machines gets one, that host's RAC takes a beating.

https://www.gpugrid.net/forum_thread.php?id=5125&nowrap=true#55153

Ian C wrote:

The VillinAdaptive WUs also pay a lot less credit reward as compared to the pablo tasks, per time invested.

Ian C wrote:

The VillinAdaptive WUs also pay a lot less credit reward as compared to the pablo tasks, per time invested.

Found it!
Now, these are labeled ADRIA_NTL9Bandit100ns.

It's been happening since long time ago.
I found this older "Immensive credits difference: ADRIA vs. PABLO tasks" thread even expresely mentioning it.

And regarding this rod4x4 thread "Performance Tab still broken" topic, I agree that Adria work units may be a good tool for a performance comparative, due to their consistence in number of calculations per task.

I'm all for that, as long as times are used instead of points per day for rating the cards. That would negate the effects that errors and WU granted credit variances have on a host's RAC. I see credit acquisition rates as a function of the entire host machine's output, rather than just the GPU.

I get it now, mate. Users can't access the level of data that Grosso can, and the existing server webpage interface is no longer providing pertinent data so it is essentially useless. I hope somebody on the team actually reads your post and corrects this problem.

Thanks for your kindly responses and thanks double for your contribution to benchmarking GPU performance while Processing ACEMD platform based tasks.

I have calculated that scanning 100 tasks for 20 hosts on each of the 150 volunteer pages scanned could take several days. As a comparison, a SQL query on the backend data would only take a few minutes to run (or less)

I've catched on past week in my systems some ADRIA work units, few, but enough for me to confirm some suppositions.

* Graphics card ASUS TUF-GTX1650S-4G-GAMING, based on GTX1650 SUPER GPU, running at PCIE gen 3.0 x16 slot, got three. Execution times (seconds): WU#1: 18628,53 ; WU#2: 18660,36 ; WU#3: 18627,83
* Graphics card ASUS DUAL-GTX1660TI-O6G , based on GTX1660 Ti GPU, running at PCIE gen 2.0 x16 slot, got two. Execution times (seconds): WU#1: 16941,52 ; WU#2: 16998,69
-1) So far, execution times for ADRIA tasks are relatively consistent when executed at the same card model and setup.

* Graphics card ASUS ROG-STRIX-GTX1650-O4G-GAMING, based on a factory overclocked GTX1650 GPU, got three. Execution times (seconds): WU#1: 23941,33 ; WU#2: 23937,18 ; WU#3: 32545,90
-2) Ooops!!! What's happened here? Execution time for WU#3 is very different to WU#1 and WU#2, being executed at the same graphics card model.
The explanation: WU#1 and WU#2 were executed on a card installed at a PCIE gen 3.0 x16 slot, while WU#3 was executed on a card installed at a PCIE gen 3.0 x4 slot.
Both are installed at the same multiGPU system, but due to mainboard limitations, only PCIE slot 0 is running at x16, while two more slots 1 and 2 are runing at x4, thus limiting performance for ADRIA tasks.
Conclusion: Execution times are not only showing a particular graphics card performance itself, but also its particular working conditions at the system where it is installed.
rod4x4, in my opinion, your decision to discard hosts with multiple GPUs makes complete sense due to this (and other) reason(s).

It's a pity that ADRIA work units availability is unpredictable, and usually very transient...

-3) rod4x4, thank you very much again for your labour.

More importantly, we would not have this dilemma if the Performance tab was working.

That GPUGrid tasks are heavily dependent on PCIE bus speeds has been posted and commented multiple times by Ian.

I too see a noticeable slowdown on all GPUG tasks when comparing the same RTX 2080 cards (x3) from the two cards at X8 speeds compared to the card running at X4 speed.

Not as big a difference as your test comparing X16 to X4 though.

Einstein sees the same kind of differences though not as extreme compared to GPUG.

From my testing with my cards (RTX 2070 and faster) I’ve determined that PCIe 3.0 x4 should be the lowest link you should use. Preferably PCIe 3.0 x8 or better for full speed crunching. I noticed no difference in 3.0x8 and 3.0x16.

When comparing link speeds make sure you’re accounting for both link gen and link width. Saying just x4 or x8 or x16 alone is rather meaningless. PCIe 2.0 is half the speed as 3.0 at the same width, and 1.0 is half the speed as 2.0 (1/4 speed as 3.0).

Also if you’re using an 3.0x4 slot on an Intel board, that’s likely from the chipset. In which case you will likely have LESS than 3.0x4 actually available to the device depending on what other devices are being serviced by the chipset. The DMI link between the chipset and the CPU only has the equivalent of PCIe 3.0x4 available for ALL devices total (other PCIe slots, SATA devices, networking, etc). You really won’t get the full speed from a chipset based slot because of this.

Don’t forget to account for CPU load also. If your CPU is maxed out, you’ll see slow and inconsistent speeds.
____________

Is anyone else seeing these?

0_2-GERARD_pocket_discovery_08522dfc_f243_41c4_b232_587af6264fe7-0-3-RND1589

Good point to emphasize where the slot is serviced by. My card at the bottom of the motherboard is a PCIE Gen2 X4 slot serviced by the chipset.

Only can manage 5.0 GT/second compared to the other cpu fed slots that can do 8.0 GT/second.

...

I have captured data from the top 1200 hosts - Volunteer / Hosts tab.

Breakdown of top 1200 host's data:
- 213 hosts have multiple GPUs (which are excluded from further analysis)
- 108,000 completed tasks captured from remaining 987 hosts
- 1322 completed tasks are ADRIA (1.2% of total tasks captured)

Script runtime: 2 hours 43 minutes (Bit longer than anticipated, but still reasonable when considering the volume of data captured)
Scan started: 11th November 2020, 23:20 UTC

Below is a summary of ADRIA task runtimes for each GPU type.

NOTE:
Sorted by fastest Average Runtime
All Run Times are in seconds

Same as previous post, this time for GERARD tasks

Only 129 GERARD tasks completed in the last 10 days for the hosts sampled. (Hosts keep task list for the last 10 days)

NOTE:
Sorted by fastest Average Runtime
All Run Times are in seconds

Thanks a million for your efforts here, rod4x4!

This IMHO is the most useful analysis yet.

Thanks a million for your efforts here, rod4x4!

This IMHO is the most useful analysis yet.

This IMHO is the most useful analysis yet.

This IMHO is the most useful analysis yet.

+1
My admiration for your work.

Too bad that BOINC doesn't individually identify multiple GPUs the way FAHCore does (slots), so data from those hosts could be used also. It would give an idea of how much multiple cards in a host suffer from performance loss.

Unfortunately, F@H has poor statistical reporting in comparison to BOINC.

From the same dataset used in the previous ADRIA and GERARD comparisons, here is a comparison of MDAD task using average runtime.

Due to the variablility of runtimes for the MDAD tasks, this comparison should not be taken too seriously. (although ranking is consistant with expections)

NOTE:
GPU types with less than 500 tasks have been excluded
Runtimes are in seconds

I saw where Toni has responded to ServiEngineIC on the wishlist thread (https://www.gpugrid.net/forum_thread.php?id=5025&nowrap=true#55709) so I trust that we we will be placated eventually and our toy restored to educate and amuse us.

Meanwhile, thanks to the loyal, diligent efforts of rod4x4 in this thread, we have a pretty good model of relative performance and efficiency on which to base a GPU buying decision.

Three virtual (social distanced) cheers for rod4x4! 🍺🍺🍺

I saw where Toni has responded to ServiEngineIC on the wishlist thread (https://www.gpugrid.net/forum_thread.php?id=5025&nowrap=true#55709) so I trust that we we will be placated eventually and our toy restored to educate and amuse us.

Meanwhile, thanks to the loyal, diligent efforts of rod4x4 in this thread, we have a pretty good model of relative performance and efficiency on which to base a GPU buying decision.

Three virtual (social distanced) cheers for rod4x4! 🍺🍺🍺

* Graphics card ASUS TUF-GTX1650S-4G-GAMING, based on GTX1650 SUPER GPU, running at PCIE gen 3.0 x16 slot, got three. Execution times (seconds): WU#1: 18628,53 ; WU#2: 18660,36 ; WU#3: 18627,83
* Graphics card ASUS DUAL-GTX1660TI-O6G , based on GTX1660 Ti GPU, running at PCIE gen 2.0 x16 slot, got two. Execution times (seconds): WU#1: 16941,52 ; WU#2: 16998,69
-1) So far, execution times for ADRIA tasks are relatively consistent when executed at the same card model and setup.

* Graphics card ASUS ROG-STRIX-GTX1650-O4G-GAMING, based on a factory overclocked GTX1650 GPU, got three. Execution times (seconds): WU#1: 23941,33 ; WU#2: 23937,18 ; WU#3: 32545,90
-2) Ooops!!! What's happened here? Execution time for WU#3 is very different to WU#1 and WU#2, being executed at the same graphics card model.
The explanation: WU#1 and WU#2 were executed on a card installed at a PCIE gen 3.0 x16 slot, while WU#3 was executed on a card installed at a PCIE gen 3.0 x4 slot.
Both are installed at the same multiGPU system, but due to mainboard limitations, only PCIE slot 0 is running at x16, while two more slots 1 and 2 are runing at x4, thus limiting performance for ADRIA tasks.
Conclusion: Execution times are not only showing a particular graphics card performance itself, but also its particular working conditions at the system where it is installed.

If anybody has time, look at my Host 514522. https://www.gpugrid.net/show_host_detail.php?hostid=514522

It's a
Z-270A ASUS Prime MB,
i7-7700K,
G.Skill DDR4-3200 (2x8GB),
Samsung Evo 970 1st gen. 500GB PCIE3 M.2 SSD (Win10 pro),
(2) ASUS Dual OC GTX 1060 3GB GPUs,

According to Afterburner's HW Monitor, my GPU1 (x16 slot) draws around 90-97 watts depending on the WU. GPU2, an identical (newer) card in the x8 slot below draws 95-105W.
The onboard Intel graphics (GPU3) are the windows graphics, leaving both GPUs 1&2 unhindered (or so I have assumed).

The combined RAC of the 3GB GTX 1060s is presently hovering around 589,000 (with 5 threads of WCG apps running on the CPU). I find that encouraging, yet I wonder what factors contributed the most to it.

Does the Slot have anything to do with the wattage? I know these cards get additional power from the PSU, but what I see here mystifies me. Should I assume that the GPU drawing more power is doing more work?

If anybody has time, look at my Host 514522. https://www.gpugrid.net/show_host_detail.php?hostid=514522

It's a
Z-270A ASUS Prime MB,
i7-7700K,
G.Skill DDR4-3200 (2x8GB),
Samsung Evo 970 1st gen. 500GB PCIE3 M.2 SSD (Win10 pro),
(2) ASUS Dual OC GTX 1060 3GB GPUs,

According to Afterburner's HW Monitor, my GPU1 (x16 slot) draws around 90-97 watts depending on the WU. GPU2, an identical (newer) card in the x8 slot below draws 95-105W.
The onboard Intel graphics (GPU3) are the windows graphics, leaving both GPUs 1&2 unhindered (or so I have assumed).

The combined RAC of the 3GB GTX 1060s is presently hovering around 589,000 (with 5 threads of WCG apps running on the CPU). I find that encouraging, yet I wonder what factors contributed the most to it.

Does the Slot have anything to do with the wattage? I know these cards get additional power from the PSU, but what I see here mystifies me. Should I assume that the GPU drawing more power is doing more work?

Have you checked the temperatures of the GPUs?
I have also observed that WUs will having varying power draw at different stages in the WU processing, so it will be hard to have a direct comparison just by the power draw.

My host 556495 shows a similar scenario for my two Alienware GTX 1650s.

Host 556495 data grabbed same time as your other host

By the way rod4x4, can a script be written to send Adria's PeanutButterBANDIT WUs on to somebody else?
Just kidding, but they really seem to put a dent in the RAC of the host that's running them.

I don't mind ADRIA tasks. There is always the possibility they could be used for future Scientific publications and go toward a badge.

You're right. It all comes out even in the wash, anyway.

I've realized just now that errors are hurting worse. My 750ti card might be dying, as I've underclocked it and it still produces "Error invoking kernel: CUDA_ERROR_LAUNCH_FAILED" errors.

You're right. It all comes out even in the wash, anyway.

I've realized just now that errors are hurting worse. My 750ti card might be dying, as I've underclocked it and it still produces "Error invoking kernel: CUDA_ERROR_LAUNCH_FAILED" errors.

Thanks Ian, the card is only 4 years old (pulled from a Dell and sold to me for $60 on eBay) and stays reasonably cool (55-60C) so I hope you found the problem for me. I'll sure try it.
Sorry for straying off topic here.

Results of a survey of the top 1500 hosts from the Volunteer Tab performed 7th December 23:20 UTC.

4 hosts have no Nvidia GPU (removed), hence not included in below statistics
243 hosts have Multiple GPUs and are not included in below statistics

Remaining 1253 hosts have all tasks (136,565 tasks), runtime and credit captured for the last 7 days (Gpugrid retain task history for 7 days for each host).

The Average Credit Per Day has been calculated from the Runtime Total and Credit Total.
NOTE: this is different from BOINC Recent Average Credit (RAC) which is computed as an exponentially weighted average with a half-life of one week. See here for further details on BOINC RAC calculation (point 4.1 Total and recent average credit): https://boinc.berkeley.edu/boinc_papers/credit/text.php

since it's well known that the windows app performs about 15-20% less than the Linux app, you should probably split these statistics by OS/app to differentiate from Windows/Linux.

I also see that you have a 3070 in that list, which shouldnt be there as the apps dont work for the Ampere cards yet and only produce errors. whatever host you picked that one up from likely had history data from a different model and recently swapped it out for a 3070 (causing BOINC to report it as such).
____________

I also see that you have a 3070 in that list, which shouldnt be there as the apps dont work for the Ampere cards yet and only produce errors. whatever host you picked that one up from likely had history data from a different model and recently swapped it out for a 3070 (causing BOINC to report it as such).

since it's well known that the windows app performs about 15-20% less than the Linux app, you should probably split these statistics by OS/app to differentiate from Windows/Linux.

From the data capture performed 7th December 23:20 UTC

1,275 ADRIA_PBBandit tasks were completed

maybe if the admins can get a new app with Ampere support released sometime soon, you can get some real data from the 30-series cards ;P. I've got a 3070 host waiting (with a 1660 Super secondary card running GPUG, and 3070 @Einstein)
____________

rod4x4:
It's awesome how, regarding low end GPUs, your study is nailing previously observed data at original GPUGrid's Performance tab.

Quoting my own comment from one year ago, at Low power GPUs performance comparative thread:

I've made a kind request for Performance tab to be rebuilt
At the end of this tab there was a graph named GPU performance ranking (based on long WU return time)
Currently this graph is blank.
When it worked, it showed a very useful GPUs classification according to their respective performances at processing GPUGrid tasks.
Just GT 1030 sometimes appeared at far right (less performance) in the graph, and other times appeared as a legend out of the graph.
GTX 750 Ti always appeared borderline at this graph, and GTX 750 did not.
I always considered it as a kind invitation for not to use "Out of graph" GPUs...

Can we retitle this thread to "Performance Tab Temporarily Located Here", maybe?

Excellent that you calculated the avg. per diem credit instead of plugging in BOINC RAC figures.

Bravo, rod4x4!

ServicEnginIC wrote:

Both GTX 750 Ti and GT 1030 are still shown respectively at penultimate and last position in your current general list, and GTX 750 is not.

I think that Titan Xp Collectors Edition was the card leading the classification (?)

Can we retitle this thread to "Performance Tab Temporarily Located Here", maybe?

Excellent that you calculated the avg. per diem credit instead of plugging in BOINC RAC figures.

Performance efficiency comparison.

The data shows the performance of the GPU Model when Average Credit / Day is compared to reference TDP Watts.

Please let me know if there is an error in the TDP Watts column. I have tried to get the correct reference TDP Watts for each GPU Model.

the 1650 is only a 75W TDP card. (but really only pulls 60-65W while running from my experience)
____________

the 1650 is only a 75W TDP card. (but really only pulls 60-65W while running from my experience)

Some random stats on successfully completed Experimental RAIMIS tasks.

- Data was captured 22nd December 2020, 9:00am UTC
- Top 1600 hosts (Including hosts with Multi GPU) from Volunteer/Hosts Tab were checked
- Processing time was lost during the period checked due to a Disk Full error at GPUGRID
- These Experimental tasks currently only run on Linux hosts.

NOTE:
All stats are just informational and have no true meaning as the tasks are still Experimental with numerous bugs. But could serve as a baseline to gauge improvements in future task performance.


Random Stat...1 - Top 10 longest tasks

Tasks awarded 20.83 credit - 147

Tasks awarded 20.83 credit - 147

as a reminder you will hit this "penalty" if your total credit reward is calculated to be greater than about 1,750,000 or so (I havent seen higher than this awarded).

ADRIA D3RBandit task runtime summary

The following table has been created from data gathered from the top 2000 hosts in the Volunteer tab.

NOTE
- Hosts with Multiple GPUs - these hosts have been included, but only tasks from the primary GPU (Device 0) have been used.
- Data was captured on 24th February 2021 at 2:30 UTC
- Task data history is only retained on the Gpugrid website for 7 days.