PrimeGrid Challenges 2023

[mmonnin03]

I have been getting 1.2m PPD with a 3080Ti in Windows at 80% power running CW sieve.

 

[StefanR5R]

You can take these relative scores as an orientation:
www.primegrid.com -> Fastest GPUs -> CW-Sieve

These are not totally exact, but quite well in the ballpark.

Edit:
These tables don't have entries for FP64-heavy Nvidia GPUs right now. But going by the AMD GPU table, FP64 hardware support is evidently not helping in CW-Sieve.

 

[Markfw]

Ok, after some test runs, 560 seconds on a 3080TI, and 780 seconds on a Titan V, but the Titan uses 1/2 the power.

C:\Users\Mark>nvidia-smi
Thu Jul 6 11:18:22 2023
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 531.61 Driver Version: 531.61 CUDA Version: 12.1 |
|-----------------------------------------+----------------------+----------------------+
| GPU Name TCC/WDDM | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+======================+======================|
| 0 NVIDIA GeForce RTX 3080 Ti WDDM | 00000000:01:00.0 On | N/A |
| 81% 83C P2 306W / 350W| 1091MiB / 12288MiB | 100% Default |
| | | N/A |
+-----------------------------------------+----------------------+----------------------+
| 1 NVIDIA TITAN V WDDM | 00000000:08:00.0 Off | N/A |
| 95% 89C P2 145W / 250W| 356MiB / 12288MiB | 100% Default |
| | | N/A |
+-----------------------------------------+----------------------+----------------------+

 

[StefanR5R]

That's indeed superior perf/W on the older Volta vs. the newer Ampere...

Also interesting that you are getting 100% SM (streaming multiprocessors) utilization but only 87 % power utilization (Ampere) and 58 % power utilization (Volta), on Windows. On Linux, I have seen 100% SM utilization at 100% power utilization, both with Pascal cards and with a Lovelace card (a.k.a. Linda) in CW-Sieve.

Can you dual-boot this computer into Linux to compare task times and power draw on the same cards, Windows versus Linux?

 

[Markfw]

That's indeed superior perf/W on the older Volta vs. the newer Ampere...

Also interesting that you are getting 100% SM (streaming multiprocessors) utilization but only 87 % power utilization (Ampere) and 58 % power utilization (Volta), on Windows. On Linux, I have seen 100% SM utilization at 100% power utilization, both with Pascal cards and with a Lovelace card (a.k.a. Linda) in CW-Sieve.

Can you dual-boot this computer into Linux to compare task times and power draw on the same cards, Windows versus Linux?

This computer is the one I am typing on, and runs BOINCTASKS and is my central hub for controlling/watching my fleet. It never gets rebooted for any reason other than mandatory updates.

 

[StefanR5R]

Nevermind. I found some computers on the PrimeGrid web site according to which Windows and Linux performance in CW-Sieve appears the same. It's not certain though if we merely look at same hardware (of different users, unknown background load, unknown TDP setting), rather than one and the same hardware.

I suspect the reason why my cards used 100% of their power budget while yours did not is this one:
Pascal: ≤55 °C, Linda: ≤70 °C — vs. — Volta: 89 °C, Ampere: 83 °C

 

[StefanR5R]

GPU	task duration	throughput	board power	power efficiency
RTX 3080 Ti	560 s​	1,200 kPPD​	306 W​	3.9 kPPD/W​
Titan V	780 s​	850 kPPD​	145 W​	5.8 kPPD/W​
GTX 1080 Ti	1,550 s​	430 kPPD​	250 W​	1.7 kPPD/W​
RTX 4090	210 s​	3,100 kPPD​	450 W​	7.0 kPPD/W​

 

[mmonnin03]

I run 2x tasks, mainly for WUProp hours so per task time is half of my BOINCTasks estimated run time of 17:19. Or about 520s per task.
My 3080Ti is the only one in the system and is running around 71C at 75% fan speed. 328 max wattage its all being used at 100% utilization. So yeah, his high temps.

 

[StefanR5R]

GPU	task duration	throughput	board power	power efficiency
mmonnin's 3080 Ti	2× 520 s /2​	1,300 kPPD​	328 W​	3.9 kPPD/W​

 

[biodoc]

4090@275 watt power limit: 247 seconds.

 

[StefanR5R]

CW-Sieve

GPU	task duration	throughput	board power	board power efficiency	operating limit
RTX 4090	210 s​	3,100 kPPD​	450 W​	7.0 kPPD/W​	power
RTX 4090	247 s​	2,700 kPPD​	275 W​	9.7 kPPD/W​	power
RTX 3080 Ti	520 s​	1,300 kPPD​	328 W​	3.9 kPPD/W​	power
RTX 3080 Ti	560 s​	1,200 kPPD​	306 W​	3.9 kPPD/W​	temperature?
Titan V	780 s​	850 kPPD​	145 W​	5.8 kPPD/W​	temperature?
GTX 1080 Ti	1,550 s​	430 kPPD​	250 W​	1.7 kPPD/W​	power

 

[TennesseeTony]

RTX 2080 Super, reaches "Utilization Limit", 766 s , 250W TDP -- 863kPPD , 3.45 kPPD/W

RTX 2080 Ti , 250W TDP, 635s , Power Limited -- 1041kPPD , 4.16 kPPD/W

RTX 2080 , 821s , forgot to check limitations, 215W TDP -- 805 kPPD , 3.74kPPD/W

 

[StefanR5R]

Challenge scores are currently 1/7th of the actual credits. That's presumably because CW-Sieve gave 1/7th of the credits per result years back before it went into hiatus (so I have read elsewhere). Let's see if they leave the challenge this way, or switch it up to modern credits.

 

[mmonnin03]

It must have changed? I'm at 267k almost 4 hours in.

 

[StefanR5R]

Yes, they switched to current credit after 22:45 UTC/ before 23:00 UTC = two hours after the start. :-)

 

[Ken g6]

Day 1 stats:

Rank___Credits____Username
4______21482685___[TA]Skillz
6______20013784___w a h
7______17037728___crashtech
8______16180869___Pokey
9______16173218___xii5ku
27_____4161887____biodoc
46_____2509373____Orange Kid
84_____1224084____mmonnin
99_____1048122___10esseeTony
106____872160_____johnnevermind
108____856859_____Skivelitis2
132____596741_____waffleironhead
141____543187_____Fardringle
217____130059_____Letin Noxe
223____122408_____Ken_g6

Rank__Credits____Team
1_____102930219___TeAm AnandTech
2_____77706424___Antarctic Crunchers
3_____44128252___SETI.Germany
4_____42536942___BOINC@AUSTRALIA

I'm so slow at this. Did everybody upgrade from a 1060 but me?

 

[StefanR5R]

I partially upgraded from 1080Ti, obviously. Initially I wanted to use only the upgrades in this challenge, due to the Summer heat. But then I was afraid to drop out of the arithmetic progression in which I found myself early on, so I entered my remaining Pascals as well.

Yesterday I tortured the GPUs with up to 35°C intake air temperature. Today temperatures should be a little lower, but not by much.

 

[StefanR5R]

I wonder how much of it will be left after the challenge.

Subproject status page: https://www.primegrid.com/stats_cw_sieve.php
Right now (1.6 days into the challenge) it lists 102,070 total "tasks", 88,267 completed¹ "tasks" (i.e. 13,803 "tasks" to be done).

At the same time, the sum of the rightmost column at http://www.primegrid.com/challenge/2023_5/top_teams.html is 117,718 results (valid and pending).

This doesn't add up. I suppose the status page lists workunits, not tasks. (The status page also claims 3,825 "tasks" queued for download, while the front page shows 7,186 available tasks. Edit, and https://www.primegrid.com/server_status_subprojects.php shows 102,070 workunits/ 214,961 tasks in the database. Hence, the status page definitely lists workunits, not tasks.) Still, if there are really only 13,803 workunits left to be done, that won't be a lot relative to the progress rate of the challenge.

________
¹) Give or take. The row of the 3860T - 3870T range claims "-59" completed "tasks", which makes it all even weirder.

 

[StefanR5R]

Subproject status page: https://www.primegrid.com/stats_cw_sieve.php
Right now (1.6 days into the challenge) it lists 102,070 total "tasks",

And now it shows 112,204 total (workunits), because…

New rows are added to the table as the leading edge advances. IIRC CW was originally sieved to 3500T for the original <25M range, so we are passed that point with the sieve currently. I'm not sure where the optimal depth for 25M-50M is determined to be, but I imagine the challenge will be getting a lot of progress towards the optimal depth.

 

[Ken g6]

Day 2 stats:

Rank___Credits____Username
4______46683528___[TA]Skillz
5______33180345___xii5ku
6______32759565___crashtech
7______32193426___Pokey
8______31979211___w a h
24_____10259359___biodoc
49_____5034048____Orange Kid
78_____2960754___10esseeTony
92_____2494072____mmonnin
101____2241605____johnnevermind
117____1667815____Skivelitis2
144____1239385____waffleironhead
154____1147579____Fardringle
201____543187_____kiska
250____175962_____Ken_g6
264____130059_____Letin Noxe

Rank__Credits____Team
1_____204666956___TeAm AnandTech
2_____148259604___Antarctic Crunchers
3_____96932204___SETI.Germany
4_____95126679___BOINC@AUSTRALIA

I've limited my contribution, because it's not doing much anyway, and it makes my computer freeze up frequently for up to a minute or two!

 

[StefanR5R]

The thumb up is for the stats, not for the freezes… I was about to say that my computers aren't doing that, but then, I am almost only interacting with them through ssh and remote boincmgr. So if there were freezes, they might easily escape my attention.

 

[Skillz]

I'm running it on my 3090 primary rig. Not lock-up/freezes on my host at least.

 

[mmonnin03]

I've been running it fine as well on my main PC.

 

[Ken g6]

Day 3 stats:

Rank___Credits____Username
4______66559603___[TA]Skillz
5______50202772___xii5ku
6______49896751___w a h
7______48634413___crashtech
8______48236586___Pokey
26_____16479239___biodoc
49_____7596975____Orange Kid
66_____5393623____cellarnoise2
74_____4888688___10esseeTony
94_____3764060____mmonnin
99_____3603399____johnnevermind
149____1759621____waffleironhead
154____1667815____Skivelitis2
155____1660164____Fardringle
181____1231735____kiska
261____229515_____Ken_g6
278____130059_____Letin Noxe

Rank__Credits____Team
1_____311904423___TeAm AnandTech
2_____218965795___Antarctic Crunchers
3_____165725762___SETI.Germany
4_____146140408___BOINC@AUSTRALIA

Turns out there's a bug in my version of BOINC so it never detects when I'm idle. I could not find a way to download a later Linux binary without the installer either. But I wrote a script workaround, so I'm finally doing work again.

 

[Skillz]

I think the newest version of BOINC for Linux is on a GitHub page.