Einstein - Improved apps for nvidia cards

[pututu]

RFGuy crunched Einstein recently and notice almost double the output with the new nvidia apps.

The main thread is here: https://einsteinathome.org/content/all-things-nvidia

The new apps for the popular gamma ray pulsar are listed here: https://einsteinathome.org/apps.php

I think it is auto-detect but I haven't run it yet.

The new consumer gpu king for Einstein seems to be 3080 Ti belonging to Ian&Steve: https://einsteinathome.org/host/12830750/tasks/0/0. Current output is about 2.7M PPD.

[biodoc]

The main thread is here: https://einsteinathome.org/content/all-things-nvidia

According this thread, the new apps require OpenCL 3.0.

I had the nvidia driver 460 installed (OpenCL 2.1) on 2 of my rigs (1080Ti on one and 2070S on another) so I ran a few tasks using the old einstein app. Then I installed the 470 nvidia driver (OpenCL 3.0) on both rigs and ran a few tasks using the new einstein app (downloaded and used automatically).

EDIT: All the tests are on Linux. I was also running TN-Grid on the CPUs but one full core was reserved for GPU support during testing.
"Old app": Gamma-ray pulsar binary search #1 on GPUs v1.20 () x86_64-pc-linux-gnu
"New app": Gamma-ray pulsar binary search #1 on GPUs v1.28 () x86_64-pc-linux-gnu

RTX 2070S:
task completion time using the old einstein app: ~473 seconds
task completion time using the new einstein app: ~290 seconds

GTX 1080Ti:
task completion time using the old einstein app: ~390 seconds
task completion time using the new einstein app: ~272 seconds

[pututu]

Awesome. Looks like my Radeon VII is now only good for MW project (other than ETH mining). Not very clear about the crunching/power efficiency, maybe a slight edge in Radeon VII versus 3080 Ti which consumes humongous power and rated at 350W.

[biodoc]

Apparently AMD doesn't have a driver that supports OpenCL 3.0 yet. The AMD driver situation on linux is a mess. The new drivers only support rcr (rocm) and legacy installation options. They dropped support for PAL which is required for MilkyWay app support on the Radeon VII. I rebooted my Radeon VII rig into Windows 10 and downloaded the latest driver and ran a few einstein tasks. @200 watts, the einstein apps completed in 170 seconds.

[Fardringle]

Is the OpenCL version purely dependent on the driver, or does it also depend on the card?

For instance, I'm trying to figure out if my GTX 1060 and Quadro P4000 cards support OpenCL 3.0 or not. They both have the appropriate GTX and Quadro version 472.12 drivers installed.

[biodoc]

For instance, I'm trying to figure out if my GTX 1060 and Quadro P4000 cards support OpenCL 3.0 or not. They both have the appropriate GTX and Quadro version 472.12 drivers installed.

According to the techpowerup database, both your cards support OpenCL 3.0. https://www.techpowerup.com/gpu-specs/g ... 6-gb.c2862

Check your coproc_info.xml file in the boinc directory to confirm. In linux you can run clinfo to check.

On windows you can run GPU-Z and in the "advanced" tab choose "OpenCL" in the drop-down menu to get the details.

[Fardringle]

Thank you for the suggestion. I had looked in CPU-Z but didn't realize there was a drop-down list on the Advanced tab..

It says "OpenCL 3.0 CUDA 11.4.136" for the Platform Version, and "OpenCL C 1.2" for the C Version. Interestingly enough, GPU-Z says the same thing for my really old Quadro K2200 cards.

The BOINC coproc_info.xml file also shows OpenCL 3.0 on all of the computers so I guess that means it should work even for the really old cards...

I'll switch them all over to Einstein as soon as I get to 100M in SRBase, to see how they do with the new application.

[biodoc]

The BOINC coproc_info.xml file also shows OpenCL 3.0 on all of the computers so I guess that means it should work even for the really old cards...

According to the thread above, the server only distributes the new app to computers with Nvidia compute capability 6.1 or higher. It might be worth testing to be sure.

[Fardringle]

Google says the K2200 only has compute capability 5.0 so I guess it probably won't work for those cards. I'll try it and see what happens.

[StefanR5R]

I had the nvidia driver 460 installed (OpenCL 2.1) on 2 of my rigs (1080Ti on one and 2070S on another) so I ran a few tasks using the old einstein app. Then I installed the 470 nvidia driver (OpenCL 3.0) on both rigs and ran a few tasks using the new einstein app (downloaded and used automatically). [...]

Today it is a bit cooler in my kitchen than usual. There happens to sit a dual 1080Ti computer in a corner of the kitchen. So I decided to fire up Einstein@home on it, first with the currently installed 460 driver.

I see that I had the boinc client configured to run two Einstein GPU tasks at once per card. With the GRPBS#1 work which it downloaded just now, this results in 100 % GPU utilization but still merely ~180 W power usage (while the power limit of these cards is 250 W). It's possible that I formerly took care that the two tasks per card are started at different points in time, to bridge parts of the tasks which use CPU but not GPU — but I am not sure if that's really the case why I have this 2/GPU setting. I need to write down notes of such things more often.

TBD: Update to the 470 driver and see how the newer application version fares with 1/GPU and 2/GPU.

PS, 470 driver, 1/GPU: SM utilization and power usage fluctuate between ~95…99 % and ~160…230 W, with 99 % and ~210 W being typical figures. My app_config.xml for the former application does not apply to the new version, it will need to be updated if I want to try 2/GPU.

PPS, 470 driver, 2/GPU: SM utilization is pegged again at 100 %, power usage fluctuates in narrower band of ~200…220 W now, again with ~210 W being the typical consumption.

Added in 50 minutes 50 seconds:
My conclusion:
Based on the 2-simultaneous-tasks/GPU setting, the update from hsgamma_FGRPB1G version 1.20 to version 1.28 brings

• ~1.5 times the throughput

at merely

• ~1.15…1.2 times the board power consumption

on this 1080Ti equipped computer. The efficiency increase is even higher if you consider that the rest of the system power consumption (CPU power and so on) stays basically the same.

[Fardringle]

I didn't want to uninstall the current drivers, install old drivers, test, and then reinstall the new drivers again, so I've only had the K2200s run a few tasks with the new drivers. I'm not 100% certain, but it does look like the old cards are not getting new new optimized applications since they are running the "Gamma-ray pulsar binary search #1 on GPUs 1.22 (FGRP0pencl-nvidia)" while I believe the new application is version 1.28...

[StefanR5R]

it does look like the old cards are not getting new new optimized applications since they are running the "Gamma-ray pulsar binary search #1 on GPUs 1.22 (FGRP0pencl-nvidia)" while I believe the new application is version 1.28...

Yep, 1.22 (FGRPopencl-nvidia) was installed in June 2019, but 1.28 (FGRPopencl2Pup-nvidia) in August 2021. (apps.php)

[biodoc]

I installed my new 3080Ti yesterday (350 watt TDP) and am running Einstein on the "new" opencl 3.0 app. The OS is Mint 20.3 and TnGrid is running on 30 threads (5950X). The power limit is set using 'sudo nvidia-smi -pl 275' as an example for setting the power limit at 275 watts.

Default power limit (350 watts). According to nvidia-smi, the card is using 324 watts.
task completion time: 120 seconds
275 watt power limit:
task completion time: 124 seconds
250 watt power limit:
task completion time: 126 seconds
225 watt power limit:
task completion time: 134 seconds

I also have a 3070 Ti running Einstein with 30 threads on TNGrid (3950X)
225 watt power limit
task completion time: 165 seconds

[crashtech]

Eyeball math makes the 250 watt setting look good.