Rx570 works fine here.

Hello All,

I have an AMD Radeon HD 6670 card that I am attempting to use on MW, but I can't seem to get any work units.
Is anyone else having luck using this old card?
Thanks,
Allen

Hello All,

I have an AMD Radeon HD 6670 card that I am attempting to use on MW, but I can't seem to get any work units.
Is anyone else having luck using this old card?
Thanks,
Allen

The HD 6670 does not have double precision required by MW.

Joe

Milkyway@Home: Notice from server
ATI GPU R600 (R38xx) does not support OpenCL

I'm quite certain that it does though. Any idea what's going on?

Usually it means the OpenCL driver isn't loaded but in your case it is, at least it is on the Darwin based pc, the Windows based pc has no drivers loaded for the gpu at all.

Try running the TF tasks at the project SRBase and see if it works there, they don't require OpenCL but do require drivers so if you are trying to use the Windows pc you will need to loads drivers for that pc. I don't think MilkyWay upped the minimum requirement for the version of OpenCL here to 1.2 but if so that could be your problem too, I just checked my pc's and all mine are running version 1.2 or better.

Thanks, just my luck.

Have been running MW on an Asus mobo, Intel Core i7-3930K Hexa-Core CPU, Windows 10 Enterprise x64, NVIDIA GTX 1070 Ti, and its Cuda cores were crunching numbers very well. Last week I built a new system especially for targeting this kind of task and others like it. Moved my NVIDIA GPU to the new system, installed MW but now, no Cuda number crunching. Here are the differences between the two configurations.

The new system is ASRock 570 Creator mobo, AMD Ryzen 9 3900xt CPU, Windows 10 Enterprise x64, NVIDIA GTX 1070 Ti, and this time when I installed BOINC I did so without installing the Oracle VM bundle.

Is leaving the Oracle VM out, the cause for no Cuda processing? Or could there be another reason?
Any guidance would be helpful from someone who's seen this kind of issue before.

Have been running MW on an Asus mobo, Intel Core i7-3930K Hexa-Core CPU, Windows 10 Enterprise x64, NVIDIA GTX 1070 Ti, and its Cuda cores were crunching numbers very well. Last week I built a new system especially for targeting this kind of task and others like it. Moved my NVIDIA GPU to the new system, installed MW but now, no Cuda number crunching. Here are the differences between the two configurations.

The new system is ASRock 570 Creator mobo, AMD Ryzen 9 3900xt CPU, Windows 10 Enterprise x64, NVIDIA GTX 1070 Ti, and this time when I installed BOINC I did so without installing the Oracle VM bundle.

Is leaving the Oracle VM out, the cause for no Cuda processing? Or could there be another reason?
Any guidance would be helpful from someone who's seen this kind of issue before.

In case the question is asked, yes, the OpenCL drivers are installed.

Have been running MW on an Asus mobo, Intel Core i7-3930K Hexa-Core CPU, Windows 10 Enterprise x64, NVIDIA GTX 1070 Ti, and its Cuda cores were crunching numbers very well. Last week I built a new system especially for targeting this kind of task and others like it. Moved my NVIDIA GPU to the new system, installed MW but now, no Cuda number crunching. Here are the differences between the two configurations.

The new system is ASRock 570 Creator mobo, AMD Ryzen 9 3900xt CPU, Windows 10 Enterprise x64, NVIDIA GTX 1070 Ti, and this time when I installed BOINC I did so without installing the Oracle VM bundle.

Is leaving the Oracle VM out, the cause for no Cuda processing? Or could there be another reason?
Any guidance would be helpful from someone who's seen this kind of issue before.

In case the question is asked, yes, the OpenCL drivers are installed.

I've solved my problem but not sure why it works this way, maybe someone with greater insight can enlighten me. On the last installation of BONIC, I installed it as a service in addition to installing it without Oracle Virtual Box. After reinstalling as a desktop app, not a service, the GPU Cuda cores are now crunching numbers again. I've done some queries on this site in an effort to find something that explains this but found nothing that speaks to this directly.

I've solved my problem but not sure why it works this way, maybe someone with greater insight can enlighten me. On the last installation of BONIC, I installed it as a service in addition to installing it without Oracle Virtual Box. After reinstalling as a desktop app, not a service, the GPU Cuda cores are now crunching numbers again. I've done some queries on this site in an effort to find something that explains this but found nothing that speaks to this directly.

Long time since I ran BOINC on Windows, but I seem to recall that if you run it as a service it can't see the GPU(s). Indeed, that's confirmed at https://boinc.berkeley.edu/wiki/GPU_computing where it says

Warning: On Windows do not install BOINC in Protected Access Execution (PAE) mode aka service mode (6.4.5 - 7.0.28) or Service Install mode (7.0.64 and above). If you do, BOINC will not be able to detect or use your GPU.

(working on the assumption that that is still correct...)

Hope that helps - Al.

I've solved my problem but not sure why it works this way, maybe someone with greater insight can enlighten me. On the last installation of BONIC, I installed it as a service in addition to installing it without Oracle Virtual Box. After reinstalling as a desktop app, not a service, the GPU Cuda cores are now crunching numbers again. I've done some queries on this site in an effort to find something that explains this but found nothing that speaks to this directly.

Long time since I ran BOINC on Windows, but I seem to recall that if you run it as a service it can't see the GPU(s). Indeed, that's confirmed at https://boinc.berkeley.edu/wiki/GPU_computing where it says

Warning: On Windows do not install BOINC in Protected Access Execution (PAE) mode aka service mode (6.4.5 - 7.0.28) or Service Install mode (7.0.64 and above). If you do, BOINC will not be able to detect or use your GPU.

(working on the assumption that that is still correct...)

Hope that helps - Al.

You beat me to it, YES that's true AND it's also true that if you use multiple log ins for your pc the gpu will stop working as soon as the first person is not the person running the desktop. For example a lab where multiple people use the same pc...the gpu will run for the original person each day but as soon as some one else logs in the gpu stops and won't run again until the original person logs back in again.

Thanks for the info. I finally realized I missed the clear and direct answer to this question when installing the app. The Advanced dialog clearly states that GPU crunching won't work if the app is installed as a service. My bad! Just goes to show how getting in a hurry, ends up wasting more time than reading the directions beforehand.

So hear it is 2021 and soon to be 2022, so I just thought I would ask if we are ever going to see MW be able to run a WU with the Intel UHD series graphics processor. Unless there is a reason why it's not possible, I'm rather surprised that some gifted programmer hasn't written a program to use it yet.

Thanks for any remarks.
Allen

First I'm not on linux but OS X ;-)
Second, I don't have the hand on the drivers!! I use those that OS X installs!
It's ok for the 5770, I don't try to compute with it.

If you need more détails on my config, ask me.

Thx for your response!

P.S: The problem was same with my old HD4870!!

Can somebody tell me the name the Modified Fit app uses in client_state.xml, I am going to attempt making an app_info file that hopefully works for OSX and allows GPU crunching.

Modified Fit ? ? ? ? We haven't had that work for years. Only work is for Separation and N-Body. Separation is the only gpu application. Its app_name is milkyway in client_state.
https://milkyway.cs.rpi.edu/milkyway/apps.php
Current work is Separation and the app version is 1.46. You can pick your poison for the application and its name here.
https://milkyway.cs.rpi.edu/download/

---

There seems to be so much miss-information about GPU's and the ability to number crunch.
I have 5 computers running Darwin and another W10 i7 working on MilkyWay@Home. My situation with regard to Video cards is fortunate as I have several boxes of them all what gamers call 'good to have' but few are suitable for FP64 and MilkyWay@Home.
My computer configuration is visible and the changes I make to improve performance can be seen.
Interesting results so far with more changes coming.
The GTX Titan is good in W10 (Drivers installed) and Darwin as long as the Darwin 15.6.0 is used and drivers installed as the drivers will not install in 18.7.0. The difference in compute speed with both OS's without drivers is -2.5x.
With Darwin Mojave you need a Metal compatible card and the fastest doing a WU is the HD 7950 so far, some 75% faster than the Titan and a GTX 680 with drivers installed.
Lots of older AMD cards tried and they just seem to be on steroids, the later ones just don't do FP64 as well. NVIDIA seem to be behind the 8 ball in every GPU swap I have done.
Some of the cards I have played with:
HD 7950 FP64 GF 716.6 1:4
HD 5870 544 1:5
GTX 780 Ti 222.7 1:24
GTX 480 168 1:8
GTX 680 135.4 1:24
GTX 970 122.5 1:34
GTX 570 Ti 105.3 1:12
R9 270 168.1 1:8
GT 640 28.86 1:24
GTX Titan 1570 1:3
HD 7870 748.8 1:4

Nowhere do I see information on consumer math capable cards but look and you will find cards 7-10 years old that suit this project by a wide margin unless your a gamer.
More card swaps coming and btw, in MacPro 5,1 using Triple Channel memory makes no difference to compute times so I went back to all slots filled with ram.
Found that changing the CPU usage to 60% keeps the thermal stress down with not much difference in outcome.
Still playing, any thought or advice?

The TechPowerUP GPU database is the best resource for finding the FP64 capability of every known card.
https://www.techpowerup.com/gpu-specs/
If you search for a card or click on a card, the Theoretical Performance panel shows the FP16, FP32 and FP64 ratings.

---

Yes it's a good site but the specs give no indication of how well they work in various OSX versions. Like drivers for GTX Titan work in El Capitan but not in Mojave. When it is working it gets through a WU at about the same speed as OSX and W10 with FP64 enabled but slower than the HD 7950 by quite a margin. There will be a work around, I have the programming skills but not the motivation to bother.

Another thing with Mojave is that it requires a Metal capable card and the listing on sites leave out so many cards that will work. So it's a matter of finding a fast FP64 Kepler GPU that works with Mojave and MW.
A lot of Nvidia cards won't work at full potential as Apple dropped the support in Majove, that's why I run the MacPro 3,1 as well the 5,1 because good cards are supported.

As for AMD lots of cards work straight out of the box, far more than the listings suggest. Again Majove requires Metal capable cards, most of the HD 7xxx work in the 5,1 and the HD 58xx-59xx work well on 3,1.

So it's been interesting to look at the performance when swapping out cards from W10 to OSX.

I can also confirm that some Nvida cards work together on MW in both x16 slots although the power budget has to be watched unless the power leads are tapped before the main board. MacPro cheese grater has close to 1kw available from the power supply to run the computer.

Gaming GPU's are not value for money running MW in my opinion as some of the $40 old cards are quite fast, what's outdated for games is great for MW.

The Top GPU list For OSX doesn't add up to me, TechPowerUP list some of the FP64 performance as quite poor.

I've never had any inclination to touch Macintosh, not even with a 10 foot pole. From what I have heard from fellow crunchers and posts such as yours, the hardware/software lockdown by Apple is stupidly restrictive.
You must have an appetite for frustration. Good luck in your search.

---

For those that would like to use the older Mac's 3,1-5,1 these are the cards I found give reasonable FP64 GPU performance, averaged time over 5 WU's.

The 3,1 running El Capitan, Darwin 15.6.0: HD 5870 105 sec, GTX Titan 91 sec.

The 5,1 running Majove, Darwin 18.6.0: HD 7870XT, 89.09 sec, HD 7950 62.1 sec.

I will continue to test and post as further cards become available.

The TechPowerUP GPU database is the best resource for finding the FP64 capability of every known card.
https://www.techpowerup.com/gpu-specs/
If you search for a card or click on a card, the Theoretical Performance panel shows the FP16, FP32 and FP64 ratings.

This list shows double precision in a column where one can easily pick out the leaders. Unfortunately, it cannot be sorted and some of the much older cards way down in the list are better at Milkyway than much newer cards
For example HD 5850 at 417 FP64 (2009 released) is better than GTX 1080TI and ranks up there with even the newest RTX 3xxx series.

https://www.geeks3d.com/20140305/amd-radeon-and-nvidia-geforce-fp32-fp64-gflops-table-computing/

Only one of the professional AMD cards is shown.

Not listed are the S9000 and S9050 which use the same CHIP as the HD7950. Those S cards used to go in the $100 range but miners have driven all the prices back up to the MSRP or higher.

For those that would like to use the older Mac's 3,1-5,1 these are the cards I found give reasonable FP64 GPU performance, averaged time over 5 WU's.

The 3,1 running El Capitan, Darwin 15.6.0: HD 5870 105 sec, GTX Titan 91 sec.

The 5,1 running Majove, Darwin 18.6.0: HD 7870XT, 89.09 sec, HD 7950 62.1 sec.

I will continue to test and post as further cards become available.

I have an old Mac desktop, a 3.1, can it take a standard gpu like a 5870 that also works in a Windows pc? Someone once said the cards had to be Mac something or other to work in a Mac