I've noticed the numerous GPU variants for nVidia devices and I'm curious why? Is it related to generation of hardware, CUDA driver support, both? Why would someone have more than one version of the app on hardware with identical GPUs?

Also related to this the AMD variants. I can guess that _cat132 is related to the broken OpenCL driver in the 13.2 Catalyst package but what about _ati Vs _ati5?

And while we are at it AstroPulse V6. What is the difference between ati_opencl_100 and opencl_ati_100 for AMD and cuda_opencl_100 and opencl_nvidia_100 fro nVidia?

Just being inquisitive.

---

"Life is just nature's way of keeping meat fresh." - The Doctor

There are three versions of CUDA, plus OpenCL100. These each have their own "family" of apps.
The reason we are seeing them all turning up is so the servers can work out the best version for your cruncher. Once this initial "calibration" has been done you should only see one type, barring the odd "recalibration" task that will be sent out just to make sure everything is running correctly.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

For cuda yes.
No for OpenCL.

---

With each crime and every kindness we birth our future.

I've noticed the numerous GPU variants for nVidia devices and I'm curious why? Is it related to generation of hardware, CUDA driver support, both? Why would someone have more than one version of the app on hardware with identical GPUs?

The reason there are Five different Cuda apps is because different GPUs favour different apps, as well as providing apps for hosts running older driver/Cuda versions.

Keplars tend to favour the Cuda5 app, Fermis the Cuda42 app, legacy GPUs are either going to favour the Cuda32 or Cuda23 app depending on driver version, (there is a slowdown on Cuda32 and Cuda42 apps with post 301.xx drivers on legacy GPUs)
and the Cuda22 app is for hosts that are for some reason driver limited, or are are particularly low on GPU memory,

Eric has decided to send the different apps out to almost all GPUs that support the respective Cuda version, and let the host do work and find out which app is fastest,
then the scheduler should only tend to send out work to the fastest app version.

Also related to this the AMD variants. I can guess that _cat132 is related to the broken OpenCL driver in the 13.2 Catalyst package but what about _ati Vs _ati5?

No, there are two separate plan classes for Cat 12.10 and earlier and Cat 13.2 and later, that is to stop the apps and Wu's being sent to Cat 12.11 Beta and Cat 13.1 because of their Buggy compiler.

The ati app uses fast local memory on the GPU and ati5 version doesn't.

And while we are at it AstroPulse V6. What is the difference between ati_opencl_100 and opencl_ati_100 for AMD and cuda_opencl_100 and opencl_nvidia_100 fro nVidia?

Boinc 6 clients can only tell if hosts have CUDA or CAL support, and not OpenCL support, so Eric did a special plan class to send out work for those clients using a few assumptions.

Boinc 7 can detect whether Nvidia and AMD/ATI GPUs have OpenCL support (as well as CUDA & CAL support), so they can get work from both sets of plan classes.

Claggy

Thank you, I know you are busy.

I was just curious because peaking around the results on the mega-crunchers I found it odd that cuda32 is running on a Titan card (running BOINC 7.0.64) or a GTX 580 is running cuda32, cuda42 and cuda50 (running BOINC 6.10.58).

---

"Life is just nature's way of keeping meat fresh." - The Doctor

Initially hosts will be running all permissible apps until the server has enough data to pereferentially send out the fastest one.

---

A person who won't read has no advantage over one who can't read. (Mark Twain)

Yes I got that from Claggy's post.

---

"Life is just nature's way of keeping meat fresh." - The Doctor

............
No, there are two separate plan classes for Cat 12.10 and earlier and Cat 13.2 and later, that is to stop the apps and Wu's being sent to Cat 12.11 Beta and Cat 13.1 because of their Buggy compiler.

The ati app uses fast local memory on the GPU and ati5 version doesn't.
......

Claggy, I just thought that ati5_app uses GPU memory and ati_app (without 5) does it not. So I drive the wrong app.

............
No, there are two separate plan classes for Cat 12.10 and earlier and Cat 13.2 and later, that is to stop the apps and Wu's being sent to Cat 12.11 Beta and Cat 13.1 because of their Buggy compiler.

The ati app uses fast local memory on the GPU and ati5 version doesn't.
......

Claggy, I just thought that ati5_app uses GPU memory and ati_app (without 5) does it not. So I drive the wrong app.

No, the HD5 app, with HD5 in the filename (or mentioned in the stderr.txt) uses local GPU memory, Eric used a different api to determine the target device, and gave the HD5 app the ati planclass, and the non_HD5 app the ati5 planclass.

Claggy

............
No, there are two separate plan classes for Cat 12.10 and earlier and Cat 13.2 and later, that is to stop the apps and Wu's being sent to Cat 12.11 Beta and Cat 13.1 because of their Buggy compiler.

The ati app uses fast local memory on the GPU and ati5 version doesn't.
......

Claggy, I just thought that ati5_app uses GPU memory and ati_app (without 5) does it not. So I drive the wrong app.

No, the HD5 app, with HD5 in the filename (or mentioned in the stderr.txt) uses local GPU memory, Eric used a different api to determine the target device, and gave the HD5 app the ati planclass, and the non_HD5 app the ati5 planclass.

Claggy

Thank you for the info, Claggy.