I am hoping I can get each of these cards to crunch two work units at a time without a drastic performance decrease.

The Seti documentation seems to think you can. I am not sure how fussy it will be about having enough cpu cores for each task though.

Tom

---

A proud member of the OFA (Old Farts Association).

Before you spend all that money on GPUs that can't run the CUDA Special App, you might want to consider running Linux with the CUDA Special App. I currently have 9 nVidia GPUs running on a Gigabyte FinTech board without using any special settings. The board will handle up to 12 GPUs, without using any 1x4 switches. The 1x4 switch will work if you wanted to add more.

Try comparing the Run-Times the Special App produces against anything the OpenCL Apps produces. The Special App is anywhere from 2 to 3 times faster on the same hardware.
Coprocessors: [9] NVIDIA GeForce GTX 1060 3GB (3011MB) driver: 410.57 OpenCL: 1.2

Well, I wasn't talking about that at all to him. I figured once he got that pallet load of Gtx 1050's that he ordered from overseas in he could "try" the upgrade out. After he canceled the long-delayed shipment he didn't mention thinking about this until he did.
I think he has already ordered the Radeon GPUs.

I agree with your principle but right now he is a Windows person. Maybe we can get him to try out the Lunatix Amd speedups after a while?

He also has a dual e5-2670 that he could drop a couple of gtx 1060 3 GB's into and do the Linux/Cuda90+ upgrade. He is running it with a very low-end Quadro card just for a placeholder right now (I think).

Tom

---

A proud member of the OFA (Old Farts Association).

I currently have 9 nVidia GPUs running on a Gigabyte FinTech board without using any special settings. The board will handle up to 12 GPUs...

Those boards have almost all PCIe x1 tiny slots for dedicated miner cards. Are you using x1 miner 1060s or risers/adapters to fit x16 full-size cards to them?

---

I am using the risers as required by the ASUS B250 but I have since learned that miner boards don't pose any significant bottleneck for Seti as the X1 lane is only used to send and get data from the GPU's. Once the GPU has the job, it crunches it until done and then sends it back for another. I get the sense the GPU and core interaction is minimal once the GPU has the work unit. What is not clear to me is whether the GPU caches the crunching algorithm or does that stay resident in the mobo RAM and core.

I must have missed class the day CUDA and the special Linux app was revealed so I dumped the order of Radion cards and have a pack of eight GTX 1060's coming. (3GB)

I am using the risers as required by the ASUS B250 but I have since learned that miner boards don't pose any significant bottleneck for Seti as the X1 lane is only used to send and get data from the GPU's. Once the GPU has the job, it crunches it until done and then sends it back for another. I get the sense the GPU and core interaction is minimal once the GPU has the work unit. What is not clear to me is whether the GPU caches the crunching algorithm or does that stay resident in the mobo RAM and core.

I must have missed class the day CUDA and the special Linux app was revealed so I dumped the order of Radion cards and have a pack of eight GTX 1060's coming. (3GB)

No, you didn't miss class. We never talked about it in this thread. If you didn't have the time or were not reading in the Messages -> Crunching area you would not have run into any of the CUDa90/Linux discussion.

I am beginning to believe that only the Windows SOG task for the gpu requires a lot of bandwidth between the CPU and the gpu. None of the others do.

Tom

---

A proud member of the OFA (Old Farts Association).

That is also my impression Tom.
I am not afraid of Linux and have fires it up a few times.

We shall see.

By the way, I have 3 of the Radion cards coming as well so i will be able to test both.

By the way, I have 3 of the Radion cards coming as well so i will be able to test both.

1+

---

A proud member of the OFA (Old Farts Association).

I'm excited. The 1060's are waiting on me at home!

Hey Tom, does seti know to download the CUDA stuff for windows when it see's the new GPU's? I take it that's a manual operation in Linux?

Hey Tom, does seti know to download the CUDA stuff for windows when it see's the new GPU's? I take it that's a manual operation in Linux?

The CUDA90+ stuff is a manual setup.

You can set up the "standard" Linux Seti setup and then add the CUDA90+ stuff if you want or you can try my simple-minded setup at: https://setiathome.berkeley.edu/forum_thread.php?id=83274 I had a typo in the original thread starting message so I asked the moderator to hide it and re-posted with the typo fixed.

My setup will get you up and running but it doesn't have any "autostart" features for the BoincManager or any of the other really useful stuff that are talked about in the "From Windows to CUDA90+" thread in the messages -> crunching area.

Tom

---

A proud member of the OFA (Old Farts Association).

So, just to confirm, there is no CUDA for the windows platform?

I got 8 of the 9 1060's up and running!

1/2 the power

1/2 the heat

Almost twice as fast per work unit?

Cant tell if # 9 was a bad connection or a bad GPU. I'll tinker that out in the coming days.

If the CUDA on Linux is really twice as fast as Windows, then I need another 10 1060's to play with ASAP!

You guys (and gals) must think I'm nuts but I have an ulterior motive besides Seti.

I crunch electromagnetic sims for dollars and my simulation PKG uses GPU's!
If what I find on the net is right, 8 1060's give 30.4 Teraflops and the i7 7700K gives 45.81 Gigaflops...

I better sign off for the night. I'm starting to get excited...
10 more 1060's? 65 teraflops?

One thing to share that is relevant...
On this ASUS B250, the PCIeX16 slot on the board, you can only use a short single width card in that slot otherwise you cant get to the connectors on the board!! I have an FX 570 in there now and I find its better to NOT install the driver for that as it screws up the BIG GPU's. Let windows call it a generic video display and leave it alone to get your BIG GPU's going. At least at my house.

So, just to confirm, there is no CUDA for the windows platform?

I got 8 of the 9 1060's up and running!

1/2 the power

1/2 the heat

Almost twice as fast per work unit?

Cant tell if # 9 was a bad connection or a bad GPU. I'll tinker that out in the coming days.

If the CUDA on Linux is really twice as fast as Windows, then I need another 10 1060's to play with ASAP!

You guys (and gals) must think I'm nuts but I have an ulterior motive besides Seti.

I crunch electromagnetic sims for dollars and my simulation PKG uses GPU's!
If what I find on the net is right, 8 1060's give 30.4 Teraflops and the i7 7700K gives 45.81 Gigaflops...

I better sign off for the night. I'm starting to get excited...
10 more 1060's? 65 teraflops?

One thing to share that is relevant...
On this ASUS B250, the PCIeX16 slot on the board, you can only use a short single width card in that slot otherwise you cant get to the connectors on the board!! I have an FX 570 in there now and I find its better to NOT install the driver for that as it screws up the BIG GPU's. Let windows call it a generic video display and leave it alone to get your BIG GPU's going. At least at my house.

There is no CUDA90 or + for Windows. There are slower CUDA32, 42, and 52 for Windows. Normally when you first start crunching the Seti scheduler will offer some of those CUDA tasks. After a while, for the newer, the scheduler usually, eventually, starts feeding your SOG tasks.

It is entirely possible you could make it onto the first page or at least the 2nd page of the Seti Leaderboard using the Windows SOG tasks and a lot of Gpu's. Your problem would be, SOG at least, seems to want a lot more hand shaking with the cpu than any of the other gpu-based processing tasks.

You might want to run Windows for a while to "shake down" the system and see what your Windows-baseline looks like. I expect you to be getting 7 minute SOG tasks. If they are running 15 minutes, something isn't quite right.

Tom

---

A proud member of the OFA (Old Farts Association).

Cool. Sounds good. Thanks for all your wisdom Tom.
I will try and get Linux running tomorrow.
The miner RIG is running and will all day.

Dave

Oh and just to be clear?
SOG stands for...

SOG stands for...

From the author:

Signals on GPU.

It means spikes, autocorrs and gaussians are stored completely on GPU now and retrieved only on checkpoints. Much less synching for some ARs and hence much less issues with NV's realisation of that synching in their OpenCL runtime.
And perhaps better GPU usage after all.

---

Thank you Sir!

Okay swim fans!

I got a second batch of 9 1060's on the way! (Just dont tell my wife)
If the one I think had issues last night was just a connection issue, I will be able to play with a fully populated ASUS mining rig for Seti with 18 1060 GPU's in it!
68.4 teraFLOPS in one machine!

Okay swim fans!

I got a second batch of 9 1060's on the way! (Just dont tell my wife)
If the one I think had issues last night was just a connection issue, I will be able to play with a fully populated ASUS mining rig for Seti with 18 1060 GPU's in it!
68.4 teraFLOPS in one machine!

Just don't make them "belly flops" or it will HURT!







Running and hiding..... more or less.

Tom

---

A proud member of the OFA (Old Farts Association).

SoG is significant faster than cuda 32,42,50. Cuda 90 is only for Linux. The only developer able to import it to windows is MIA. So don’t look for it anytime soon. SoG requires 1 full code per work unit. The cuda don’t. But SoG is significant faster. OpenCl is restricted to only 26% of all gpu Ram. So bigger the ram on a gpu the more we you can use. Higher end gpus have faster rams so they are quicker than lower end. Multiple work units on higher end will deliver better results. 3 in 15 minutes or 5 minutes per work unit. I was able to get it down to 12 minutes. Linux cuda special now is around 75 seconds so it no contest.

---