Cool, I'll take a look at that, though mine seems pretty happy with the manual setting. I'll probably leave it like that for a few days to make sure this build is happy on my machine. So far so good, guppies (blc02) running at around 9.5 minutes as opposed to 11.5...

Thanks,

Chris

---

It seems your machine is responding nicely to the new App. The Inconclusives have suffered a dramatic reduction, it would be nice if our machines ended up with a similar ratio as this machine, http://setiathome.berkeley.edu/results.php?hostid=6906726&offset=200 Shame it takes so long to work through all the existing Inconclusives. I haven't been able to compile an App in Sierra, so, I suppose I'll have to go back to El Capitan and xCode 6.2 to compile another CUDA 8 App. I hate leaving Sierra, it seems to be working very well with my MacPro 3,1....must be the new 367 driver. Yosemite and El Capitan are still stuck with the old 346 driver.

Two weeks down with a nasty cold, and I see there are some things to try on my sierra/El-capitan installs :) Will see if the 1050ti decides to play ball with any of the newer webdrivers/cudadrivers, then see if I run into similar XCode issues using the single mac makefile approach.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

So, have you had any luck with the newest CUDA driver? I'm still using the previous version which works very well. The post I made last week is about to drop off the front page without any responses;
Poor performance with OSX CUDA Driver 8.0.81 * 0 Replies * 21 Views * T-Bar * 1 week ago
I keep checking for a new CUDA driver but nothing so far. It's been almost a month since 8.0.81 was released.

Not that far with OSX yet, but will know soon enough. I'd expect it to be rather full of diagnostic/debug code and relatively unoptimised though. The trick to verify if this is the case, will be to compare some of the basic GPU Compute SDK bandwidth/compute samples on the same GPU under each of the OSes. Windows and Linux Cuda builds should be relatively mature, and OSX playing catchup. At least initial focus will likely be on Pascal, so my 1050tis should say something. If it comes down to a different driver model altogether (i.e. blame Metal) then it could take some time to mature. If it's more just they are focussing on Pascal first, perhaps to try grab the media/graphics market that are the most vocal about wanting NV GPUs, then the differences should be less pronounced on the newer hardware.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Hmmm, all that may sound semi-reasonable if you were just talking about Sierra & Pascal, however, cuda driver 8.0.81 also shows up in El Capitan as a recommended update. There haven't been any Webdriver Updates to EC, it's still using the same old 346.xx version as is Yosemite. As far as I know, there aren't any plans for any updates to El Capitan or Yosemite past 346.xx or Maxwell. So, there isn't any need for 'diagnostic/debug code'. Yosemite is still using the same 7.5.30 cuda driver as it was almost a year ago, I doubt it will ever see 8.0. The only OSes that can use cuda 8 is El Capitan & Sierra and EC will probably never see Pascal. Yet, the recommended cuda driver 8.0.81 works just as badly in EC as it does in Sierra. I just ran a task in EC with driver 8.0.71, then updated to 8.0.81 and ran it again;

Running on TomsMacPro.local at Sun May 7 03:45:18 2017
---------------------------------------------------
Starting benchmark run...
---------------------------------------------------
Listing wu-file(s) in /testWUs :
16fe08aa.12502.25021.6.33.13.wu

Listing executable(s) in /APPS :
setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80

Listing executable in /REF_APPs :
MBv8_8.05r3344_sse41_x86_64-apple-darwin
---------------------------------------------------
Current WU: 16fe08aa.12502.25021.6.33.13.wu
---------------------------------------------------
Skipping default app MBv8_8.05r3344_sse41_x86_64-apple-darwin, displaying saved result(s)
Elapsed Time: ………………………………… 3565 seconds
---------------------------------------------------
Running app with command : setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80 -unroll autotune -device 0
unroll limits: min = 1, max = 256. Using unroll autotune.
      148.20 real        21.60 user        15.80 sys
Elapsed Time : ……………………………… 149 seconds
Speed compared to default : 2392 %
-----------------
Comparing results
Result      : Strongly similar,  Q= 99.72%
---------------------------------------------------
Done with 16fe08aa.12502.25021.6.33.13.wu.

Done with Benchmark run! Removing temporary files!

Running on TomsMacPro.local at Sun May 7 03:50:02 2017
---------------------------------------------------
Starting benchmark run...
---------------------------------------------------
Listing wu-file(s) in /testWUs :
16fe08aa.12502.25021.6.33.13.wu

Listing executable(s) in /APPS :
setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80

Listing executable in /REF_APPs :
MBv8_8.05r3344_sse41_x86_64-apple-darwin
---------------------------------------------------
Current WU: 16fe08aa.12502.25021.6.33.13.wu
---------------------------------------------------
Skipping default app MBv8_8.05r3344_sse41_x86_64-apple-darwin, displaying saved result(s)
Elapsed Time: ………………………………… 3565 seconds
---------------------------------------------------
Running app with command : setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80 -unroll autotune -device 0
unroll limits: min = 1, max = 256. Using unroll autotune.
      335.09 real        32.66 user       243.12 sys
Elapsed Time : ……………………………… 336 seconds
Speed compared to default : 1061 %
-----------------
Comparing results
Result      : Strongly similar,  Q= 99.72%
---------------------------------------------------
Done with 16fe08aa.12502.25021.6.33.13.wu.

It certainly looks as though it is just a very bad driver release to me.
I had just installed 10.12.4 with the Pascal driver, was playing around with it, looking at MacRumors, and people started talking about not having CUDA with the Pascal driver. A couple hours later, up pops 8.0.81...and the crowd cheers. Except...it works very badly, and has for about a month.
;-(

Oh well, back to 8.0.71 for now. 8.0.81 works about 5 times worse in BOINC than it does in the Benchmark for some reason.

I found a benchmark tool that is supposed to test CUDA here, https://render.otoy.com/octanebench/
According to that, there is a noticeable difference between driver 8.0.81 and 8.0.71, it's just not as bad as with the SETI Apps.
It's a little worse than 5x when run in BOINC in both El Capitan & Sierra. It's 8 times worse when you compare the 8.0.81 times, but 18 times as bad with 8.0.81 when compared to 8.0.71.
That means what takes 2.5 minutes in BOINC with 8.0.71 takes 45 minutes with 8.0.81. In Octane it goes from a GTX 960 score of 51 to 46, the GTX 950 difference is about the same.
So, for some reason nVidia has released a driver for Everyone running El Capitan & Sierra when it was only needed for those few dozen people who are running the Pascal GPUs in 10.12.4. No One else needed this driver that totally breaks the SETI CUDA Apps with the last 2 generations of OSX. For anyone Not trying to run a Pascal GPU, I suggest you use the Previous driver here, http://www.nvidia.com/object/macosx-cuda-8.0.71-driver.html

So, for some reason nVidia has released a driver for Everyone running El Capitan & Sierra when it was only needed for those few dozen people who are running the Pascal GPUs in 10.12.4. No One else needed this driver that totally breaks the SETI CUDA Apps with the last 2 generations of OSX. For anyone Not trying to run a Pascal GPU, I suggest you use the Previous driver here, http://www.nvidia.com/object/macosx-cuda-8.0.71-driver.html

Could it be the same trick again that this and other large companies practice quite often: "see how slow your old hardware is - buy new one!".
And older hardware intentionally forced to be slow by new software updates in this trick.

---

SETI apps news
We're not gonna fight them. We're gonna transcend them.

For now I'm clinging to the theory they just released it too soon to try to get the conversation back on the desired track. The conversation at MacRumors was shifting from how many Pascal cards I'm going to buy, to, why isn't there a cuda driver for this card I was thinking of buying. They had to stop that line of thinking quickly. I was there because, like everyone else, I had just found the new Pascal Webdriver broke the existing cuda driver and was looking for a fix. They had just released 8.0.71 two weeks earlier to work with OS 10.12.4 and probably didn't realize it was going to be broken so soon. 8.0.71 is Not broken for all those using the Apple driver, or the Webdriver-367.15.10.45f01 released just two weeks prior to work with 10.12.4. Hopefully the next driver will fix what 8.0.81 broke, but, it's taking much longer than I expected for the next driver. If the next driver doesn't work, then I might change my theory.

Bah, neither of the newer Pascal supporting webdrivers recognise the 1050ti yet (in Sierra), so back to Linux then the virtualisation passthrough afterwards, while we play the waiting game.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

I decided to give it a try, just so I could say I did. I pulled the middle 950 and inserted the 1050. The Maxwell driver of course just said it was an nVidia card. So I installed the current Pascal driver and it saw the EVGA 1050 without any problem, but, it said I had to update the cuda driver. You know where this is going...
The benchmark times aren't that terrible, but every time I try cuda 8.0.81 in BOINC it becomes simply horrid. The normal times for the below would be around 142 seconds for the Shorty and 390 for the BLC2. For some reason the benchmark App said the 1050 was device 0 instead of 1, and Device 2 is the 950 in a x4 PCIe slot. The 1050 didn't like the cuda 7.5 App at all with the Pascal drivers, so it ran the 8.0 App.

setiathome_CUDA: Found 3 CUDA device(s):
  Device 1: GeForce GTX 1050, 2047 MiB, regsPerBlock 65536
     computeCap 6.1, multiProcs 5 
     pciBusID = 1, pciSlotID = 0
  Device 2: GeForce GTX 960, 2047 MiB, regsPerBlock 65536
     computeCap 5.2, multiProcs 8 
     pciBusID = 2, pciSlotID = 0
  Device 3: GeForce GTX 950, 2047 MiB, regsPerBlock 65536
     computeCap 5.2, multiProcs 6 
     pciBusID = 5, pciSlotID = 0
********************************************
Current WU: 18dc09ah.26284.16432.6.33.125.wu
---------------------------------------------------
Skipping default app MBv8_8.05r3344_sse41_x86_64-apple-darwin, displaying saved result(s)
Elapsed Time: ………………………………… 3517 seconds
---------------------------------------------------
Running app with command : setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80 -unroll autotune -device 0
unroll limits: min = 1, max = 256. Using unroll autotune.
      226.68 real        28.87 user       138.26 sys
Elapsed Time : ……………………………… 226 seconds
Speed compared to default : 1556 %
-----------------
Comparing results
Result      : Strongly similar,  Q= 99.70%
---------------------------------------------------
Done with 18dc09ah.26284.16432.6.33.125.wu.
Current WU: blc2_2bit_guppi_57403_HIP11048_0006.17091.831.22.45.71.wu
---------------------------------------------------
Skipping default app MBv8_8.05r3344_sse41_x86_64-apple-darwin, displaying saved result(s)
Elapsed Time: ………………………………… 4797 seconds
---------------------------------------------------
Running app with command : setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80 -unroll autotune -device 0
unroll limits: min = 1, max = 256. Using unroll autotune.
      468.84 real        73.31 user       183.70 sys
Elapsed Time : ……………………………… 469 seconds
Speed compared to default : 1022 %
-----------------
Comparing results
Result      : Strongly similar,  Q= 99.37%
---------------------------------------------------
Done with blc2_2bit_guppi_57403_HIP11048_0006.17091.831.22.45.71.wu.

Running on TomsMacPro.local at Tue May 9 06:25:56 2017
---------------------------------------------------
Starting benchmark run...
---------------------------------------------------
Current WU: 18dc09ah.26284.16432.6.33.125.wu
---------------------------------------------------
Skipping default app MBv8_8.05r3344_sse41_x86_64-apple-darwin, displaying saved result(s)
Elapsed Time: ………………………………… 3517 seconds
---------------------------------------------------
Running app with command : setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80 -unroll autotune -device 2
unroll limits: min = 1, max = 256. Using unroll autotune.
      235.08 real        29.60 user       137.89 sys
Elapsed Time : ……………………………… 235 seconds
Speed compared to default : 1496 %
-----------------
Comparing results
Result      : Strongly similar,  Q= 99.70%
---------------------------------------------------
Done with 18dc09ah.26284.16432.6.33.125.wu.
Current WU: blc2_2bit_guppi_57403_HIP11048_0006.17091.831.22.45.71.wu
---------------------------------------------------
Skipping default app MBv8_8.05r3344_sse41_x86_64-apple-darwin, displaying saved result(s)
Elapsed Time: ………………………………… 4797 seconds
---------------------------------------------------
Running app with command : setiathome_x41p_zi3t2b_x86_64-apple-darwin_cuda80 -unroll autotune -device 2
unroll limits: min = 1, max = 256. Using unroll autotune.
      504.53 real        69.34 user       184.64 sys
Elapsed Time : ……………………………… 505 seconds
Speed compared to default : 949 %
-----------------
Comparing results
Result      : Strongly similar,  Q= 99.37%
---------------------------------------------------
Done with blc2_2bit_guppi_57403_HIP11048_0006.17091.831.22.45.71.wu.

The 1050 is back in the Linux box and the 950 back in the Mac with the Original 10.12.4 drivers, the Maxwell drivers work much better.

This is the line I used for the Cuda 7.5 App, do you see any reason the 1050 would choke on it? The task would finish in a few seconds with all bad signals;
NVCCFLAGS = -O3 --use_fast_math --ptxas-options="-v" --compiler-options "$(AM_CXXFLAGS) $(CXXFLAGS) -fno-strict-aliasing" -m64 -gencode arch=compute_32,code=sm_32 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_53,code=sm_53 -gencode arch=compute_53,code=compute_53

Well.......
I made another post at nVidia here, https://devtalk.nvidia.com/default/topic/1006380/cuda-setup-and-installation/poor-performance-with-osx-cuda-driver-8-0-81/
Filed a Bug report... Things seemed to be going normal. They asked for instructions on using KWSN-OSX-bench-MB. I gave the instructions and links to the old Baseline Apps at Berkeley. A few hours later the Bug report seemed to disappear, even though it is still listed as Open in my Profile. A few hours later and the Developers site is just saying it is down for maintenance;

NVIDIA Developer Site is under going maintenance.
The site will be back by shortly.
We apologize for any inconvenience.

It's been that way since yesterday...for me. Is anyone else with a Developers account able to access the site? Seems odd to be down for so long.
Anyway, there use to be a Bug report here, https://developer.nvidia.com/nvidia_bug/1923801 ...use to be....

Don't forget many services may be pulled offline over the weekend for safety, as the mass ransomware attack happened late, and full damage may be unknown until during the working week.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

The site came back up early Sunday, however, I haven't heard anything from nVidia in a week. Nothing since them asking for instructions on the Benchmark App. They released CUDA driver 8.0.83 but it doesn't work any better on my machine. So, I'll just stay at 10.12.4 with the Maxwell drivers for now, they seem to be working very well with both the 7.5 and 8.0 Special App. The 7.5 Special App works well on my machine in Yosemite with the 7.5.30 driver and in El Capitan with 8.0.71. I might be at 10.12.4, 367.15.10.45f01, and 8.0.71 for a while.

The site came back up early Sunday, however, I haven't heard anything from nVidia in a week. Nothing since them asking for instructions on the Benchmark App. They released CUDA driver 8.0.83 but it doesn't work any better on my machine. So, I'll just stay at 10.12.4 with the Maxwell drivers for now, they seem to be working very well with both the 7.5 and 8.0 Special App. The 7.5 Special App works well on my machine in Yosemite with the 7.5.30 driver and in El Capitan with 8.0.71. I might be at 10.12.4, 367.15.10.45f01, and 8.0.71 for a while.

This cone of silence tends to descend prior to a new architecture coming, so with Volta V100 existing (supposedly) then the Driver and Cuda team may be sequestered (prevents leaks). Once the large die has support, I suspect the pattern of opening up new Cuda version + Drivers may happen suddenly/quickly, though the addition of Deep learning and the new Tensor processors may see a different timeframe than the last few generations. We'll see.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

I might have been mistaken, as I haven’t had a chance to download the new beta Xcode and tools, but it looked like at the developers conference, Apple finally gives you the ability to look in detail of what is going on in the GPU. I’m not sure how much of it is available for non-metal code but at least some visibility is finally being made accessible.

Chris

---

The site came back up early Sunday, however, I haven't heard anything from nVidia in a week. Nothing since them asking for instructions on the Benchmark App. They released CUDA driver 8.0.83 but it doesn't work any better on my machine. So, I'll just stay at 10.12.4 with the Maxwell drivers for now, they seem to be working very well with both the 7.5 and 8.0 Special App. The 7.5 Special App works well on my machine in Yosemite with the 7.5.30 driver and in El Capitan with 8.0.71. I might be at 10.12.4, 367.15.10.45f01, and 8.0.71 for a while.

This cone of silence tends to descend prior to a new architecture coming, so with Volta V100 existing (supposedly) then the Driver and Cuda team may be sequestered (prevents leaks). Once the large die has support, I suspect the pattern of opening up new Cuda version + Drivers may happen suddenly/quickly, though the addition of Deep learning and the new Tensor processors may see a different timeframe than the last few generations. We'll see.

As someone that has done software testing as a profession for ~18 or so years. There is defiantly a big hush between the time the final software is delivered and the planned release date.
The software company I currently work for delivered a software release to the client in early April but their expected announcement date isn't until this Friday. Unless they change it, but since it is a US government agency they generally tend to stick to the initial dates the give us.
Unlike the Canadian or EU agencies we work with. That change their mind at a drop of a hat without notice or updating their documentation...

---

SETI@home classic workunits: 93,865 CPU time: 863,447 hours
Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[

Just to break the cone of silence a little, I've been asked to participate in Cuda 9 testing (as I did with prior versions). So I'll look distant for a bit longer, since carefully running through the mill caught some problems before.

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

Just to break the cone of silence a little, I've been asked to participate in Cuda 9 testing (as I did with prior versions). So I'll look distant for a bit longer, since carefully running through the mill caught some problems before.

I was kept in silence. Nice to know.
But anyway, I'll be busy this summer.
When I run out of WUs and I have time (and especially the specs) I'll devote a day or two to catch up.

Thanks Jason! I'm listening. WoW.

---

To overcome Heisenbergs:
"You can't always get what you want / but if you try sometimes you just might find / you get what you need." -- Rolling Stones

There may be Hope! After months of new CUDA drivers that basically didn't work with the Special App along comes the CUDA 9.0 Toolkit with CUDA Driver 9.0.103 that appears to Work.
I first tried it with the GTX 9xx GPUs in OSX 12.4 with the 367 driver and it worked OK, so, I then tried it with the 378 driver without any trouble. Swapped out cards to the 10 series and it still worked. Next was to update to OSX 12.6...it still works! At least with the CUDA 8.0 version of the Special App, the CUDA 7.5 version seems to have some weird Memory Hole, I wouldn't call it a leak, it's much worse. The CUDA 8 version works though, and it seems to work a little faster with all the updates.
Finally.
Coprocessors: [3] NVIDIA GeForce GTX 1050 Ti (4095MB) driver: 5899.27
Operating System: Darwin 16.7.0

I'm still trying to decide what to do with the Top card. They changed the spacing on the faceplate between the 950 & 1050 and now the trick I was using won't work. I may have to take the tin-snips to the 1050 faceplate to make it fit in the top slot...

I'm still seeing that BOINC Error as well;
Sun Sep 3 08:58:28 2017 | | Starting BOINC client version 7.6.34 for x86_64-apple-darwin
Sun Sep 3 08:58:28 2017 | | Data directory: /Volumes/Mov1/BOINC/Yosemite/BOINC Data
Sun Sep 3 08:58:29 2017 | | NVIDIA GPU 1: GeForce GTX 1050 Ti cannot be used for CUDA or OpenCL computation with CUDA driver 6.5 or later
Sun Sep 3 08:58:29 2017 | | NVIDIA GPU 2: GeForce GTX 1050 cannot be used for CUDA or OpenCL computation with CUDA driver 6.5 or later
Sun Sep 3 08:58:29 2017 | | CUDA: NVIDIA GPU 0: GeForce GTX 1050 Ti (driver version 9.0.103, CUDA version 9.0, compute capability 6.1, 4096MB, 3465MB available, 2255 GFLOPS peak)
Sun Sep 3 08:58:29 2017 | | CUDA: NVIDIA GPU 1: GeForce GTX 1050 (driver version 9.0.103, CUDA version 9.0, compute capability 6.1, 2048MB, 1992MB available, 1976 GFLOPS peak)
Sun Sep 3 08:58:29 2017 | | CUDA: NVIDIA GPU 2: GeForce GTX 950 (driver version 9.0.103, CUDA version 9.0, compute capability 5.2, 2048MB, 1999MB available, 2022 GFLOPS peak)
Sun Sep 3 08:58:29 2017 | | OpenCL: NVIDIA GPU 1: GeForce GTX 1050 Ti (driver version 10.18.5 378.05.05.25f01, device version OpenCL 1.2, 4096MB, 4096MB available, 729 GFLOPS peak)
Sun Sep 3 08:58:29 2017 | | OpenCL: NVIDIA GPU 2: GeForce GTX 950 (driver version 10.18.5 378.05.05.25f01, device version OpenCL 1.2, 2048MB, 1999MB available, 2022 GFLOPS peak)
Sun Sep 3 08:58:29 2017 | | OpenCL: NVIDIA GPU 2: GeForce GTX 1050 (driver version 10.18.5 378.05.05.25f01, device version OpenCL 1.2, 2048MB, 2048MB available, 619 GFLOPS peak)
Sun Sep 3 08:58:29 2017 | | OpenCL CPU: Intel(R) Xeon(R) CPU E5472 @ 3.00GHz (OpenCL driver vendor: Apple, driver version 1.1, device version OpenCL 1.2)
Sun Sep 3 08:58:34 2017 | | OS: Mac OS X 10.12.6 (Darwin 16.7.0)
Sun Sep 3 08:58:34 2017 | | Local time is UTC -4 hours

The CUDA driver has the cards in the correct order, the OpenCL part is wrong, and the 10 series are Not Pre-Fermi cards!