Posts by petri33

1) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1962264)
Posted 28 Oct 2018 by Profile petri33
Post:
he's also stated that the GPU temps are over 80C and he has to remove overclocks in the summer to keep things in thermal limits.

i'm not saying the card wont run at all. just that they will run faster with better cooling. i personally dont want my cards running 80+C all the time. which is why I and many other invest in water cooling, or using risers to space the cards out to get better airflow.


Computer case side panel removed and a duct fan aimed directly at the 4 GPUs.
The cards are FE with single blower-type (squirrel cage) cooling fan @100% that blows the hot air out at the back.
Once a month compressed air does wonders removing the dust from the inside of the cards.

===============================+======================+======================|
|   0  GeForce GTX 1080    On   | 00000000:05:00.0  On |                  N/A |
|100%   60C    P2   127W / 215W |   3067MiB /  8118MiB |     99%      Default |
+-------------------------------+----------------------+----------------------+
|   1  TITAN V             On   | 00000000:06:00.0 Off |                  N/A |
|100%   59C    P2   145W / 300W |   7734MiB / 12036MiB |     97%      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce GTX 1080    On   | 00000000:09:00.0 Off |                  N/A |
|100%   68C    P2   168W / 216W |   2802MiB /  8119MiB |     98%      Default |
+-------------------------------+----------------------+----------------------+
|   3  GeForce GTX 108...  On   | 00000000:0A:00.0 Off |                  N/A |
|100%   78C    P2   243W / 300W |   3452MiB / 11178MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
2) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1961460)
Posted 22 Oct 2018 by Profile petri33
Post:
For the Folding@Home single precision benchmark the 2070 FE matched the results from the 1080Ti FE card.

Yeah, depending on the benchmark it looks like it bounces between the GTX 1080 to GTX 1080Ti performance levels. But given current pricing, you'd be better off with a new GTX 1080.

Having said that, Phoronix have released the Linux benchmarks for the RTX 2070, that is compute benchmarks.
For Linux compute work it generally matches, or beats, the GTX 1080Ti, and it's performance per dollar actually puts it in front of most of the other video cards (RTX and 900 & 1000 series).

While it might be a humho card for gaming, it looks like it could be an excellent computing performer.
It will be interesting to see how it does perform for Seti crunching.


Nice going! Try keepp2 and clock the memory to what it is in p0. Also set power limit to 300w.
3) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1960400)
Posted 15 Oct 2018 by Profile petri33
Post:
Hi,

The same executable may work.
However here is a link to source https://drive.google.com/open?id=14SD3SU6K5iGYY3GM3aVWSb5Y19N8v5wZ
and to an executable that has sm_75 (RTX2080&Ti) enabled https://drive.google.com/open?id=1b8LYfoCmzHjAjbdCHT0ig-wHxXNOYIjZ.

My compilation may need certain versions of some system libraries. TBar can compile one that is more compatible.

Petri
4) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1960326)
Posted 14 Oct 2018 by Profile petri33
Post:
For those wondering what an RTX 2080 is capable of, check out Bruce's runtimes (there were a few issues early on, but it appears those have been resolved).

From the APR value, it looks like it's running 2 WUs at a time and it's running SoG with period_iterations_num=20
For comparison, my GTX 1070s are running 1 WU at a time using SoG, aggressive settings including period_iterations_num=1

BLC23s are taking me 7min 22sec one at a time. On the RTX 2080, 2 at a time, with period_iterations_num=20 (not like my 1), 4min 10sec.
So (very roughly) 7min 20sec v 2min 10 sec for a given WU.

Arecibo VLARs are around 8min 45+ for me, around 2min 30 for the RTX2080.


Not so good ;)
On linux, running Petri's app with a GTX 1070, BLC23 take 2min 15.....
Does anyone have the BLC23 run time for a RTX 2080 using Petri's app ?


On GTX1080Ti BLC23 is 67 seconds. The Ti is a power hog and a way slower than the Titan V with 49 seconds and Volta having constantly 90W less power draw.

I hope someone with interest in Linux would get a RTX so that we would get nvidia-smi -l output of the power draw and some facts about the run times. I'll compile an executable for who ever can get a RTX2080 or a RTX2080-Ti to try with. EDIT: Corrected another typo from RTX1080 to 2080.
5) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1960321)
Posted 14 Oct 2018 by Profile petri33
Post:
Hi,
I have not compiled one for RTX yet. TBar may have done so or Gianfranco,
It may be possible to run the same code on RTX that runs on Titan V, but if it throws errors or does not run, I'll compile one.
Just ask.
6) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1957362)
Posted 26 Sep 2018 by Profile petri33
Post:
If the letters are RTX, the number will start with a 2.


I'm sorry if you are having a bad day. My typo was a 1 instead of a 2 after the RTX.
7) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1957359)
Posted 26 Sep 2018 by Profile petri33
Post:
I think that the reviews published and linked from this thread since I posted the message you quoted - disagree with you.


They can and they have the right to do so.

How do they disagree?
Did any of those reviews run Seti special app? or SoG?

Those 5120 cores that my Volta has - they do twice the job that my GTX1080 with 2560 cores. And my Volta is running at lesser clock speed 1552MHz vs 1923MHz.
The RTX has 34 sm and 128@time or 68 sm and 64@time - about 4352 cores. Both configurations can and will be used.

My power readings are from my daily driver and both the performance and the power draw can be estimated for the Turing being a derivative or a successor of the Volta architecture. I do not use RTX nor do I use tensor cores. Cuda cores, SMs and the cache/mem subsystem count.

I'll admit I'm wrong if proven so whenever I have a chance to compile and test with one [RTX] . I'm not going to buy one though until I've destroyed a total of four 1080GTX cards. (two still at the shelf).
8) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1957355)
Posted 26 Sep 2018 by Profile petri33
Post:
Did anyone knows if there is any RTX2080 actualy crunching SETI? So we could follow the performance in the real world.


If anyone could provide me with an access via terminal window (rlogin) to a machine with a RTX-2xxx/Ti and let me do the developing remotely ....
9) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1957353)
Posted 26 Sep 2018 by Profile petri33
Post:
Ask Petri to post his first 30 lines of his Event Log to show how BOINC reports his Titan V?
Wrong way round. I'm trying to re-write BOINC so that it tells Petri the truth.


Yes. And No. I do not need any version of Boinc telling me the truth about anything. You do not either.

Volta 80 sm. Each sm has 64 threads that have access to the fpu simultaneously (5120). And all 80 have their own cache and shared variables and 65k registers.

Pulse find algorithm can use all 80 SMs at a time to run 32 warps of 32 threads when scanning for pulses. The memory subsystem slows things more than the fact that there are only 5000+ cores executing fp32 instructions of those 80 thousand it could issue. And the Volta and I guess Turing can do address calculations (integer math) at the same time it is executing fp ops.

When I reboot next time I'll grab the log and post it here if anyone is interested. I'm sorry in advance if forget to do that. Please ask if you absolutely need it.

My guess is that the Turing performance will fall in the middle of the 1080Ti and Volta using less power than the Volta (144W) and way less than the 1080Ti (234W).

"Let me see of my 1080Ti. It has the capability to compete with the best of them. The power draw is just amazing - biggest I've ever seen. The 1080Ti I have is a big and heavy space heater. That's what I think."
10) Message boards : Number crunching : User achievements thread......... (Message 1957351)
Posted 26 Sep 2018 by Profile petri33
Post:
AND of course Petri's expertise in coding the 'Habanero Sauce' app :D


I LIKE IT! MMmmmm. spicy!

Next is the "ghost peppers" app!


You'll have to wait for that sauce to brew some more time. I hope I can distill the essence of it and put it into a nice little bottle with an ear-worm (jingle) suitable for both the existing and forthcoming generations. I'm thinking at my best when I walk outdoors...

p.s.
My days and or evenings go now with Pokemon go app walking with my wife and children (when the weather allows for it and sometimes even if it does not). My friend invite code is .... I joined in some three weeks ago - now at level 24. I think I look like the Snorlax.
11) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1953899)
Posted 5 Sep 2018 by Profile petri33
Post:
Hi Petri, will you be able to "virtual" test on a Turing card without physically having one in your possession? Or will you have to debug on actual owners cards when they attempt to use the special app?


I could use X-term window remotely if the machine had CUDA development environment and gnu C++ and I could deliver source code and off-line benchmark tool with WUs.
12) Message boards : Number crunching : GTX is dead long live RTX !! (Message 1953877)
Posted 5 Sep 2018 by Profile petri33
Post:
Petri has said he is happy with his Titan V and is not looking to upgrade to Turing.
Yes, but those 3 - 1080 series cards are getting worn out by now :D


Two 1080's on the shelf (spare) and two more plus one 1080Ti installed in the machine.
I'd be happy to toss all 1080's and Ti out if could get more Volta at an affordable price.

I may try a TuringTi next summer.

When my old cards fans have worn out and died I'll replace the cards.
13) Message boards : Number crunching : Arecibo Overflow Tasks Run With CUDA Special App Sometimes Filled With Triplets Only (Message 1953875)
Posted 5 Sep 2018 by Profile petri33
Post:
This is interesting. I found two BLC Overflows also filled with Triplets;
http://setiathome.berkeley.edu/workunit.php?wuid=3123251218
Download http://boinc2.ssl.berkeley.edu/sah/download_fanout/35a/blc11_2bit_guppi_58227_21475_HIP66640_0059.30937.0.22.45.81.vlar
http://setiathome.berkeley.edu/workunit.php?wuid=3123251224
Download http://boinc2.ssl.berkeley.edu/sah/download_fanout/4c/blc11_2bit_guppi_58227_21475_HIP66640_0059.30937.0.22.45.84.vlar
However, on these the other Apps mostly agree. So, why don't they agree on the Arecibo files when they agree on the BLCs?


I downloaded the first one. I'll run a test. The second one was gone before I could DL it.


The reference on CPU had 23 triplets and then 7 pulses. GPU reported 7 more triplets and stopped at 30/30 before reporting those 7 pulses.

On a noise bomb you can find all kinds of spikes, pulses and triplets. Triplets get more common when more high spikes are found because the odds that three spikes are evenly spaced increases. And the longer the sample the more there is those noise induced spikes. No alien signal, just cell phone, radar, satellite, ... electromagnetic noise from Tellus and some wasted seconds of processing time.
14) Message boards : Number crunching : Arecibo Overflow Tasks Run With CUDA Special App Sometimes Filled With Triplets Only (Message 1953849)
Posted 5 Sep 2018 by Profile petri33
Post:
This is interesting. I found two BLC Overflows also filled with Triplets;
http://setiathome.berkeley.edu/workunit.php?wuid=3123251218
Download http://boinc2.ssl.berkeley.edu/sah/download_fanout/35a/blc11_2bit_guppi_58227_21475_HIP66640_0059.30937.0.22.45.81.vlar
http://setiathome.berkeley.edu/workunit.php?wuid=3123251224
Download http://boinc2.ssl.berkeley.edu/sah/download_fanout/4c/blc11_2bit_guppi_58227_21475_HIP66640_0059.30937.0.22.45.84.vlar
However, on these the other Apps mostly agree. So, why don't they agree on the Arecibo files when they agree on the BLCs?


I downloaded the first one. I'll run a test. The second one was gone before I could DL it.
15) Message boards : Number crunching : Arecibo Overflow Tasks Run With CUDA Special App Sometimes Filled With Triplets Only (Message 1953349)
Posted 2 Sep 2018 by Profile petri33
Post:
Boincmgr could reschedule a noise bomb (any 30/30 task) to CPU. I think that would be waste of resources.

The definition of a pulse or a triplet has not changed. The special app finds valid pulses and valid triplets that are there in the WU.

There are just so many of them. 30 or more to choose from in a noise bomb. To begin with the data in a noise bomb is invalid. You can not make a valid report from invalid data. 30 anything means the WU has noise and it should be discarded and the computation stopped. Noise bombs have no usable data. They need not report anything. Just label the WU invalid.

Petri
16) Message boards : Number crunching : Arecibo Overflow Tasks Run With CUDA Special App Sometimes Filled With Triplets Only (Message 1953303)
Posted 2 Sep 2018 by Profile petri33
Post:
// WAS 11 in V0.97 before b2
#define MAX_TRIPLETS_ABOVE_THRESHOLD 13
that is in pulsefind cuda code.

Someone a long time before me has decided a hard limit. 11 or 13 or some value is the maximum of number of ...
If I make that number bigger it slows down triplet finding but reports more of them. If I make that smaller it misses triplets.

Easy to fix. Make it 11. Miss the rest of them.

The problem doesn't appear to be finding the signals, it's about their classification.

GPU application Triplet count: 30
CPU application Pulse count: 30
The CPU is the reference, so the GPU needs to match it's classification of Pulses, not Triplets.


No It Does Not.

EDIT: There are 30 or more of both of them. END EDIT

There are thousands of signals in a 30/30 noise bomb, When the limit is reached the search stops. It is a noise bomb.
Some software stop at some point and some other software stop at some other point.
Crippling a parallel find requiring serial order is something that no one wants.
We're going to the future.
The after processing software or a new type of WU is needed to check that a noise bomb is a noise bomb which it is anyway.
We need no consensus of by what signals (autocorr, triplets, gaussians, pulses, spikes) or whatever is searched for.
It does not matter if any SW says it is a noise bomb and any of them agree/disagree.
It is a noise bomb.
Let the system send it to a GPU to make someone happy <3.
17) Message boards : Number crunching : Setting up Linux to crunch CUDA90 and above for Windows users (Message 1953300)
Posted 2 Sep 2018 by Profile petri33
Post:
"bad arg: -pfl
bad arg: 512"

The flag -pfl N is not supported any more.
Please do not try it with UPPERCASE. It is still there so that I can test it on offline benchmarks and bug hunting.

Same applies to -pfe.

Petri

Hey Petri,
With all these Overflowing BLC11 tasks is there something different we can do to find out why some overflows end up Invalid filled with Triplets?

If anyone can find an Overflow filled with Triplets before it becomes Invalid and disappears, you need to grab it. Just look for an Overflow and see if it's all Triplets.
Unfortunately, it seem those filled with Triplets are Arecibo Tasks....oh well.



Hi,

A noise bomb has so many signals that it makes a standard SW and a special SW hard to decide when to stop recording them.
The WU is processed from start to finish and the collecting part of the software has to have a limit what to store in CPU RAM.
One round is to check signals that have been time-delayed or -forwarded and divided to lengths of 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, ....
Each round could provide hundreds or thousands of "found" signals.
Imagine that to be done in parallel.
Quite hard to select the "correct ones" since everything is "Correct" but only 30 can be reported.

A noise bomb is a noisebomb.
18) Message boards : Number crunching : Arecibo Overflow Tasks Run With CUDA Special App Sometimes Filled With Triplets Only (Message 1953297)
Posted 1 Sep 2018 by Profile petri33
Post:
One from me, sorry no fanout, it's gone ..
http://setiathome.berkeley.edu/workunit.php?wuid=3116515390


Thanks Brent for trying!
19) Message boards : Number crunching : Arecibo Overflow Tasks Run With CUDA Special App Sometimes Filled With Triplets Only (Message 1953295)
Posted 1 Sep 2018 by Profile petri33
Post:
// WAS 11 in V0.97 before b2
#define MAX_TRIPLETS_ABOVE_THRESHOLD 13
that is in pulsefind cuda code.

Someone a long time before me has decided a hard limit. 11 or 13 or some value is the maximum of number of ...
If I make that number bigger it slows down triplet finding but reports more of them. If I make that smaller it misses triplets.

Easy to fix. Make it 11. Miss the rest of them.
20) Message boards : Number crunching : Arecibo Overflow Tasks Run With CUDA Special App Sometimes Filled With Triplets Only (Message 1953292)
Posted 1 Sep 2018 by Profile petri33
Post:
A noise bomb is a noisebomb.

It filled with results (probable signal candidates).

It is about of reporting order.
Which one of them to stop recording 30 of them and
where to start printing/writing them (max 30) to a file to submit to the servers.

Every one of found signals from a noise bomb could be claimed "valid"!

Ones, Tens, Hundreds, Thousands, Tens of thousands ...

So No help.

Petri.


Next 20


 
©2018 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.