The Ryzen 3950X is Zen 2 based silicon with much greater FPU performance over the older Ryzen 2700X Zen+ architecture. Really night and day difference in the designs. That is primarily the difference in observed results.
On the machines I closely supervise, I'm not currently running any GRP CPU tasks. However, my daughter has an old quad core iMac sitting on her desk that is crunching GRP. Out of interest, I decided to take a look at the current tasks list. The results are quite revealing. Fortunately, there were results still showing for quite a range of different examples of recent past LATeah102nF.dat data files and I've listed these below.
Task Name (shows frequency)
Data File Name
Run Time
CPU Time
LATeah1020F_24.0_126_...
LATeah1020F.dat
55,750sec.
52,680sec.
LATeah1022F_1512.0_39249_...
LATeah1022F.dat
62,452sec.
58,369sec.
LATeah1023F_344.0_13860_...
LATeah1023F.dat
23,522sec.
22,184sec.
LATeah1026F_88.0_36_...
LATeah1026F.dat
55,496sec.
52,506sec.
LATeah1027F_584.0_61000_...
LATeah1027F.dat
20,593sec.
19,440sec.
LATeah1028F_40.0_432_...
LATeah1028F.dat
20,131sec.
19,047sec.
LATeah1028F_808.0_105551_...
LATeah1028F.dat
21,864sec.
20,641sec.
LATeah1029F_56.0_114_...
LATeah1029F.dat
19,845sec.
18,784sec.
So, the particular data file does make a huge difference to crunch times. It is likely that the disparate results for two different CPUs presented earlier may well have been mostly due to different data files rather than some "night and day difference" in CPU architecture. Unless you know which data file was being used, you just can't justify that claim. Sure, there'll be a difference, but rather unlikely to be as large as the original figures suggested.
The frequency being analysed also makes a much smaller difference. To show that, there are two results for LATeah1028F for frequencies of 40Hz and 808Hz. This gives an idea of the change in run time with frequency.
The most recent LATeah1029F tasks (for which there are available results on the iMac) seem to be some of the fastest running of all.
Actually Gary, I can make the claim. My two hosts represent the two different architectures. The 3950X is a Zen 2 cpu and the 2920X is a Zen+ cpu.
Both running at 4100Mhz and same memory speeds. Same LATeah1029f tasks average of 10 tasks.
3950X average = 1.047 hrs
2920X average = 1.38 hrs
Both run with a wisdom file.
I know that Tom only runs his cpus at base clock so the numbers he had in his post didn't seem unreasonable to me. I also know he has a tendency to overcommit the cpu and his numbers are always higher than they could be if he just ran with 90% cpu usage instead of 98% or 100% usage.
I run my hosts at 90% BOINC cpu usage loading all the time.
[Edit] One of the primary differences in the two architectures is that Zen 2 has a full 256 bit wide FPU register for each thread while the Zen+ architecture has only a 128bit wide FPU register that is shared between the two threads of the core. It takes twice the cycles to complete a math function on Zen+ compared to Zen 2. Zen 2 also has an added FMA function.
I also see the same disparity in crunching times for Universe@home cpu tasks. There is not the big difference in data types on that project as we see here at Einstein.
So, your 20 min improvement in crunch time due to the superior architecture proves that Tom M's ~5 hour improvement - down from ~8hrs to ~3hrs must also be primarily for the same reason? After all, your claim was, "That is primarily the difference in observed results."
Are you not willing to concede (or at least allow the possibility) that such a big change is probably mostly related to data?
My worry is that someone might actually believe the hype and think they'll get a massive output boost from a very expensive new CPU.
No of course not. I have no idea what tasks Tom was running at the time that he posted his figures.
Tom regularly leaves important details out and we always have to ask twenty more questions to ascertain what he really is talking about.
I just take exception to your assertion that advanced cpu architecture has no bearing in the compute times and ALL differences can be attributed to different tasks of different frequencies.
I posted my times with the exact same task type on cpus running the same speed and showed that the Zen 2 cpus are faster than the Zen+ cpus. At any speed.
Take a look at the top computers list at Universe. It is dominated by Zen 2 cpus.
I just take exception to your assertion that advanced cpu architecture has no bearing in the compute times and ALL differences can be attributed to different tasks of different frequencies.
You'd better calm down a bit before you give yourself a heart attack! :-)
My 'assertion' was, and still is,
Quote:
Sure, there'll be a difference, but rather unlikely to be as large as the original figures suggested.
How can you honestly claim, "... your assertion that advanced cpu architecture has no bearing in the compute times and ALL differences ..." is a true and factual representation of what I actually wrote? Did you even bother to read what I wrote?
If I wanted to be equally blind in interpreting your messages, I could declare that I too "... just take exception to your assertion that the data file being used or individual task frequencies have no bearing in the compute times and ALL differences can be attributed to different CPU architecture ..." - but I won't since I can understand that it would be quite dishonest and unfair to make such a claim.
Why do you have such difficulty in acknowledging that there are several contributing factors, including CPU architecture, and that the most important one may well have been the different crunch times for tasks based on different data files? Lots of people lurk on these boards. It's helpful to give them a balanced picture by including all possibilities.
Why did you feel like jumping in and hijacking my directed reply to Tom?
Tom was a long time member of GPUUG team and I have lots of experience dealing with his posts.
I simply gave him one of the possible reasons why he saw the large differences in times. I pointed out the architectural differences in his cpus. Nothing more.
I am currently trying the "process the command line" with BOINC shut off.
When I tried it yesterday with Boinc running it appears to have just timed out without ever creating a "end of job" message at the prompt.
At least I think that is what I understood was supposed to happen.
In any case.
"...Happy crunching..." not grumpily crunching. :)
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
"...Happy crunching..." not grumpily crunching. :)
Tom M
Apparently the command line prompt just comes back. But when I examined the subdirectory in question the wisdom file was no longer 0% long.
So I will see if I can get through the 1,000+ cpu tasks I have been "blessed" with. Eventually I would like to run 4 thread cpu tasks along with the gpu tasks for Gamma Ray without feeling overwhelmed by the cpu tasks.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Because your pc needs to stop crunching and be disconnected from the net etc to get the best results, how does an online program do that? Follow the links in the early part of the discussion and you will see how it's done.
BUT, I built my wisdom file while crunching a CPU task and crunching FGRBP GPUs tasks and saw a big time decrease when wisdom was applied (earlier posted here). Same thing for crunching CPU tasks on a couple of old MacBooks. Maybe I got lucky with those three cases and hit a string of#5 tasks that didn't change from before and after creating the wisdom file, so an apples-to-apples wisdom comparison was possible.
On an AMD Phenom II from ca. 2009, however, while also crunching FGRBP on two GPUs, Gamma-Ray#5 CPU time was slooow to begin with and wisdom seemed to make it even slower; slow enough that I aborted the task before it was 20% completed.
Ideas are not fixed, nor should they be; we live in model-dependent reality.
I am currently trying the "process the command line" with BOINC shut off.
When I tried it yesterday with Boinc running it appears to have just timed out without ever creating a "end of job" message at the prompt.
At least I think that is what I understood was supposed to happen.
In any case.
"...Happy crunching..." not grumpily crunching. :)
Tom M
Tom, you won't get any kind of message. You run the Terminal command, your cursor disappears for however long your cpu needs to profile your cpu as to its math capabilities, and then the Terminal cursor reappears.
That is all that happens. No message or anything.
If you want to monitor the process, open the Task Monitor and find the fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu process with running status and just watch until it disappears. It only uses a single thread of the cpu to do the profile. You should see only a small cpu percentage of cpu running, 3 or 5% or whatever and only one cpu at 100% in the cpu activity graph.
But after the Terminal process ends you will find the FGRPB1wisdom.dat file in the project directory.
Keith Myers wrote:The Ryzen
)
On the machines I closely supervise, I'm not currently running any GRP CPU tasks. However, my daughter has an old quad core iMac sitting on her desk that is crunching GRP. Out of interest, I decided to take a look at the current tasks list. The results are quite revealing. Fortunately, there were results still showing for quite a range of different examples of recent past LATeah102nF.dat data files and I've listed these below.
So, the particular data file does make a huge difference to crunch times. It is likely that the disparate results for two different CPUs presented earlier may well have been mostly due to different data files rather than some "night and day difference" in CPU architecture. Unless you know which data file was being used, you just can't justify that claim. Sure, there'll be a difference, but rather unlikely to be as large as the original figures suggested.
The frequency being analysed also makes a much smaller difference. To show that, there are two results for LATeah1028F for frequencies of 40Hz and 808Hz. This gives an idea of the change in run time with frequency.
The most recent LATeah1029F tasks (for which there are available results on the iMac) seem to be some of the fastest running of all.
Cheers,
Gary.
Actually Gary, I can make the
)
Actually Gary, I can make the claim. My two hosts represent the two different architectures. The 3950X is a Zen 2 cpu and the 2920X is a Zen+ cpu.
Both running at 4100Mhz and same memory speeds. Same LATeah1029f tasks average of 10 tasks.
3950X average = 1.047 hrs
2920X average = 1.38 hrs
Both run with a wisdom file.
I know that Tom only runs his cpus at base clock so the numbers he had in his post didn't seem unreasonable to me. I also know he has a tendency to overcommit the cpu and his numbers are always higher than they could be if he just ran with 90% cpu usage instead of 98% or 100% usage.
I run my hosts at 90% BOINC cpu usage loading all the time.
[Edit] One of the primary differences in the two architectures is that Zen 2 has a full 256 bit wide FPU register for each thread while the Zen+ architecture has only a 128bit wide FPU register that is shared between the two threads of the core. It takes twice the cycles to complete a math function on Zen+ compared to Zen 2. Zen 2 also has an added FMA function.
I also see the same disparity in crunching times for Universe@home cpu tasks. There is not the big difference in data types on that project as we see here at Einstein.
So, your 20 min improvement
)
So, your 20 min improvement in crunch time due to the superior architecture proves that Tom M's ~5 hour improvement - down from ~8hrs to ~3hrs must also be primarily for the same reason? After all, your claim was, "That is primarily the difference in observed results."
Are you not willing to concede (or at least allow the possibility) that such a big change is probably mostly related to data?
My worry is that someone might actually believe the hype and think they'll get a massive output boost from a very expensive new CPU.
Cheers,
Gary.
No of course not. I have no
)
No of course not. I have no idea what tasks Tom was running at the time that he posted his figures.
Tom regularly leaves important details out and we always have to ask twenty more questions to ascertain what he really is talking about.
I just take exception to your assertion that advanced cpu architecture has no bearing in the compute times and ALL differences can be attributed to different tasks of different frequencies.
I posted my times with the exact same task type on cpus running the same speed and showed that the Zen 2 cpus are faster than the Zen+ cpus. At any speed.
Take a look at the top computers list at Universe. It is dominated by Zen 2 cpus.
Keith Myers wrote:I just take
)
You'd better calm down a bit before you give yourself a heart attack! :-)
My 'assertion' was, and still is,
How can you honestly claim, "... your assertion that advanced cpu architecture has no bearing in the compute times and ALL differences ..." is a true and factual representation of what I actually wrote? Did you even bother to read what I wrote?
If I wanted to be equally blind in interpreting your messages, I could declare that I too "... just take exception to your assertion that the data file being used or individual task frequencies have no bearing in the compute times and ALL differences can be attributed to different CPU architecture ..." - but I won't since I can understand that it would be quite dishonest and unfair to make such a claim.
Why do you have such difficulty in acknowledging that there are several contributing factors, including CPU architecture, and that the most important one may well have been the different crunch times for tasks based on different data files? Lots of people lurk on these boards. It's helpful to give them a balanced picture by including all possibilities.
Cheers,
Gary.
Why did you feel like jumping
)
Why did you feel like jumping in and hijacking my directed reply to Tom?
Tom was a long time member of GPUUG team and I have lots of experience dealing with his posts.
I simply gave him one of the possible reasons why he saw the large differences in times. I pointed out the architectural differences in his cpus. Nothing more.
Keith and Gary, I believe
)
Keith and Gary,
I believe you two are "talking past" each other.
So give it a rest!
I am currently trying the "process the command line" with BOINC shut off.
When I tried it yesterday with Boinc running it appears to have just timed out without ever creating a "end of job" message at the prompt.
At least I think that is what I understood was supposed to happen.
In any case.
"...Happy crunching..." not grumpily crunching. :)
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
Tom M wrote: "...Happy
)
Apparently the command line prompt just comes back. But when I examined the subdirectory in question the wisdom file was no longer 0% long.
So I will see if I can get through the 1,000+ cpu tasks I have been "blessed" with. Eventually I would like to run 4 thread cpu tasks along with the gpu tasks for Gamma Ray without feeling overwhelmed by the cpu tasks.
Tom M
A Proud member of the O.F.A. (Old Farts Association). Be well, do good work, and keep in touch.® (Garrison Keillor) I want some more patience. RIGHT NOW!
mikey wrote:Because your pc
)
BUT, I built my wisdom file while crunching a CPU task and crunching FGRBP GPUs tasks and saw a big time decrease when wisdom was applied (earlier posted here). Same thing for crunching CPU tasks on a couple of old MacBooks. Maybe I got lucky with those three cases and hit a string of#5 tasks that didn't change from before and after creating the wisdom file, so an apples-to-apples wisdom comparison was possible.
On an AMD Phenom II from ca. 2009, however, while also crunching FGRBP on two GPUs, Gamma-Ray#5 CPU time was slooow to begin with and wisdom seemed to make it even slower; slow enough that I aborted the task before it was 20% completed.
Ideas are not fixed, nor should they be; we live in model-dependent reality.
Tom M wrote: Keith and
)
Tom, you won't get any kind of message. You run the Terminal command, your cursor disappears for however long your cpu needs to profile your cpu as to its math capabilities, and then the Terminal cursor reappears.
That is all that happens. No message or anything.
If you want to monitor the process, open the Task Monitor and find the fftwf-wisdom_FGRP5_1.08_x86_64-pc-linux-gnu process with running status and just watch until it disappears. It only uses a single thread of the cpu to do the profile. You should see only a small cpu percentage of cpu running, 3 or 5% or whatever and only one cpu at 100% in the cpu activity graph.
But after the Terminal process ends you will find the FGRPB1wisdom.dat file in the project directory.