An obvious question (sorry).

Aramis Wyler · 2020-09-02, 03:56

So I'm pretty excited to have gotten colab to compile and execute gpuowl, and am happily chugging along at about 25% of the way through my first PRP (on gpuowl). Now, the output says it will take 20 hours to run this (111M range) and so I'm like, ok, that sounds pretty awesome, how many GHzD/Day is that? Some back of the napkin math tells me that it is about 600.

Now that card does TF at about 6x times that fast. Is it really that much slower to do PRPs on GPUs, or is it more likely that I've done a bad job compiling or configuring gpuowl?

EDIT: For reference, the card is a Tesla V100-SXM2

Prime95 · 2020-09-02, 04:41

You cannot use Primenet GHzDays to compare TF and PRP efficiency. The GHz days formulas were set in stone based on how fast a 2008(?) Core2 Intel CPU performed these calculations. That CPU was (relatively speaking) good at LL, bad at TF. Thus, when an architecture was developed that was good at TF, the GHz-days credited to that architecture were inflated (compared to actual wall clock time invested).

Hope that made sense :)

Aramis Wyler · 2020-09-02, 04:59

That does make sense, thank you - though now I feel like I'm abusing the Top Producers ranking every time I TF something.

Is there some other unit of work that is used to work out the value of TF bit depth vs P-1 vs the PRP itself? Flops?

Viliam Furik · 2020-09-02, 11:56

Quote:

Originally Posted by Aramis Wyler

That does make sense, thank you - though now I feel like I'm abusing the Top Producers ranking every time I TF something.

Is there some other unit of work that is used to work out the value of TF bit depth vs P-1 vs the PRP itself? Flops?

You can convert the GHz-D/D to FLOPS (FLoating Point OPerations per Second), by applying a simple formula: 500 GHz-D/D = 1 TFLOPS (10¹² FLOPS). So if you have a GPU with TF performance say 2000 GHz-D/D, that's 4 TFLOPS in FP32 (single-precision floating-point operations).

Aramis Wyler · 2020-09-02, 14:35

Interesting. I think we've already established that GHzD/Day aren't comparable with TF vs PRP, so is there a different calculation converting PRP GHzD/Days to Flops?

Prime95 · 2020-09-02, 15:57

Quote:

Originally Posted by Aramis Wyler

Is there some other unit of work that is used to work out the value of TF bit depth vs P-1 vs the PRP itself? Flops?

The unit is "wall clock time".

To decide the correct TF level we compare "how many exponents can this hardware eliminate per day by TFing to 2^N" to "how many exponents can this hardware eliminate per day by PRPing".

Since the above comparison is different for each piece of hardware we kind of guess as to the average piece of consumer hardware to determine our target TF levels.

Viliam Furik · 2020-09-02, 16:07

Quote:

Originally Posted by Aramis Wyler

Interesting. I think we've already established that GHzD/Day aren't comparable with TF vs PRP, so is there a different calculation converting PRP GHzD/Days to Flops?

Well, it's all the same, at least that's how the server treats those numbers when displaying TFLOPS throughput (here).

chalsall · 2020-09-02, 16:25

Quote:

Originally Posted by Prime95

Since the above comparison is different for each piece of hardware we kind of guess as to the average piece of consumer hardware to determine our target TF levels.

Actually, it's a bit more than guesses... James stepped up many years ago to answer the question "Just where should we TF to?"

Before this we /were/ just guessing, but with really absolutely no idea what was optimal. Please see the charts shown on each drill-down page from his GPU Lucas-Lehmer performance comparison chart. For example, for a Tesla V100 it ***used*** to be economically optimal to go to 77 "bits" at 92M or so.

One of the exciting things about the project is development is ongoing. So the economically optimal cross-over points have changed several times over the years.

Now that the Proof Mechnisim has been introduced, DCs will soon (read: in a few years) be obsolete, so the cross-over analysis will once again have to be revisited.

We live in very interesting times!!!

P.S. Oh, also... Optimal is something to be strived for, but difficult to achieve. Further complicating the calculus is different people like to do different things. Their kit, time, and electrons...

P.P.S. Perfect is the enemy of good.

Uncwilly · 2020-09-02, 18:00

Quote:

Originally Posted by chalsall

P.S. Oh, also... Optimal is something to be strived for, but difficult to achieve. Further complicating the calculus is different people like to do different things. Their kit, time, and electrons...

And throw in the volume of available work force.

LaurV · 2020-09-03, 05:10

Yep. On James' graphic, the "PRP Line" has to be somewhere in the middle between "First LL Line" and "DC Line". The reason is that in the future, we will mostly do PRP+CERT, which is a bit more than a single LL, but less than two LLs. So, click on your hardware (GPU) and see where you are, and decide how high you have to go with TF with your hardware, to eliminate the exponents faster (wall clock time).

On the other hand, James, your filters are missing the newest cards (RTX30xx).

2020-09-02, 03:56	#1
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 19B₁₆ Posts	An obvious question (sorry). So I'm pretty excited to have gotten colab to compile and execute gpuowl, and am happily chugging along at about 25% of the way through my first PRP (on gpuowl). Now, the output says it will take 20 hours to run this (111M range) and so I'm like, ok, that sounds pretty awesome, how many GHzD/Day is that? Some back of the napkin math tells me that it is about 600. Now that card does TF at about 6x times that fast. Is it really that much slower to do PRPs on GPUs, or is it more likely that I've done a bad job compiling or configuring gpuowl? EDIT: For reference, the card is a Tesla V100-SXM2 Last fiddled with by Aramis Wyler on 2020-09-02 at 03:59

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Simple problem with no obvious solution strategy	mathPuzzles	Math	0	2019-02-16 11:08
Sorry if this is obvious	robert44444uk	Miscellaneous Math	51	2018-06-18 15:23
A piece of information obvious in retrospect	fivemack	Factoring	0	2014-05-01 07:08
Col. Chemistry, General Math & Capt. Obvious	Fusion_power	Puzzles	10	2013-09-19 03:41
Area of Triangle, non-obvious case	Unregistered	Homework Help	9	2012-01-19 12:26

2020-09-02, 04:41	#2
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 1110011101100₂ Posts	You cannot use Primenet GHzDays to compare TF and PRP efficiency. The GHz days formulas were set in stone based on how fast a 2008(?) Core2 Intel CPU performed these calculations. That CPU was (relatively speaking) good at LL, bad at TF. Thus, when an architecture was developed that was good at TF, the GHz-days credited to that architecture were inflated (compared to actual wall clock time invested). Hope that made sense :)

2020-09-02, 04:59	#3
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 3·137 Posts	That does make sense, thank you - though now I feel like I'm abusing the Top Producers ranking every time I TF something. Is there some other unit of work that is used to work out the value of TF bit depth vs P-1 vs the PRP itself? Flops?

2020-09-02, 14:35	#5
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 110011011₂ Posts	Interesting. I think we've already established that GHzD/Day aren't comparable with TF vs PRP, so is there a different calculation converting PRP GHzD/Days to Flops?

2020-09-03, 05:10	#10
LaurV Romulan Interpreter Jun 2011 Thailand 3³×347 Posts	Yep. On James' graphic, the "PRP Line" has to be somewhere in the middle between "First LL Line" and "DC Line". The reason is that in the future, we will mostly do PRP+CERT, which is a bit more than a single LL, but less than two LLs. So, click on your hardware (GPU) and see where you are, and decide how high you have to go with TF with your hardware, to eliminate the exponents faster (wall clock time). On the other hand, James, your filters are missing the newest cards (RTX30xx).