PrimeNet credit CPU efficiency calculator

[James Heinrich]

What is the most efficient use of your type CPU? Now I have the answer!

http://mersenne-aries.sili.net/throughput.php

Just pick your CPU type from the list and enter the speed in MHz, and you'll get a nice table showing the relative efficiencies of each type of work.

You'll see some things like:

• Pentium IIIs should stay away from FFT work, ideally do TF <= 2^61 (but up to 2^63 is ok)
• Core i7 or Core2 are more-or-less equally efficient at FFT and 32-bit TF, but 64-bit TF is a bit faster
• etc.

Notes:
* I'm not taking multi-core issues into account, this is per-core.
* I'm not accounting for memory size or performance, which will affect P-1 and ECM (and L-L to a lesser extent)

Feedback/suggestions/bugreports are welcome.
If you CPU isn't in the list, run a benchmark and email it to me.

[James Heinrich]

After a little more playing, I'm now more confused than I was before.

It could very well be a bug in my code, but I can't spot it right now. As is shows right now, pretty much all the benchmarks seem to be heavily biased in favour of TF. I can't find a benchmark of the prototypical "GHz-Days 2.0GHz P4 Northwood" that has TF times, but these results from Northwood @ 3.0GHz should illustrate:
http://mersenne-aries.sili.net/throu...512|0&mhz=3000

Is TF really being assigned 2x-3x more credit than FFT-based work, or can someone point out my error?

[cheesehead]

Yeah, I wondered about no FFT efficiency over 43%.

Anyway,

1) Great idea!

2) How about including a short definition of what the "% Eff." number means?

[garo]

I wrote in the other thread about your numbers for PIII being correct but looking at your Core2 numbers I do think there is a bug somewhere. I get roughly the same credit per day with TF and LL on my Core2.

[James Heinrich]

Originally Posted by garo

...I do think there is a bug somewhere...

The key question is: where? What calculation have I messed up? I would be very grateful if someone can point out my error(s). Am I underestimating FFT work credit? Am I overestimating TF work credit?

(side note: A small issue that I'm ignoring (for now) is that until fairly recently (pre-v25.8) there was an off-by-one bug in the benchmarks. That may affect the relative performance of various bitdepths of TF, but does not explain the inappropriate differences between TF and FFT.)

[mdettweiler]

Originally Posted by garo

I wrote in the other thread about your numbers for PIII being correct but looking at your Core2 numbers I do think there is a bug somewhere. I get roughly the same credit per day with TF and LL on my Core2.

Yes, I see this too. For example, when I tried looking up the benchmarks for a stock Intel Q6600, a CPU for which there is surely plenty of data (considering its popularity), I see that it's 70% *worse* at FFT stuff than it is for TF. This definitely doesn't jive with what I've seen/heard for various types of Core 2's, which is approximately the same efficiency as garo said.

Edit: I'll see if I can run a benchmark on my Core 2 Duo sometime within the next day or two to see how it stacks up when checked on the website. That might help localize the error better, since it's for a CPU type (E4500) that I don't see listed at all in the drop-down menu, thus it would likely be the first (and only) submitted result.

[S485122]

Your data on LL testing is OK. The problem must be with the number of iterations for trial factoring, it is about one third of what it should be to give results consistent with my database of past results. How do you get that number of iterations ?

Jacob

[Mr. P-1]

Originally Posted by James Heinrich

If you CPU isn't in the list, run a benchmark and email it to me.

Check your email.

In fact, choosing the core 2 duo with the closest L2 cache gave results pretty close to mine.

How about allowing people to submit benchmarks via the site, and having the calculations done precisely for their machine?

All in all, a great app. I look forward to you fixing the credit balance issue.

Thank you.

[James Heinrich]

Originally Posted by S485122

The problem must be with the number of iterations for trial factoring, it is about one third of what it should be to give results consistent with my database of past results. How do you get that number of iterations?

It wouldn't surprise me if I was wrong. However, I got the numbers (as explained in this thread) from calculating number of iterations on 65 tests (from 2^58 to 2^62 each on 13 exponents from M216,091 up to M900,000,107) and my formula, rough though it is, produces numbers generally +/-2% from my counted values.

TheJudger seems to have something pointing in the right direction (maybe?) in this post but I don't quite understand what to make of it...?

[James Heinrich]

Originally Posted by Mr. P-1

Check your email.

Your benchmark is now included.

Originally Posted by Mr. P-1

How about allowing people to submit benchmarks via the site, and having the calculations done precisely for their machine?

A worthy idea, will take a little work to split out my benchmark-parsing code and let the throughput calculator pick up the one-off submitted data, but definitely doable and worthwhile. I'll get to it (sometime -- remind me in a month if I haven't made progress ).

Originally Posted by Mr. P-1

I look forward to you fixing the credit balance issue.

Using the magic multiplier of "3" seems to put TF values where I expect them to be, although the reasons behind it are still a little unclear to me. Hopefully it's correct (and someone can explain in simple terms why).

[Mr. P-1]

Originally Posted by James Heinrich

Your benchmark is now included.

Ta very much.

Using the magic multiplier of "3" seems to put TF values where I expect them to be, although the reasons behind it are still a little unclear to me. Hopefully it's correct (and someone can explain in simple terms why).

I'm not sure I really believe that the figures are comparable between the two tables, but it's reasuring anyway that I'm not being desperately inefficient doing P-1.