Floating Point Speed calculation
15 Oct 2011 7:27:49 UTC
Topic 196018
(moderation:
Hi,
the Floating point speed on the server status page shows
529.1 TFLOPS
Are these the total FLOPS for CPU+GPU or just the CPUs?
Only the CPU TFLOPS from the three apps add up to 113+176+239=528.
The total Floating point speed has the footnote
Quote:
(from recent average credit of all users).
But this can`t obviously be the case. The BRP4 has less then half the FLOPS of FGRP1 but twice the Credits.
I think for BRP4 when it was GPU-only a few weeks ago the CPU TFLOPS were < 10 but the Granted credit/100
was around 1 Million.
For FGRP1 and S6Bucket which are only CPU apps I get an ratio of 330000 credits/(TFLOPS * week) or 47000 credits/(TFLOPS * day).
If I use this ratio to calculate the TFLOPS for BRP4 with the Granted credit/100 = 1,603,745; I get 486 TFLOPS. This is the number equivalent to CPU TFLOPS.
All apps would add up to 486+176+239=901 TFLOPS.
So the project has way more calculation power then it shows on the server status page.
p.s. I noticed that BOINC-Stats only shows 240 TFLOPS.
Sascha
Floating Point Speed calculation
)
Your assumptions are mostly correct, with one exception: The "Floating point speed (from recent average credit of all users)" is indeed what it says. This is independent of the way (i.e. the WUs) this credit was gained from.
The "CPU TFLOPS (from successful tasks last week)" are actually "normalized" such that these add up to the "Floating point speed". It is true that this currently counts CPU FLOPS only, so it largely underestimates the FLOPS of BRP4. But it's the BRP4 share of the total "Floating point speed" that's underestimated, not the total computing power.
BTW: does somebody happen to know how BOINCStats calculates the computing power of a project?
BM
BM
RE: BTW: does somebody
)
I think it uses a very simple method.
BOINCStats always queries the data of each computers.
It simply adds the floating point speed of the active computers.
RE: BTW: does somebody
)
BOINC Stats FLOPS = RAC / 2.
But, if I understand right, it is a mistake - because 100 CS = 1 day * (1 GFLOPS + 1 GINTOPS) and if we have 50 000 000 CS/day performance, we have a 500 TFLOPS and 500 TINTOPS. But not "reverse" when 50 000 000 CS/day = 500 TFLOPS or 500 TINTOPS.
I think that in code of BOINC Stats mistake is present.