It shows the integer and floating point speed as determined by a run of the BOINC benchmark tool in the Manager on the Details page of a host.
This is from your link:
Measured floating point speed:3341.27 million ops/sec
Measured integer speed:9101.15 million ops/sec
The Manager has code that is supposed to benchmark the host every 30 days, but it often goes longer than that or sometimes never benchmarks the host.
You can tell that a host is brand new or that the benchmarks have never been run on the host if the Details page has 1000 million ops/sec for both values as that is what is the default value from the client_state file if the Benchmarks have never been run.
Measured floating point speed:1402.33 million ops/sec
Measured integer speed:27725.41 million ops/sec
It shows higher values for the FLOPS ratings than your 5950X example.
You surely don't believe that a RPi 3B+ is faster than a 5950X, do you?
I've seen phones benchmarked in the BILLIONS of FLOPS for both integer and floating point.
Comparing benchmark values between disparate devices is a fools errand.
The problem is the nature of the old benchmark code that is in the client. It might have been representative of the computers of the 1980's, but certainly does not reflect the true power of modern cpus.
Thanks Keith. I wasn't sure that that was the correct reading or not since both you and Ian had a small complaint (?) about Tom's use of wording being somewhat confusing.
The reason we have benchmarks in the client is because the original BOINC client code using the original credit mechanism needed to figure out how to calculate the estimated remaining time.
The original mechanism scaled against a theoretical computer capable of 1 GFLOPS in the Whetstone benchmark.
This computer produced 200 Cobblestones of credit in a day.
But the mechanism has never worked for gpus even with an attempt to correct for them later in the code when gpus started being employed.
Thank you Keith. When I looked at the floating point bench marks on universe at home and compared them to the RAC's it seemed like several of the 128 thread systems had low fp benchmarks but their results were high.
I hate it when I am fuzzy brained and don't even notice it.
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!
If you compare the benchmarks for Intel processors against AMD processors you will see a magnitude difference benefitting the Intel processors.
Benchmarks tend to be written with code compilation favoring the market dominant cpu. Which has been Intel for decades.
Different architectures and branch predictors for AMD hamper AMD in most generic benchmarks. And then of course some math libraries used in benchmarks actually intentionally hamstring performance when a AMD cpu is detected running. The MKL math library being well documented as one culprit.
But then AMD with Zen 4 actually gets an advantage over Intel main consumer class cpus with having AVX512 capability that has not been available in Intel cpus for the past couple of generations. An AVX512 benchmark will favor the latest AMD cpus but won't even run on the latest Intel cpus.
I just would not pay any attention to the benchmark figures and rely more on the RAC output of a host.
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!
I would be that AMD puts out cpu benchmarking stuff to othat shows theirs being better than Intel's stuff, I would think that someplace like Tom's Hardware or something would come up with an apples to apples comparison between the various cpu's.
Tom M wrote:Boinc benchmark
)
I'm not exactly sure what the "e@h project system details page is", but this is what I think it is, and I don't see any 'benchmark listing'.
https://einsteinathome.org/host/12949005
Plus, you originally said:
The same goes for U@H.
Proud member of the Old Farts Association
It shows the integer and
)
It shows the integer and floating point speed as determined by a run of the BOINC benchmark tool in the Manager on the Details page of a host.
This is from your link:
Measured floating point speed:3341.27 million ops/sec
Measured integer speed:9101.15 million ops/sec
The Manager has code that is supposed to benchmark the host every 30 days, but it often goes longer than that or sometimes never benchmarks the host.
You can tell that a host is brand new or that the benchmarks have never been run on the host if the Details page has 1000 million ops/sec for both values as that is what is the default value from the client_state file if the Benchmarks have never been run.
Why I say you can't trust
)
Why I say you can't trust the measured FLOPS speed is because if you look at a Raspberry Pi like this one
Raspberry Pi 3B+
Measured floating point speed:1402.33 million ops/sec
Measured integer speed:27725.41 million ops/sec
It shows higher values for the FLOPS ratings than your 5950X example.
You surely don't believe that a RPi 3B+ is faster than a 5950X, do you?
I've seen phones benchmarked in the BILLIONS of FLOPS for both integer and floating point.
Comparing benchmark values between disparate devices is a fools errand.
The problem is the nature of the old benchmark code that is in the client. It might have been representative of the computers of the 1980's, but certainly does not reflect the true power of modern cpus.
Thanks Keith. I wasn't sure
)
Thanks Keith. I wasn't sure that that was the correct reading or not since both you and Ian had a small complaint (?) about Tom's use of wording being somewhat confusing.
Now I understand completely!
Proud member of the Old Farts Association
The reason we have benchmarks
)
The reason we have benchmarks in the client is because the original BOINC client code using the original credit mechanism needed to figure out how to calculate the estimated remaining time.
The original mechanism scaled against a theoretical computer capable of 1 GFLOPS in the Whetstone benchmark.
This computer produced 200 Cobblestones of credit in a day.
But the mechanism has never worked for gpus even with an attempt to correct for them later in the code when gpus started being employed.
Computation credit
http://www.boinc-wiki.info/Recent_Average_Credit
Thank you Keith. When I
)
Thank you Keith. When I looked at the floating point bench marks on universe at home and compared them to the RAC's it seemed like several of the 128 thread systems had low fp benchmarks but their results were high.
I hate it when I am fuzzy brained and don't even notice it.
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: I hate it when
)
That's what happens to fuzzy brained people. Me... I just have a brain in fog most of the time.
Proud member of the Old Farts Association
If you compare the benchmarks
)
If you compare the benchmarks for Intel processors against AMD processors you will see a magnitude difference benefitting the Intel processors.
Benchmarks tend to be written with code compilation favoring the market dominant cpu. Which has been Intel for decades.
Different architectures and branch predictors for AMD hamper AMD in most generic benchmarks. And then of course some math libraries used in benchmarks actually intentionally hamstring performance when a AMD cpu is detected running. The MKL math library being well documented as one culprit.
But then AMD with Zen 4 actually gets an advantage over Intel main consumer class cpus with having AVX512 capability that has not been available in Intel cpus for the past couple of generations. An AVX512 benchmark will favor the latest AMD cpus but won't even run on the latest Intel cpus.
I just would not pay any attention to the benchmark figures and rely more on the RAC output of a host.
I wonder if we could find a
)
I wonder if we could find a "fair" CPU benchmark?
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: I wonder if we
)
I would be that AMD puts out cpu benchmarking stuff to othat shows theirs being better than Intel's stuff, I would think that someplace like Tom's Hardware or something would come up with an apples to apples comparison between the various cpu's.