We're not talking about BOINC's built in benchmark. Which in itself is worthless. We're talking about benchmarking software such as Cinebench, 3DMark and others.

Some people are just too dumb to read about those benchmarks and see what they are actually benchmarking. They just look at the numbers like wow, that must be good.

Tired of repeating myself to someone who refuses to learn anything.

I you are interested in CPU-x verses CPU-y buy one of each and experiment to your hearts content. The rest of us mostly don't care.

The problem is do you run Boinc at 100% cpu usage or 90% or even 80 or less percent? That all needs to be factored in when you are using someone elses benchmark numbers. We can 'assume' alot of things but if you don't do it yourself there are ALOT of variables that can affect how benchmark numbers are arrived at.

I don't know what your argument is. I posted a screenshot that explained how that web site got their benchmark totals. It claims to use some "top secret formula" which basically means, they have no idea. Just plugging in numbers. I avoid sites that don't go into detail on how they get their results because it's poppy cock at best.

There are BILLIONS of possible combinations. You want to write something that can emulate all that. Go ahead.

Benchmark software developers are only going to write benchmarks for their target audience. As I have said, multiple times. They're just for gaming, office work and other similar stuff.

No reviewers buy or get free samples of EPYC (or any server grade CPU) because that's not their target audience. The benchmark you think you found for the 7742, the site doesn't even own it. They're just "guessing" what it can do, which is the wrong way.

There is no such thing as a "end all, be all" software suite that can test every possible thing a CPU can do.

As I said before. Benchmarks you find on sites like that do not benchmark for the niche market BOINC is in.

They target gamers mostly, but also office work and audio/video production/editing.

Oh sorry, Toms Hardware does that

Why is a benchmark not simply running a test suite? When I find a list of benchmarks on a website, they should have run a variety of programs to come up with that value.

In the beginning because people didn't want to run something they weren't getting credit for, nor did it help the Science the Project was trying to do as everyone essentially ran the same mini task.

Bench marks are not the be all and end all for the testing of a CPU, as no single bench mark does every thing (at least as far as I have seen), it there was a single bench mark that could test every thing then they would not have to run a "Test Suite" to see how a CPU preforms they could just run one programme.

personally I think he should just purge/delete them to cleanup anything in this kind of pending-limbo state.

Amicable numbers for gpu used 10 gb of ram

Small cache? LMAO

Dude, Google is like... right here. www.google.com

3900X = 64MB L3 cache
3900XT = 64MB L3 cache
7742 (EPYC, 64-core, ROME) = 256MB L3 cache

As I have told you before and I'll say it again. Benchmarks are useless metrics. If you want to know how a CPU performs, then get real world data from what you actually intend to use the CPU for.

The EPYC is good on every BOINC project. Not just Universe@home. I only picked Universe@home because I knew GPUUG users have EPYCs that run it nearly 24/7. So I knew I'd be getting good results from their EPYC hosts.

You are right. EPYCs are good at one thing. Number crunching which is what 99% of these BOINC projects need. What they are NOT good at is gaming which is what 99% of all benchmarks revolve around.

It sounds like you don't have enough RAM in the system to use it fully for that project. Add more RAM and watch RAM utilization.

A good rule to follow is having 1GB of RAM per thread. So you'll want at least 128GB of RAM for that 5995 TR Pro CPU.

Just keep in mind, 1GB is a good rule of thumb, but not the end all. Some projects can and will use more than 1GB per task. 1GB is the bare minimum. I've recently been going with 2GB or 4GB per thread on my hosts due to the increase in memory foot prints at some projects that I run.

then i guess you'll never know.

Because they don’t need the extra precision. And implementing it would ostracize a large portion of their user base, reduce a very large chunk of their overall processing power, and gain nothing in the quality of their science results.

So what are researchers running on the pro cards and why isn't that sort of work done in Boinc?

Ask them.

Because they don’t need the extra precision. And implementing it would ostracize a large portion of their user base, reduce a very large chunk of their overall processing power, and gain nothing in the quality of their science results.

No point being pedantic about verbiage. You did the wrong thing. That’s what matters.

Universe tasks are not standardized. You cannot compare single tasks to each other without knowing they are the same size computationally. I told you this from the beginning and you went and immediately did exactly what I told you not to do.

Your other mistake is relying on a single benchmark. There is no “one-size-fits-all” benchmark. You need to either pick a benchmark that’s more similar to the intended workload (or just use the intended workload as a bench) or take a collection of relevant benchmarks and compare those.

or Titan V which is better than all of those.

but Einstein uses DP also. and it is needed for the calculations, but it's only a small part of the overall computation (not 100% like MW) and they just coded the app in a way to revert that calculation to the CPU if the GPU doesn't support it.

it takes the tiniest bit of effort to google the specs. you don't have to theorize if an EPYC has "tiny" cache when you can go to the product page and see that it has 256MB L3 cache. not that Cinebench is very sensitive to cache size anyway.

your confusion is because you did exactly what I told you not to do. you compared two cherry picked tasks that are not comparable. if you look at systems that have been running continuously for a long time and have stabilized RAC, you'll see the 64-core EPYCs are about 4x, maybe 5x, faster than your 3900X/T, not 10x. Skillz was right, you need to be told everything twice.

Folding@home uses single precision

https://fahbench.github.io/details.html

it's two 64-core CPUs. HT is on, so that's 128 threads each, and 256 threads total for the system.
the only project the system runs is Universe, and it crunches at 90% CPU utilization (roughly 230 threads) 24/7
~3.0M ppd total, ~1.5M ppd each CPU.

comparing cherry picked tasks from hosts is pointless since universe has varying task lengths. they are not all the same so it's easy to end up comparing two tasks that aren't comparable.

go nuts and cope.