AMD Ryzen Threadripper PRO 7995WX Linux Performance Benchmarks

[sophisticles]

qarium


LOL, that is so funny. Puget Systems is a reputable vendor of custom computer systems that has been in business for years with a specialty in content creation.

They also have excellent articles and blogs with benchmarks and they have an HPC blog that is maintained by a guy with a PhD in Chemistry, that has done some excellent articles on HPC using CUDA on Nvidia GPUs.

It's possible that you did get some "infection", but considering I have visited that site from Windows and Linux, on various browsers and nothing has happened, and no one else is reporting similar issues, you are either lying or experienced the issue due to other causes.

Having said that, I do think it's hysterical that you think I work for an Israeli intelligence company.

[sophisticles]

So what you are saying is that you think HPC processors are a scam because some tasks need a lot of ram and run slowly if they don't have enough ram? Isn't that blindingly obvious?

What i am saying is that the benchmarks are misleading because they fail to take into account the total workflow, they run for a short length of time that takes place entirely in ram but in real world scenarios workloads take hours, days and swapping takes place.

Assume you have 2 vehicles, one is a pickup truck and one is a box truck footer. The cargo areas are the system ram.

You have to transport 5000 bags of sand 10 miles. With the pickup truck you can load and unload the bed much faster and you can drive much faster, but you have to make multiple trips. With the box truck, it takes longer to load and unload, but it only takes 2 trips. If you also have an 18-wheeler available, you can do the whole delivery on one trip but it takes you longer to load and load.

The HPC benchmarks that reviewers do are the equivalent of only looking at the total travel time without taking into account the amount of time it takes to load and unload and dock.

If you look at only travel time, you would conclude that the 18-wheeler is the fastest choice, because you do one 10 mile trip. But if you factor in the amount of time it takes to dock, the load and unload time, you may discover that the pickup is the better choice because even though you have to make 4 trips you can load and unload it so fast and you can take a faster route that it more than makes up for the 18-wheelers extra cargo capacity.

Similar concept with these HPC processors, their strength is in the ability to use more ram, but there is no such thing as a free lunch, a system with 192gb of data in ram has to do 192gb worth of writes to disk at some point and this needs to be factored into the total time for test completion.

These benchmarks do not take that into account.

Are you able to understand this?

[DavidBrown]

What i am saying is that the benchmarks are misleading because they fail to take into account the total workflow, they run for a short length of time that takes place entirely in ram but in real world scenarios workloads take hours, days and swapping takes place.

That is nothing remotely like what you said earlier - you claimed HPC cpu vendors were scamming people because everyone using a high-end workstation runs Windows and has Windows Defender (or similar) running, limiting their performance.

I think pretty much anyone except the most enthusiastic salesperson will agree that benchmarks can be misleading. You have to understand what is being tested, and how it relates to what you personally want to do, and the way you want to do it, in order to judge how applicable the benchmarks can be. Independent benchmarks, such as those on this site, are not deliberately misleading - Phoronix tries hard to cover a variety of tests and situations. But they will never match exactly with what any one user actually does with their system.

(If you are doing real work, in real-world scenarios, and swapping is taking place on the work you are doing, you are doing things wrong. It doesn't matter if, say, your 100-tab browser session gets swapped out while you are running your simulation overnight, but if you are using some demanding software for a real task, and that task is swapping while working, you have screwed up. You need more ram, or better software, or you need to split your task into smaller parts.)

Assume you have 2 vehicles, one is a pickup truck and one is a box truck footer. The cargo areas are the system ram.

Assume not everyone in the world is American, and that other people have no idea what a "box truck footer" is. But I can still understand the point you are trying, and failing, to make.

You have to transport 5000 bags of sand 10 miles. With the pickup truck you can load and unload the bed much faster and you can drive much faster, but you have to make multiple trips. With the box truck, it takes longer to load and unload, but it only takes 2 trips. If you also have an 18-wheeler available, you can do the whole delivery on one trip but it takes you longer to load and load.

Why are you still misunderstanding this? No one is suggesting you replace your current system with a machine that has twice as many cores and half as much ram. You seem to have done that in the past, and been unaware that on most systems you can upgrade the ram. Buy a machine that has as much ram as you need. That's what everyone else does - especially if they are considering a 64-core monster cpu. No one spends $10k on a cpu and fails to spend $2k getting 256 gb ram - or more if they need more.

Similar concept with these HPC processors, their strength is in the ability to use more ram, but there is no such thing as a free lunch, a system with 192gb of data in ram has to do 192gb worth of writes to disk at some point and this needs to be factored into the total time for test completion.

These benchmarks do not take that into account.

Are you able to understand this?

That's simply nonsense. Total and complete rubbish.

There are some tasks that involve a significant amount of writing to disk (and if you are doing that, you buy several good PCIx SSD's to run in parallel - these processors come with a lot of PCIx lanes). But virtually any serious computing task, for which massive multi-core cpus come in useful, the working memory needed (including temporary files, which don't need to be written to disk if you have enough ram) outweighs the data output by orders of magnitude. I just checked with a build of my current project (which is not huge, taking about 30 seconds to build on my 8 year old 4C/8T processor) . The total source is about 14 MB. The working ram used during build is about 800 MB, though it's hard to tell as there are other programs running. It generates about 650 MB of temporary files, all in a tmpfs directory (so they don't need to be written to disk). The output files, that are the actual useful result of the build, total 0.6 MB. The data that must be written to disk is a couple of percent of the ram space needed for the build.

Back here in the real world, rather than whatever planet you inhabit, more ram generally means less disk work because more of your files are cached when you want to read them (especially if you use an OS that does a good job of caching filesystem data). And it means less temporary file writes, because you can put them on a tmpfs system. (Well, you can if you use a decent OS.) And it means writes are not bottlenecks, because they can be cached and then written out in parallel with more calculations (if you use a good OS).


Oh, and did you read Phoronix's other recent article https://www.phoronix.com/review/thre...-windows-linux ? The one noting that Linux is the dominant OS for systems using very high processor counts? The one measuring that Linux is 20% faster on this system than the Windows supplied by HPE? My comments about Windows and Linux is based on fact and personal experience - I use both systems, and see advantages and disadvantages of each.

Going back to the original AMD Ryzen Threadripper processors, Linux has long possessed a performance lead over Microsoft Windows. With Linux typically being the dominant OS of HPC systems and other large core count servers, the Linux kernel scheduler has coped better than various flavors of Windows when dealing with high core count processors. Paired with some early Windows issues that Microsoft and AMD has since worked through. Linux has gained a reputation of handling these HEDT systems better than Windows.

[qarium]

qarium


LOL, that is so funny. Puget Systems is a reputable vendor of custom computer systems that has been in business for years with a specialty in content creation.

They also have excellent articles and blogs with benchmarks and they have an HPC blog that is maintained by a guy with a PhD in Chemistry, that has done some excellent articles on HPC using CUDA on Nvidia GPUs.

It's possible that you did get some "infection", but considering I have visited that site from Windows and Linux, on various browsers and nothing has happened, and no one else is reporting similar issues, you are either lying or experienced the issue due to other causes.

Having said that, I do think it's hysterical that you think I work for an Israeli intelligence company.

Its very simple other people do not report
a problem because they only target single
individuals the so called "Targeted Individuals"
in an operation like this they have no interest
to target any random people that would cause
a high possibility to uncover the secret operation.

you maybe do not work for them and also maybe
Puget Systems does also not work for them .
but their servers are clearly infected by israel based trojan horse and they do carry out attacks from their website.

and your Puget Systems article who claims intel is faster than AMD Threadripper is clearly a Hoax because of course more cores need more RAM for the same task to benchmark with the same amount of ram of course is cheating for the intel cpu.

as soon as you give the threadripper enough ram it will beat intel in any benchmark. with the
exception of the TensorFlow benchmark maybe.

[sophisticles]

That is nothing remotely like what you said earlier - you claimed HPC cpu vendors were scamming people because everyone using a high-end workstation runs Windows and has Windows Defender (or similar) running, limiting their performance.

I defy you to show me where i actually said this.

I said that these Threadrippers, and similar high core count CPUs are a scam, that is true, but i have explicitly explained why i stated this. No where did i say it was because "everyone using a high-end workstation runs Windows and has Windows Defender (or similar) running, limiting their performance".

I did link to an article where an AMD engineer said that they were running into I/O bottlenecks with the 64-core Threadrippers, but you grabbed onto Windows Defender and ran with it.

If you can't understand what is being written, that's on you.

[sophisticles]

That's simply nonsense. Total and complete rubbish.

There are some tasks that involve a significant amount of writing to disk (and if you are doing that, you buy several good PCIx SSD's to run in parallel - these processors come with a lot of PCIx lanes). But virtually any serious computing task, for which massive multi-core cpus come in useful, the working memory needed (including temporary files, which don't need to be written to disk if you have enough ram) outweighs the data output by orders of magnitude. I just checked with a build of my current project (which is not huge, taking about 30 seconds to build on my 8 year old 4C/8T processor) . The total source is about 14 MB. The working ram used during build is about 800 MB, though it's hard to tell as there are other programs running. It generates about 650 MB of temporary files, all in a tmpfs directory (so they don't need to be written to disk). The output files, that are the actual useful result of the build, total 0.6 MB. The data that must be written to disk is a couple of percent of the ram space needed for the build.

LOL, this is so funny! This is what you are using as a frame of reference?

I am going to give you an actual example that you can try yourself and is something that mathematicians do.

i assume you are familiar with Pi, correct?

One of the things that mathematicians have done over the years is calculate Pi out to 100 trillion decimal places, originally to corroborate the various proofs that Pi is irrational and now to waste CPU cycles and waste electricity breaking world records:

https://cloud.google.com/blog/products/compute/calculating-100-trillion-digits-of-pi-on-google-cloud

This is something I have actually done on my home PC, write a program that approximates Pi using a variety of methods, from the simplest 22/7 to more complex calculations.

Let the calculation run into the millions of decimal places and then save the results to a spreadsheet for analysis.

The HPC benchmarks being run are the equivalent of only looking at the processing time, nut if you are running a simulation, a math analysis, a physics or chemistry analysis, or similar, you want to save the data.

You seem to think that people are going to buy a 64 core monster system with 128+gb of ram, run millions of calculations and then simply wipe the results from ram.

That's not the case, what is actually going to be done is the results will be written to disk and stored for further analysis.

This is the difference between benchmarks and real world, with benchmarks you look at how long it took to calculate 22/7 to 10 million digits, in the real world you save that data to a disk and the reality is you are probably saving it as you calculate, rather than at the end, so the difference between a benchmark result were everything is calculate in ram and then flushed so it is gone and reality where you calculate 1000 digits and constantly update a file on disk is significant.

But i don't expect you, and most AMD faithful, to be able to understand this.

You have demonstrated an appalling lack of ability at rational thought.

In fact, one could argue that your thoughts are Pi like.

Mmmm, pie.

[sophisticles]

Back here in the real world, rather than whatever planet you inhabit, more ram generally means less disk work because more of your files are cached when you want to read them (especially if you use an OS that does a good job of caching filesystem data). And it means less temporary file writes, because you can put them on a tmpfs system. (Well, you can if you use a decent OS.) And it means writes are not bottlenecks, because they can be cached and then written out in parallel with more calculations (if you use a good OS).

So you have no idea what you are talking about? In your mind, data is copied from disk to ram instantaneously and nothing is ever copied back, it just disappears.

Glad you are not a computer engineer.

[sophisticles]

Its very simple other people do not report
a problem because they only target single
individuals the so called "Targeted Individuals"
in an operation like this they have no interest
to target any random people that would cause
a high possibility to uncover the secret operation.

you maybe do not work for them and also maybe
Puget Systems does also not work for them .
but their servers are clearly infected by israel based trojan horse and they do carry out attacks from their website.

and your Puget Systems article who claims intel is faster than AMD Threadripper is clearly a Hoax because of course more cores need more RAM for the same task to benchmark with the same amount of ram of course is cheating for the intel cpu.

as soon as you give the threadripper enough ram it will beat intel in any benchmark. with the
exception of the TensorFlow benchmark maybe.

So you think that an Israeli Intelligence Company infected Puget Systems servers with malicious software capable of infecting Firefox running on fedora 39, in the hopes that I would one day post a link on a forum where you participate, with the goal of infecting your computer?

This despite the fact that they have no way of knowing who i am, who you are, that Michael was going to post this article or that either one of us was going to be members of this forum.

But despite all that, they "targeted" you.

I am going to share a true story with this forum. Years ago i met a guy that used to were a knit football hat and inside he had it lined with aluminum foil, so that "no one could read his thoughts".

When I asked him how someone would read his thoughts, he "explained" to me that aliens had kidnapped him and placed a chip in his head so they could read his thoughts. Bot only that, but the aliens had conspired with the CIA and the Mafia to pull it off, because none of them had the ability to do it on their own.

I asked him what he did for a living to make him such an attractive target for the CIA, Mafia and aliens and he told told me he pumped gas for a living.

I wish i was making this up.

This guy reminds me of that guy.

You need help my friend. I hope you get it.

Oh, and for the record, in the articles i linked, all systems had 128gb of ram.

If that isn't enough for the Threadrippers, then we have a problem.

[qarium]

I don't remember the name of the benchmark but it relied on int and float math operations and this application was also used to "burn in", i.e. test stability of new builds, especially overclocked systems.

The reason i bring this up was because there was one specific benchmark within this app where the P4 could not be touched, it would smoke everything. This was because that specific test fit entirely within it's cache and was run from there. If the sample size was increased so that it had to access system ram, then the performance tanked and you could see the performance drop over time if you plotted the results on a graph.

the name of the app you search for is Prime95.

https://en.m.wikipedia.org/wiki/Prime95

and the pentium4 was the fastest because it had hand writen assembler who was very special and could not be portet to other X86 architecturs.

your claim is maybe true to but i remember it was pentium4 assembler what did make it so fast.

[yump]

My god, sophisticles you are a masterful troll. Pi calculation, specifically, is notorious for being an unusually IO-intensive distributed computing problem. I am on the edge of my seat for your next choice of cherry-picked benchmark. Will you manage to once more cite something that has a result you like, but for a completely different reason that what you say?

Now, in all the excitement I kind of lost track of who's fanboying for which team here, and I sure as hell ain't gonna read the last 6 pages again to find out. But the funny thing is, you're actually right about memory being a potential problem for the tests in this article. The Threadripper systems all have only 128 GiB of memory, which puts them as low as 0.7 GB/thread for the 96-core. Amazon EC2, for comparison, provisions 4 GB/thread on "general purpose" instances, 2 GB/thread on "compute" instances, and 8 GB/thread on "high memory" instances. The Intel box with its slightly-more-than 2GB/thread is within the normal range, but based on market demand there must be common workloads that require even more than that.

qarium, please take your meds. You appear to be having some kind of episode, and are not thinking logically.

If it helps, consider that Puget Systems is a well-reputed workstation builder and reviewer and has been around for like, a decade, and if Sophisticles is a government agent posting links to pages with trojans hosted on their compromised servers, targeted at you specifically... he may be practicing Zersetzung, in which case beclowning yourself by schizoposting is playing right into his hands.