[Simbelmynë]

So initial Linux benchmarks shows how this can potentially be a beast for budget oriented systems that also do pre- and post processing.


https://www.phoronix.com/scan.php?pa...3d-linux&num=3

[flotus1]

It is worth noting that these results seem to be using this benchmark suite: Thoughts on Openbenchmarking.org
Sooo...I won't be spending 480€ on this "budget" CPU before seeing any other CFD-related benchmarks. More than a 2x performance increase from just tripling last level cache sounds too good to be true. I expect some improvement from it, but not this much.
What I find hilarious is the discrepancy in the level of detail about the test setup. Going into detail about the microcode version and display resolution, while noting "memory: 16GB". Just one step removed from "memory: yes"
I wonder if the 30M benchmark even fits into 16GB.

[Simbelmynë]

Agreed. Hopefully there will soon be more CFD benchmarks, in between the gaming related benchmarks. Even if the improvement is not as dramatic as Openbenchmarking.org would suggest, I do believe a larger cache can benefit some CFD loads. Milan-X seemed to benefit a lot from the extra cache in many HPC loads.

[flotus1]

Absolutely, larger caches/another hierarchy level are long overdue. We got a glimpse of it with Broadwell for desktop in 2015, and that was it.
Now finally both Intel and AMD will follow through with it, fingers crossed. The fact that it coincides with the transition to DDR5 makes for very exciting CPUs ahead.

[gnwt4a]

intel is the king and it seems will stay that way until zen845683246587324655687444787846X6D


https://www.igorslab.de/en/amd-ryzen...-some-light/8/

[Simbelmynë]

There are large discrepancies between Windows and Linux.


5800X3D is really good on Windows gaming, however for scientific workloads it seems like a really poor choice.


It is the opposite for Linux, where Linux games are rather unaffected by the extra cache, but scientific workloads benefit a lot.

[gnwt4a]

gaming differences between win and linux are due to gpu drivers. but cpu perf diffs when using win/linux are not detectable - unless u know otherwise which would be interesting.

[Simbelmynë]

Originally Posted by gnwt4a

gaming differences between win and linux are due to gpu drivers. but cpu perf diffs when using win/linux are not detectable - unless u know otherwise which would be interesting.

I referred to the relative performance of 5800X3D on Windows vs Linux, i.e. how the 5800X3D compares to other CPUs on Windows and how it compares to other CPUs on Linux. GPU performance is less relevant in this case for gaming, and really of no relevance when it comes to scientific benchmarks.



What you can see from many benchmarks on windows is that there is very little benefit of the extra cache when it comes to scientific workloads. But you can see that Windows gaming benefits from the extra cache on the 5800X3D.


On the other hand, you can see that Linux gaming is not affected much by the extra cache, while scientific workloads see massive improvements in some cases. The 5850X3D can compete with the 5950X in a number of scenarios where the 5800X is far behind. This is pretty amazing.



I think your question is answered pretty well in the initial article I posted. Phoronix is one of the best Linux hardware/software centered sites out there. Imho.

[Simbelmynë]

Originally Posted by flotus1

It is worth noting that these results seem to be using this benchmark suite: Thoughts on Openbenchmarking.org
Sooo...I won't be spending 480€ on this "budget" CPU before seeing any other CFD-related benchmarks. More than a 2x performance increase from just tripling last level cache sounds too good to be true. I expect some improvement from it, but not this much.
What I find hilarious is the discrepancy in the level of detail about the test setup. Going into detail about the microcode version and display resolution, while noting "memory: 16GB". Just one step removed from "memory: yes"
I wonder if the 30M benchmark even fits into 16GB.

I could not resist buying one since I am already invested in an AMD platform at home. I absolutely love the longevity of AM4 so this is sort of a tip to AMD as well


Anyways, since I have to run down in the basement to reset the system with each failed memory overlocking attempt I only have some simple memory settings for now:


5800X3D, DDR4 1R @ 3200 MT/s (14-14-14-14-28, 1T, FCLK=1600), OpenFOAM V9, OpenSUSE Tumbleweed, GCC 11.2.1


For 6 cores it finished in 138 seconds. This is an improvement over my previous 3700X with about 22% (the 3700X running memory @ 3600 and very tight timings, so not 1:1 comparison).


Compared to the 5900X the 5800X3D is about 12% faster looking at the result in the benchmark thread (that case was run with 3600 MT/s R2 memory but at somewhat loose timings).


It is also worth noting that it is faster than quad channel 7940X systems with 3200 MT/s R2 memory! It is clear that pushing memory bandwidth is not the only story of CFD.


My initial impressions are very good with this chip. I will do some proper testing in the coming days.

[flotus1]

Thank you for your sacrifice
My main takeaway here is solidified confidence in how utterly useless the OpenFOAM numbers on openbenchmarking.org are.
Someone with a solid grasp of OpenFOAM should look into this benchmark.

[Simbelmynë]

Originally Posted by flotus1

Thank you for your sacrifice
My main takeaway here is solidified confidence in how utterly useless the OpenFOAM numbers on openbenchmarking.org are.
Someone with a solid grasp of OpenFOAM should look into this benchmark.

Yeah, those numbers are very strange. The variation is really small which makes me wonder if there are many submissions from the same person/system, in which case the number of public results are just virtually inflated. I expect the Alder Lake systems to vary a lot since there is a huge difference between 4800 and 6000+ memory kits. Also, for Zen there has been huge advantages to fine-tuning the timings of the memory. Someone who plugs in a kit that defaults to 2133 will see perhaps 30-50% increase by setting it to 3600 and then another 10-20-ish % for fine-tuning the timings. Server systems are much more consistent for benchmarks since they are locked at certain JEDEC standards. Well I guess they are, unless you can find some custom BIOS that can tune the timings.



Anyways, right now the 5800X3D is 32% faster than my 3700X, using the same memory kit, and only a fraction of the time spent compared to fine-tuning the 3700X. I am pretty sure I can push it another 8% if I can convince myself it is actually worth the time investment