Properly test the CPU before buying

[dnb]

I am about to purchase a used AMD EPYC 7763 processor. The seller agreed to install the processor to my system and test it with benchmarks. I'm going to use the HPCG benchmark for beginning. I only have my FreeBSD system and would not like to use other systems.

Questions:

1 - What tests on FreeBSD would you use in this case?

2 - What logs to look for in search of errors? dmesg, messages?

Maybe someone will advise me a more sophisticated approach for CPU tests?

Thank you

 

[Phishfry]

Run a make world with all cores and see what it does...
You really need fast storage too for real deal.
Cores plus fast storage.

 

[covacat]

'defective' cpus are extremely rare
i've only seen one that misbehaved and i've seen lots of them

 

[Phishfry]

You are lucky to get somebody to do that for you.
AMD is doing weird stuff with processors locked to certain manufacturers.(Dell,HP,IBM)
You really have to research what you are buying.
Someone proving it works is probably enough.
Research cooling requirements.
There are high watt Epic CPU's

 

[SirDice]

Run a make world with all cores and see what it does...

Yeah, that's typically my litmus test too. Building world pushes just about everything on the system at once. CPU, memory, I/O.

 

[mer]

Funny how make world with all cores is a good test of hardware.
Pushes CPU, memory, I/O.
"It works fine when I'm editing in emacs but a make world causes reboots".
Besides cooling, I always oversize the power supply and get a higher quality one. A lot of consumer grade systems (I'm talking about you Dell) provide a barely adequate PSU for the system they ship. Add one card and it falls overs. Spending more upfront on PSU saves cost down the road.

 

[VladiBG]

Prime95 - Wikipedia

en.wikipedia.org

To maximize search throughput, most of Prime95 is written in hand-tuned assembly, which makes its system resource usage much greater than most other computer programs.

I observe the same thing. Prime95 was the only software that can cause higher CPU temp compared to all other cpu stress tests. Most likely it use some specific cpu instructions which cause higher cpu usage compared to the other benchmarks like cinebench or intel XTU.

math/mprime

 

[dnb]

Prime95 - Wikipedia

en.wikipedia.org

I observe the same thing. Prime95 was the only software that can cause higher CPU temp compared to all other cpu stress tests. Most likely it use some specific cpu instructions which cause higher cpu usage compared to the other benchmarks like cinebench or intel XTU.

math/mprime

Yes, this is a great benchmark, thanks!

 

[dnb]

As a result, I settled on 3 benchmarks:

1. High Performance Conjugate Gradients (HPCG) Benchmark.
2. mprime
3. Compilation of some LLVM projects: сlang, flang, compiler-rt (in general, compilation of LLVM is often used as part of testing Epyc processors).

All data will be stored in tmpfs if possible.

Now I'm going to manually compile a jail specifically designed for benchmarking, put it in a ZFS dataset, and move it between FreeBSD machines (using zfs send/receive) to test them all.

Q: What do you think about using a separate jail for this task? As far as I understand jails, the results will be almost the same as if I used the benchmarks directly on the host machine, right?

 

[Phishfry]

You noticed this right?

Default TDP 280W

That is what I would call a flame thrower. You got a good power supply I hope.

 

[George]

I think the search terms for google are: "benchmark cpu freebsd"
There is a port, too: benchmarks/ubench

Don't use current. It is slower, because it is optimized for debugging, at least according to the freebsd benchmark wiki.

 

[dnb]

That is what I would call a flame thrower. You got a good power supply I hope.

I have a good old Supermicro PWS-865-PQ.

Default TDP 280W

In general, this is the usual TDP for most Threadrippers, for example. More specifically, AMD's website says "AMD Configurable TDP (cTDP) 225-280W".

https://www.amd.com/en/product/10906

In addition, there is an interesting performance review of FreeBSD 13.0 vs DragnFlyBSD vs Linux:

FreeBSD 13.0 vs. DragonFlyBSD 6.0 vs. Linux On AMD EPYC 7003 Series "Milan" - Phoronix

www.phoronix.com

It all shows that it's worth it)))

 

[Phishfry]

I was running the value of 280W versus several small benchtop devices in my head.
16 quad core NUC's or one big whopper.

Back to what I said earlier. You really need a performant disk system to compliment this beast.
Quad NVMe at least.

 

[Phishfry]

865W Power supply for 280W Max CPU sounds OK if you dont have some power draining GPU.
I don't think fully loading your power supply is good. I would say 2/3 is my power budget. 3/4 Max. ie. 650watts max.

 

[dnb]

Back to what I said earlier. You really need a performant disk system to compliment this beast.
Quad NVMe at least.

Unfortunately, I have to save a lot on storage. At first I wanted a Samsung 980 Pro without knowing about the known firmware issues. When I realized this, I sold them and bought a pair of Samsung PM9A3 (m.2 22110) 1Tb. It is too expensive to buy more, and even more so, only two m.2 22110 drives can be put in a typical Supermicro motherboard (if using PCI 4.0 without a special dedicated m.2 device like AOC-SLG3-2M2). I hope that Godly ZFS will save me from data loss.

Since this is a FreeBSD forum, it should be mentioned as well. I took two disks and installed FreeBSD 13.2 on them with a ZFS mirror. FreeBSD performance is amazing on this system (with 16-core EPYC)! Certainly, the existing hardware for AMD Epyc is extremely FreeBSD friendly.

 

[Phishfry]

I sold them and bought a pair of Samsung PM9A3 (m.2 22110) 1Tb.

That is good choice. Mid-Enterprise class drives. You can always add more later.

 

[bob2112]

Besides cooling, I always oversize the power supply and get a higher quality one. A lot of consumer grade systems (I'm talking about you Dell) provide a barely adequate PSU for the system they ship. Add one card and it falls overs. Spending more upfront on PSU saves cost down the road.

It's not straightforward, oversizing PSUs typically results in lower energy efficiency which can be expensive in the long term. In some cases the TCO can be dominated by energy costs.

 

[bob2112]

"To maximize search throughput, most of Prime95 is written in hand-tuned assembly, which makes its system resource usage much greater than most other computer programs."

I observe the same thing. Prime95 was the only software that can cause higher CPU temp compared to all other cpu stress tests. Most likely it use some specific cpu instructions which cause higher cpu usage compared to the other benchmarks like cinebench or intel XTU.

I suspect it's more to do with the nature of the problem rather than the use of assembly. It may be useful as a stress test of the CPU and cooling, but it's so abnormal that it sounds like a very poor benchmark.

 

[mer]

It's not straightforward, oversizing PSUs typically results in lower energy efficiency which can be expensive in the long term. In some cases the TCO can be dominated by energy costs.

Agreed and I wasn't very clear. Personal experience with a Dell, from the factory everything was fine. But since I'm capable, I tried to add more RAM, an extra HD and the system fell over because the PSU was overwhelmed by the addition of a single device.

My comment was more of "consumer devices typically spec a barely adequate PSU so a first thing to try is a higher quality PSU, even of the same specs".

 

[taiwan740]

My CPUs are generally bought very used and attached to old hardware that usually turn up smelling of tobacco and dog-hair, so if they survive a hosing of Febreeze and still work after drying overnight, I'm calling that good enough