Hacker News Reader

Modern CPUs are not designed for a program that is so small it fits into uop cache and runs AVX nonstop.

This is a slippery slope towards CPUs which are only warranted to run certain approved apps. Overclocking aside, I'd consider a CPU broken if it can't run any sequence of instructions 24/7 without damage at the stock speed and voltage.

Besides, Prime95 isn't what I'd consider a pathological case anyway --- FFTs are common in scientific computing and signal processing, and having a CPU continuously perform them is not at all an unusual workload.

It somewhat reminds me of https://news.ycombinator.com/item?id=7205759

I agree with you, but with the caveat that CPUs should be able to run flat-out, indefinitely at rated voltage. In this case, people are over-volting the CPU by up to 40% so running the hardest workloads simply pushes an amount of power (heat) through the CPU that it was never intended to handle.

The way I see it is an over clock is only stable if it can run the same applications as a stock clocked chip could at full bore for the same time periods. So if you could take a stock 4770k or something and run prime95 small FFTs albeit at 90C for 24 hours with no issues then your OC should as well. Having a high standard like that means I have more confidence in my OCs and my systems.

> The way I see it is an over clock is only stable if it can run the same applications as a stock clocked chip could at full bore for the same time periods.

AVX has been on a separate clock/offset for a while now, and many OC'ing motherboards will let you tweak it.

Given how much power AVX consumes I wouldn't be surprised if throttling it even let you clock the rest of the core higher.

> run prime95 small FFTs albeit at 90C for 24 hours with no issues then your OC should as well

That's nice in theory but in practice if it's gonna crash it's gonna do it quick. I doubt many CPUs are suddenly erroring out for the first time at 23 hours. So why not just test for an hour and call it good? An hour of super abnormally high stress is more than enough to validate it for standard usage in my opinion. More than OK at 30 minutes too. Usually fails within the first 5 seconds, if you make it to 1 min then you have a very solid chance of making it the whole way. 30-60m is plenty.

If your data assurance requirements exceed one "proof test" shot[0], so to speak, then you have no business using overclocked CPUs at all. Prime95 SmallFFT is way way more stress than you would ever put a CPU under in any real-world load, and like an overpressured proof shot it's not really the best thing in the world for the processor to do this nonstop. Electromigration is a thing that exists.

[0] https://en.wikipedia.org/wiki/Proof_test

I think it's possible with exotic cooling solutions like custom water cooling. The costs are often ludicrous and not something you'd put into a cloud server but for a home workstation a good water cooling setup can get you better than stock performance such that you could run Small FFTs for hours or days and be fine. That sounds like overkill but it's peace of mind for me and I need that or I'll just keep the thing stock which I've been doing lately. My current 1620v2 xeon based workstation with a lot of enterprise Sas SSDs and tons of ram is at stock because the xeon line is locked down and it's plenty fast.

> If your system will handle 15-30m of Prime95 SmallFFT then you're pretty much good to go.

Wait, SmallFFT? Isn't that the test that focuses heavily on CPU and very little on the memory?

Since this news is about memory timings/overclocking, I'd rather put much more focus on tools like Memtest86+, running that for a few hours to be sure that it touches all the addressable regions and uses various access patterns.

Whatever test you do, you want to exercise all the memory, or else you'll end up with a computer that always seems fine after boot, but accumulates errors and crashes based on uptime and how much concurrent work you throw at it.

Memtest86+ tests the memory itself. I think Prime95 mixed mode is best if your goal is to stress the memory controller and all related subsystems.

Update: I mean blend mode

I'm curious as to your recommendation of SmallFFT in response to the grandparent's post. prime95, iirc, says SmallFFT does not test the RAM much. I think "blend mode" was the recommendation that includes RAM testing, while SmallFFT was more for stress-testing CPU overclocks?

Is it because the question is not whether or not a particular bit of RAM is damaged and rather whether or not the overclock to the communications bus between CPU and RAM is affected?

Yeah, blend is probably better if you're going for RAM overclocking. Which is really what's going on here. Or maybe something that's really hammering throughput or something (database loads come to mind).

Obviously you want to focus your testing on whatever subsystem you're trying to OC. I was just giving generic advice on overclocking. There is pretty broad consensus that Prime95 is a good stability test for CPUs.

Someone also suggested Intel Burn Test as a test that might catch some instabilities that Prime95 misses in a reply and then deleted it for some reason (why?).

In particular though, it's not just the memory subsystem you're overclocking here, you're really overclocking the interconnect as well. It will have its own "silicon lottery" and a particular point of instability, in addition to your individual memory kit's own stability characteristics. There are a lot of moving parts here.

I do wonder how much of this is getting the memory controller stable versus getting the interconnect stable. It seems possible they may have knobs they can twiddle on the interconnect as well.

[Serious question here:] I wonder if Prime95 or burn-test type tools would stress the interconnect with enough cross-core communication to be worth running.

I remember when I built my last gaming PC I unknowingly ran a version of Prime95 that could not pass an OC stress test on certain Intel processors.

The combination of instructions it used meant that even if, under sustained 100% CPU load with a normal benchmark, your cooling solution was sufficient, within 30 seconds the CPU would hit 100° and activate it's thermal protection before a shutdown.

I went through 3 water cooling setups and numerous thermal paste reapplications before I realized, since I was driving the CPU hard (4690k at 4.9Ghz with a high overvoltage I recall being considered at the limit for non-LN2 setups, 1.4?) and thought that cooling was the issue

Yup 24 hours of smallFFT or bust. And then a similar thing with any over clocked GPU.

Would Prime95 be likely to damage hardware if ran for a prolonged period on aws/gce/azure?

Do they have anything in their contracts telling you not to do that?

The hypervisor could throttle the workload even if it lit up the entire physical node (which is unlikely). The CPU could also throttle itself, the instances where Prime fried CPUs were due to overclockers forcing high frequencies and voltages AFAIK.

> Prime95 is real sensitive to instability

You mean 'really sensitive'. 'real' is the adjective form - 'really' is the adverb submodifier form, which is the one you want in this case.

It's acceptable in colloquial sense: https://en.wiktionary.org/wiki/real#Adverb ; Also, I heard it's more common in some US states.

Most of the time I dont use apostrophes.