Reconditioning 8700k, question about fixed vs offset vcore...

[eddiew]

Hey folks,

A friend has shown interest in taking on my old 8700k and most of the parts. There was a power interruption about a year ago, after which it lost it's 5ghz overclock and all the bios settings and I was too lazy to ever properly fix it and just limped on at lower clocks. Before I sell it on, I'm trying to validate that it is still a reliable computer that has some years left in it and not something that's hanging onto life by the skin of its teeth.

It's been sat running y-cruncher for 10 hours overnight @ 4.8 GHz with 1.28 constant vcore. This feels like a nice sane overclock, and I'll run some more tests, but if it continues passing then I think I'm happy to call it good.

But the odd thing is... if I switch to offset voltage, it simply isn't stable in the same way. I've taken it up to the point it hits 1.36v under max load, at only 4.6GHz and it still fails stability tests within 2 hours. It also bumps 90 degrees, which might mean I need to refresh the heat paste! Probably still the same stuff I put in on day one...

Anyway I've varied from -40mv offset with aggressive LLC to +10mv offset with gentle LLC, feeling out where each of them results a vcore near to 1.35, and I cannot get either of them happy, not even at just 4.6GHz.

Is it worth chasing higher offsets, maybe +50mv, with even lower LLC to get the same loaded vcore? If I do that am I at risk of significant voltage spikes under lower loads? And is there some reason it would be inherently less stable with dynamic voltage vs fixed? And is it actually worth it?

Mostly an esoteric technical exercise at this point. In both cases, it will drop to 800MHz at idle, and even with fixed voltage, it's only using about 16 watts when nothing is going on so I'm not sure there is a lot to gain from further tuning - but I would love to understand why it can apparently tolerate a much lower fixed voltage than it can make do with adaptively.

Also be good to know if that trait is a sign of impending system failure! (I have a barely-used Ryzen 1600X+motherboard I could substitute instead so I'd rather make sure I pass on a working computer than one that dies in 6 months.)

Thanks for your thoughts

 

[eddiew]

If you want to pursue voltage offsets the first thing to do would be to make sure you're on the latest BIOS but why bother.

It would mostly be an academic exercise at this point But at least I've got these bios settings saved safely now, so I can experiment and revert easily.

Mostly I'd just love to know why I can run stable on basically 100mv less when it's fixed voltage vs offset. Not sure if it might be something in the VRMs on the board having problems keeping up with the rapid switching?

 

[eddiew]

Well my overnight-y-cruncher-stable kit bluescreened on desktop this morning. I'm confused

Guess I'll back it down to 4.6 and just use it, see how it goes.

 

[eddiew]

From what I've seen 1.28 is quite low for an overclock, if the temperatures are decent I would personally up the voltage to 1.29 and monitor the temps and stability rather going in the opposite direction.

It peaks about 81 degrees under max load with fans at max on a 280mm rad. Was never a cool chip, but I feel like it used to hit the same temps at 5ghz and more volts when it was new. I'll get some fresh heatpaste on it in the next week or so and see if that brings things down to give some headroom

 

[eddiew]

Damn thing failed Prime95 (non-avx) after 30 mins at 4.6Ghz, same vcore it passed 10 hours of y-cruncher @4.8 with, with only 4 Prime threads + everything else I was doing at the time. I'm so confused. Not a heat issue at this point, never passed 80C.

On the point of saying this thing is busted. I'll re-test at like 1.3v and 4.5Ghz, but if it can't handle that then something is degrading hard and I need to not pass it on to someone I see regularly! Rather give him a nearly-unused 1600X than an 8700K that's about to die...

 

[eddiew]

Well, this morning was an adventure... before giving up, I decided to replace the heatpaste. Cpu out, all cleaned down with white spirit, wiped around the socket... one corner of the cloth dragged over the socket... you can see where this is going but I didn't quite catch on at the time. Sadly the cloth did.

Of course it didn't boot any more. Post code 0A suggested a memory problem, but I had my suspicions, and indeed on close inspection with my phone's macro camera, I had bent one of the pins in my moment of thoughtless hamfistery.

But with very delicate manipulation from a needle-pointed dental probe while holding the phone in my other hand to act as a magnifier and torch...

Can still see the misalignment but it's a lot better than it was!

And the thing boots again! Initial impression is that it's shaved about 5-10 degrees off.

I've been refreshing myself on 8700K overclocks and it does seem pretty normal to throw up to 1.36v at them. I am pretty sure I have an established point of stability at 1.28v and 4.4GHz, which has now survived days of on-and-off prime, gaming, and general usage. With this as my baseline I'll see if I can persuade it to come up a little higher. Used to be a 5.0 chip, would be nice if it still is, but I wonder if I might have pushed it a little too hard in the past...

 

[eddiew]

Update (although I suspect nobody is interested ) - seems solid and stable at 1.3 constant vcore and 4.7GHz. Did a little reading around VCCIO and VCCSA and settled on them as 1.1 and 1.15 respectively. The IO might come down further, the SA appears to be as low as I can get. But either way, this seems very sensible and safe for the long term and I hope it will give years more service at these settings.

Been running for several days with no odd behaviours, frequently with some stability tests loading 2-3 cores in the background, just to force continuous activity even when I'm not doing much. Prime95 with AVX is insanely hot, drawing over 240 watts and reaching 99 degrees before I backed it off. Without AVX, it sits around 75-76 degrees and 180 watts. Gaming temps more like 50, and that's with the fans and pump set to quiet mode. Currently running it while I'm at work, and if it all passes the day I will consider these settings a baseline. The cache is currently on a very sedate 4.2 and I might look at that, but it's probably more effort than its worth. Might have another go at offset voltage, but if it goes the same way as before, I'll leave it constant. Idle temps are low 30s at 25 watts anyway, it's just not that big a difference that it's worth compromising the load volts/thermals for.

If nothing else, this experience has made me realise how outright chilly a 65W AMD chip will be by comparison. Even if it does draw more like 90-100 under full steam, I can quite sensibly get away with an air cooler and not require a 280mm radiator like I have with this old girl. Although I really want to see those 7800X3D reviews before I commit to a 7700X for sure, I have a hunch it's going to be by far the most sensible of the new X3D options

 

[eddiew]

(I'm kind of just keeping a journal at this point, but who knows, maybe it will be of interest to someone else trying to revive an older system some day.)

I discovered that by dropping the ram from 3200 to 3000 I can lower the vcore by at least 20mv (haven't tested lower but will) and the VCCSA by a whopping 200mv (it needed nearly 1.3v previously), and possibly 150mv off the VCDDR. This dropped 7 degrees off the p95 (non-avx) temps. I may see about dropping VCCIO a bit too, but I don't want to add any more variables right now.

Feeling like the memory controller is the weak point here. To be fair to it, it is running 2x16gb of 3200 cas 14 Trident Z. It's not a lot in mhz, but the latency is about as low as I've ever seen, and for it's era it was considered high density. I'm not too worried about dropping to 3000, because from what I know, that memory difference will be like 0.5% overall performance and is probably well worth it to keep everything at low volts for longevity. I may try it at a more typical 3200 c16 as well, although idk if that would actually be a net loss. Going to have to actually use some performance benchmarks as well as y-cruncher and p95!

It's all turned into an esoteric but fascinating exercise

 

[eddiew]

Final entry to wrap up:

After much tuning and stress testing, the old girl runs reliably at 4.7GHz with 1.25 vcore, 1.05 vccsa, and 1.0 vccio. Memory is at 3000mhz c14, 1.35v. Currently testing the uncore/cache at 4500, already passed 150 loops of Linpack Extreme, just running Prime on it while I'm at work and I'll call it good if it passes.

Any attempt to get the memory up to 3200 massively raises the vccsa required, and bumps the vcore requirements as well. Frankly it ain't worth it for the gains. This is a solid overclock for the voltage and I'm happy without that extra 6%.

I am still curious as to the behaviour with offset voltage instead of constant. I might give it another shot with everything dialled in, but I was having a really hard time keeping it down that below 1.3v. With constant voltage, the reported VIDs are usually around 1.29-1.31, but they shoot up in adaptive mode and ask for 1.36+ which requires a negative 100mv offset. I suspect this doesn't go well when the cores downclock to some middling level at sub-1v and it's those midranges that crash it. If there was a way to tell it to only run the base 800mhz OR max, it'd probably go better

Anyway, I'm content with this now. With these volts it should live basically forever. If the IMC has suffered degradation back during the [email protected] days, it shouldn't degrade further, and it'll just be what it is until it ends up in the scrap heap. Hopefully the new owner will get some good mileage out of it though, since I know it's only for some specific light games.