
Khenglish
Member-
Posts
19 -
Joined
-
Last visited
-
Days Won
1
Khenglish last won the day on December 16 2024
Khenglish had the most liked content!
Recent Profile Visitors
The recent visitors block is disabled and is not being shown to other users.
Khenglish's Achievements
-
I am in the US. Thanks for the offer. I was thinking of the X170 vapor chamber, but I'd probably break it removing and replacing the heat pipes. Even if I didn't temps would probably only improve by 1-2C. Temps are already excellent with the mods I have done. Much better than my P150EM with a Clevo 2080 monstrosity. That thing runs around 16C hotter at the same power.
-
Without the added heatsinks and external fan temps would cap out around 77C with the power limit mod after letting temps plateau. 67C at the default 155W limit. P775 radiators are less deep than the fan depth. curving the fan shrouds into rhe radiator helps temps a little.
-
The end of the 2nd GPU test would be 59C. No vapor chamber. P750 chasis with P775 heatsinks as the baseline.
-
The 3080 Ti pulls ahead at 160W or so, but it definitely underperforms given how much bigger the core is than the 3080 core. I'm only getting 15K in TS GPU because I'm hitting the card's current limit (not the same as the power limit!), which is an unreported throttle in monitoring tools. Best 3080 score is 15077 at 1986 average core clock: https://www.3dmark.com/spy/40731969 Best 3080 Ti score is 15645 at 1866 average core clock (this is similar to my reported core clocks): https://www.3dmark.com/spy/32697164 I assume both are at 200W. The Ti is only 3.7% faster, which is pretty bad for a 22% bigger core. I considered raising my card's current limit, but since I'm already struggling with GPU power draw with the 9900K I haven't and don't plan to.
-
This is the best I can do with my current limit issues. 43 points short of the 3080 record. I can't touch the 3080 Ti record due to the BGA systems having much better power delivery: Things I have learned: Classic Throttlestop fixes all low perf at low utilization issues. CPU would improperly downclock, then GPU would downclock even more severely. This even impacted the 6600K. RTX 3060 was fine, but RTX 3080 Ti was not. It's now no longer necessary to force a VF point in Afterburner, so now the card can properly idle without crippling performance Card stability: +195 stable full voltage range +210 stable at < 0.9V +225 stable at <0.8V Card spends most of it's time here under heavy load at default TDP +270 benchable ~0.85V Lots of artifacts, but it chugs along ThrottleStop also MOSTLY fixes crash at idle/light load issues if I force high performance. Crashes are a full system lockup C-state control in BIOS and TS does not work, so I don't have a means to fully prevent idle clock drop I will try borrowing another CPU from someone at work This CPU was listed as not working for $60 on ebay System fails to boot 50% of the time This is a bit annoying since I don't have a power button. I need to pull the power cord and replug to reboot Interestingly, the ThrottleStop benchmark is a great stability test. It'd give me crashes at higher voltage than anything else TS benchmark stability: 4.2GHz -90mV 4.3GHz -85mV 4.4GHz -85mV 4.5GHz -85mV 4.6GHz -75mV 5.0GHz -55mV I could subtract another 35mV for anything other than the TS benchmark CPU or GPU current spikes could trigger a lockup: Limiting CPU clocks helps stability Reducing CPU current limit would cause windows to fail to boot, so I could not attempt this 3D workload on 3080Ti could trigger a lockup if TDP is limit raised, but not if it wasn't Still can get idle lockup My plan is to make this a janky SFF desktop, but the occasional idle crashes are a problem for that. The 6600K doesn't crash, but that is a far slower CPU than the 9900K.
-
The km vapor chamber is a very intersting idea. I didn't know it was gpu-only, so if the heat pipes are moved it would fit. I worry about the champer expanding and getting ruined when replacing the heat pies though. I never see any EC panic, but I am running a 3080 Ti with it's own set of problems. When the card has a low 3D load it's performance craters due to some internal throttling. Card also has an OCP limit that's effectively around 200W. mxm 3080 power delivery is too weak for me to go about raising it, but this means I can't match the 3080 Ti BGA setups. I'll drop the power limit back down for normal use so this isn't a big deal. 9900K is giving me a lot of hardlock problems too at low load, but I got it for $65 so it's hard to complain. I managed to get a 8700K for free that will show up in a week. Hopefully it behaves better.
-
I'm currently using 775 series heatsinks as the baseline for mine. They stick out a little, but otherwise fit. I don't actually have a whole system. I got a motherboard for $25 and a bottom case for $22.
-
So I got a 9900K: I'm having lots of issues with it crashing when transitioning from idle to load. If I have it loaded it overclocks well and I can undervolt a lot. Playing around with fixed voltage now to see if I can get it to stop crashing for things like web browsers. With the 6600K pushing the GPU I almost got 15K. I haven't pushed the GPU with the 9900K yet since I'm trying to figure out the stability issues:
-
Fixed the VRAM clocking. I had a few incorrect resistor values for one of the added VRAM chips. Now runs 17.6GHz+ like with the 3060 core. Getting close to beating the best 3080 scores. The throttling is hurting me though and it increased a little with the higher VRAM clocks. My core clocks are similar to the best 3080 Ti scores, but they beat me by 800 points. Maybe these really are HC18 VRAM chips. I'll swap the rest over and find out.
-
So I've gotten a bit of improvement: The biggest help was improving the GPU supply voltage on the motherboard. I direct soldered the cable on one end, shortened the cable, and also doubled the motherboard GPU power switch FETs. This efficiency improvement let me increase the GPU voltage from 843mV to 875mV prior to the invisible throttling engaging. This plus added capacitors also let me increase VRAM clocks by 200MHz.
-
It didn't work, but I also tried it in an older desktop. These 30 series cards are really picky and it may have just hated the old Zen+ system. I also tried it in an M2 adapter to a full pci-e slot on a Zen4 system, but it wouldn't even power on. I couldn't fogure out what signal was missing. Also 30 series is rejected by the P150EM. 20 series works though.
-
Good find on that video. My soldering thoughts: Reballing is a PITA. I have found it easiest to use a direct heat minimally sized stencil since they warp much less. I'll also do an initially heating in a toaster oven with the stencil still in place with a very thin flux coating. The slow even heating prevents solder balls from jumping out. After that the solder balls are tacked down and I do a second heat cycle with the heat gun and much more flux to get good connections. I also only use leaded solder balls. They are so much easier to work with than unleaded. They bond at much lower temp, are stronger, wet more easily, and pull and center parts more strongly when molten. In general it's worth removing solder from unleaded preballed parts and reballing with leaded. If I had done this I wouldn't have lost the 3080 board. His flux seems less corrosive than mine. Mine causes pcb damage while his seemed to cause none. I should use a different flux The GPU glue is terrible. The reason why my 3060 pcb is so discolored is because the glue was still sticking after melting the solder to remove the 3060. I had the core under high temp for quite a long time until I got it off. Once the core is off just scrape off the glue with a hot soldering iron. This is way less dangerous to the board than the way he Dremeled it. I only use aluminum foil if plastic components are underneath. Components under foil can still get molten, and the slight pressure of the foil can shift them around. My perf thoughts: I also got the weird vram clock toggling. You can fix it by forcing a fixed VF target is MSI afterburner. You can see my scores are around 1K higher than his in TS, and I was just running 0.84V, which I think is sround 180W, similar to his max. My core can benchmark at +240, and is fully stable at +210, similar to his I'm curious in is vram clocks. Something is wrong with my setup.
-
I first tried this on an RTX 3080 board. Unfortunately, the reballing from the seller was faulty and several solder balls detached from the core when soldering to the card, which was visible after removing the core from the card. The core required a reball, but the 3080 pcb died after 2nd core solder cleaning. You can see the power/ground routing got very exposed in the middle. I tried resoldering the core on anyway but it shorted. This is the 3060 in-progress next to the 3080 prior to removing the 3080Ti core: As for future mods I do hope to fix the vRAM clocking and finish upgrading the chips for 16GB. Currently only 3/8 chips are 2GB. In theory a 30 series pcb can take a 40 series core. They are electrically compatible, but my viewpoint is finding a vBIOS that would take that mod is almost certainly impossible. The 3080Ti mod only works with an Eluktronics vBIOS. All other vBIOS fail, including Clevo. Someone briefly listed an MXM clevo 3080Ti for sale on Ebay and Prema somehow identified it as having an Eluktronics vBIOS from the GPU-Z screenshot, which was extremely helpful.
-
I converted a Clevo RTX 3060 to be a 3080Ti: I can't post links because 3dmark blocks viewing scores from unrecognized hardware combinations: The score is a little bit disappointing. A shunt modded 3080 can beat this (just barely though). I did a 50% shuntmod, but this is giving me OCP throttling and I need to lock the core to 0.84V to avoid it, which is only around 180W. Doing a weaker 10% shuntmod should yield a better result to avoid needing to lock the core voltage. I'm also having vRAM overclocking issues. The vRAM can only run up to 16.3GHz before the system hard freezes like an NES. With the same vRAM chips the 3060 core could run 17.7GHz. This requires further investigation. I currently have the card set to be 8GB. I can run more benchmarks if there is something someone is curious in seeing.
-
You should be fine just replacing the blown fet. That one is driven by the main driver. I haven't seen that chip die before.