360mm good quality AIO vs custom loop for ~250-300W CPU

[User5566]

So, I've made an Alphacool order through a local distributor. They're planning a production run for the 20th of Nov, but the distributor told me it is by no means certain they will make enough of these blocks to cover whatever orders they might have already. So he's asking for confirmation of allocation and hopefully I'll find out on Monday if these are truly available at that price.

Regarding the cpu difference in temperature at multicore load I found lots of posts online about it. It is not grounds for RMA unless the chip underperformed by a lot. And to make sure it underperforms I need a fresh install of Windows... Which I wasn't planning on, but I might do.

To be fair the first time I noticed it I was already running a curve optimizer negative setting of - 15 on all cores. IDK if it happens at stock settings but I'll check today.

If this happens on stock I'll post hwinfo logs, but from what I saw so far (with all cores in question set to - 25). I looked at 4 cores. One is the highest temp, one the lowest and 2 in the middle. They all get the same voltage according to hwinfo and the same frequency. And yet at the time the hottest core was showing almost 8W ofpower consumption, the lowest one 6.5, and two remaining ones around 7W. Perhaps it's some artefact of curve optimizer use?

 

[Woomack]

The local ones buy from the Alphacool store (just a business channel), so it takes longer and usually costs more as local stores have to add their margin Usually, it's not a big difference, but I don't know what price you got. The most important is that it's on the way.
Alphacool graphics card blocks are pretty good quality, so I'm sure you will be satisfied. The last one I got from them was for RX6800XT, and after a year it's clean with no stains or other issues.

Re CPU, I know that Intel was replacing chips that had over 15°C difference between cores. I don't remember any RMA case with AMD, so it's hard to say. CPUs, in general, have a very low RMA rate.

 

[BugFreak]

I use an Alphacool block on my 3080 with no problems. It keeps the card cool and still looks brand new after over a year. You will like it.

 

[User5566]

The local ones buy from the Alphacool store (just a business channel), so it takes longer and usually costs more as local stores have to add their margin Usually, it's not a big difference, but I don't know what price you got. The most important is that it's on the way.

For me the main difference is if I buy directly from Alphacool I have to pay first and as this is made to order there is no guarantee of availability.

Buying from a local supplier I pay €7 more per block but I pay on delivery when(and if) that delivery happens. Alphacool has payment on delivery within Germany, but sadly not across the border in Poland, where I'm at.

In the unlikely scenario these blocks do not materialise I'll order Bykski blocks which are available "in 5 days", but are €25 more...

Alphacool graphics card blocks are pretty good quality, so I'm sure you will be satisfied. The last one I got from them was for RX6800XT, and after a year it's clean with no stains or other issues.

Re CPU, I know that Intel was replacing chips that had over 15°C difference between cores. I don't remember any RMA case with AMD, so it's hard to say. CPUs, in general, have a very low RMA rate.

I've ran it with stock settings and the biggest difference between cores is well over 15C. But this is AMD, not Intel. And until this impacts performance I can't complain especially if I can level it out with curve optimiser.

Interestingly when I put core temps on an excel chart I saw all 8 cores in the first die are approximately 10C hotter and are running 200mhz faster, all the cores in the second die are cooler and are running slower. Then again, this is an ancient win 10 install that had lots of crap installed on it over the years. I'm currently trying to get the per core temperature readout working in Linux to compare.

Just by using y-cruncher in Linux and windows I get about 5% better performance in Linux. But tgats all I can say for now.

 

[User5566]

I can confirm the exact same thing happens with stock settings. I even updated the bios, but it didn't make much difference. I'll post the hwinfo log later.

Unfortunately, there is no way to get per core temps in Linux with this cpu. It uses too new smu version (best bet was a project called ryzen-smu patched to support 7950x, this too isn't working on my hw version). I find it incredible amd can't just publish the specs how to get that power/temp info out. Authors of software like hwinfo have to rely on reverse engineering things like ryzen master to get the data.

Also, my r23 scores are a bit lower than expected, at the moment I can't tell if the reason is a clogged aio, old windows install or some actual problem with the cpu I highly doubt it's a cpu problem as my 5950x also scored less than expected so this points to windows.

So I have a quick question. When benchmarking with things like Cinebench r23, does it matter on the scores if one is using the igpu? I have two nvidia cards, but Im using them just for compute and my monitor it connected to the motherboard instead using the igpu. Can this cause lower scores on stock settings(everything on Auto in bios)?

 

[EarthDog]

How much lower is 'a bit'?

Also, I have a big AIO on it (3x140mm Liquid Freezer II)... so that may be why. The scores at Tom's Hardware are a little less than mine...37.9k for R23 and 2055 for R24.

What's your single-thread score? So long as it's not using your 'offset' core, heat shouldn't be an issue and should be close to others. RAM speed doesn't make too much of a difference in CB, but I'm also at DDR5-6000 CL30 so that may be it.

Intel Core i9-14900K, i7-14700K and i5-14600K Review: Ryzen X3D Stays On Top

Mid-range gaming has a new champion.

www.tomshardware.com

 

[User5566]

How much lower is 'a bit'?

Barely getting 36k when I kill all sorts of crap running in the background in Windows(corsair service, Dropbox sync, Asus updater and 10 other things).

I was pretty sure that was an old Windows's install fault, but then I got Cinebench r23 working under Linux (with wine) and I got the same 36k.

This all on stock, no expo, ram at 4800MT. Bios all on Auto. No PBO2, no curve optimizer undervolt. I'm trying to get an "as advertised" baseline first.

I suspect the default BIOS settings are far from optimal as it's pulling "only" 167W tops while I see it on various benchmark videos people talking about 230W power consumption. I'm not quite sure how's that possible with no pbo as stock TDP is 170W.

Also,no one ever mentions specific voltages resulting from bios settings. For example in my MB the Auto SOC setting always sets it to about 1.018 (yes, that zero is not in there by mistake, its barely over 1V).

Also, I have a big AIO on it (3x140mm Liquid Freezer II)... so that may be why. The scores at Tom's Hardware are a little less than mine...37.9k for R23 and 2055 for R24.

What's your single-thread score? So long as it's not using your 'offset' core, heat shouldn't be an issue and should be close to others. RAM speed doesn't make too much of a difference in CB, but I'm also at DDR5-6000 CL30 so that may be it.

Single core is at 1971 at 64C, so about 100 too low too. This is with the core boosting to about 5.6ghz and at ~70W of power.

Also, I tried to find any youtube videos where anyone does stock settings 7950x multicore Cinebench run with individual core temps showing.... There are none. The closest I got was at the beginning of the der8auer video with dry ice. There one can see his ccd1 and ccd2 temps and one is 89.9, the other is 80.9 so almost 10C there.

If anyone here has got a 7950x, that scores 38k r23 points multicore and 2070 single stock can you post your core temperatures and all the voltages and BIOS settings? (regardless of cooling tech).

 

[EarthDog]

If anyone here has got a 7950x, that scores 38k r23 points multicore and 2070 single stock can you post your core temperatures and all the voltages and BIOS settings? (regardless of cooling tech).

I got you... give me a night or two to fire that rig up and I'll get a score in. This would be with a 3x120mm AIO though as the other system is in use/testing. W11 22H2.

 

[User5566]

I got you... give me a night or two to fire that rig up and I'll get a score in. This would be with a 3x120mm AIO though as the other system is in use/testing. W11 22H2.

That would be great to compare more parameters

I mentioned before I saw the same inter-core temperature differences on stock settings (bios completely wiped and updated, RAM at 4800MT/s, no OC, no PBO2 limit overrides etc). However, after looking at this more I'm inclined to believe it might be normal as the die to die temperature difference is maybe ~7C and that looks the same as for example in the der8auer video I mentioned.

Still I attach the hwinfo log. Perhaps someone can deduct more from it.

Also, I've made two interesting observations. First, if I set curve optimizer and I set PBO2 limits higher (to about 350/200/250) I get about 37.9k r23 points which is very close to the advertised 38k. If I have time I might do a new install of windows 11 to to test (this is captured on win 10), but I have a feeling it will not matter much as I disabled autostart of anything in my current win10 install.

One more thing that made me waste 1h of time on troubleshooting. If you have a Ryzen 9 7950X and suddenly it starts "boosting" to 545 MHz tops.... please check your PPT/TDC/EDC are not actually specified in mW and mA in on the other page in the BIOS Yes, that's true. After a BIOS upgrade I now noticed my PBO2 limits in Advanced->AMD Overclocking->Precision Boost Overdrive are in mW so if you want a PPT of 350W you enter 350000. No it doesn't have anything to do with my otherwise too low scores, but I thought it will amuse someone to learn this.

And final thing I noticed... Everyone knows for some strange reason PBO2 limits are in two places in the BIOS and the settings are not equivalent in practice. Now we have a third place to set the same settings... In addition to AI Overclocking->PBO, and Advanced->AMD Overclocking->PBO, we also have Advanced->AMD CMS->SMU Settings and there all the PBO and thermal limits are replicated for the third time. Of course none of these 3 places are in sync and you can set each one of them to a different value if you get confused Anyway, hopefully this info helps someone.

 

[User5566]

I might be exchanging that cpu for another one. I've done a fresh windows 11 install. It is the same. I can hit 37073 only with curve optimiser.

Also I've done testing with y-Cruncher and I'm getting very different results for 3 of the cores (the worst for c7).

For example (both c7 and c8 are set to - 25 on Co, pbo2 limits on default).

Running single core y-Cruncher 5mln pi digit benchmark these are some numbers
C7
Freq 5500, temp 82C, total power 65W, time 72.4s
C8
Freq 5500, temp 64C,total power 85W, time 71.6

So not only c8 (and any other core than c7, C5 and c14) is very slightly faster it does so while pulling 20W more power at almost 20C less temperature!

This looks more and more like some manufacturing issue.

And I can repeat it all day long, also I disabled built in gpu to exclude it.

Edit: the power and voltage consumption difference is because c7 is on ccd0 while c8 is on ccd1 which requires 0.1v more to get to the same freq. But it doesn't explain the almost 20c higher temps (in single core workload) on c7. In comparison with c0-c6 c7s power/voltage is the same, but temperature is even higher. C0-C6 run the same test at 57C.

Edit2: I'm pretty sure the cpu is a dud. I wonder how many of them are there exactly like that and no one notices, because normal users don't actuverify promised performance.

How do I know it's a bad cpu? I tried the "dynamic OC switcher" feature. I set the current threshold at 90a, voltage at 1.25 (well below the max, or 1.3 "ai overclocker recommended") and multipliers at 52 and 50.

When I run Cinebench r23 core 7 went to 107C while all the others sat on 80C,all while the cpu consumed 190W. This is insane. I've already ordered another one and I'll be returning this one. I hope the other one will score 38k points at stock settings.

 

[EarthDog]

Here's my results.........

R23
Multi - 38,697 @ 218W, 85-90C
Single - 1,880 @ 16W, 65C

R24
Multi - 1,997 @ 211W, 82-88C
Single - 112 @ 15W, 65C

Scores above are on Gigabyte X670E Extreme with memory at DDR5-4800 (JEDEC).

Edit2: I'm pretty sure the cpu is a dud. I wonder how many of them are there exactly like that and no one notices, because normal users don't actuverify promised performance.

can't say I'm convinced... it could just be how your motherboard runs the chip. Many are different. I guess we'll see. What motherboard do you have??? Can you create a signature that lists your hardware so we know what you're working with, exactly?

As an FYI, out of 17 boards I've tested R23 on, the scores range from 34k (cheap B650M boards) to 38.7k (up to flagship-class boards) with this same 7950X, 3x120mm AIO, and RAM (DDR5-5600 CL30). Single core ranges from 1,854 to 2,038.


EDIT: I edited quite a bit above, FYI.

 

[User5566]

Here's my results.........

R23
Multi - 38,697 @ 218W, 85-90C
Single - 1,880 @ 16W, 65C

R24
Multi - 1,997 @ 211W, 82-88C
Single - 112 @ 15W, 65C

Thanks for checking, what were the core temperatures? (I mean individual cores)?

can't say I'm convinced... it could just be how your motherboard runs the chip. Many are different. I guess we'll see.

As an FYI, out of 17 boards I've tested R23 on, the scores range from 34k (cheap B650M boards) to 38.7k (up to flagship-class boards) with this same 7950X, 3x120mm AIO, and RAM. Single core ranges from 1,854 to 2,038.

But did any had a 30C difference between the coolest and hottest core on the same ccd0? That is the question? For example in r23 at 90C tclk the coolest core would have to be 60C to be as bad as my unit. And the power.... You have 218W. There is no way in hell my cpu ever exceeds 180W ever, no matter what pbo limits are set. Once that one core heats up the rest of the cpu gets throttled.

I doubt it is the motherboard. It's one of the ASUS's "flagship" products. Pro art X670e-creator wifi.

Also in a video published yesterday der8auer showed his per core temps while testing two different air coolers on an Intel 14900k (of course it's a different cpu, but the general principle is the same - a very powerful, energy wise, set of cores in a very small package). In all cases his cpu never exceed 6C between the hottest and the coolest p core.

So I'm 99% convinced a dud. Hopefully I'll have another one tomorrow.

 

[EarthDog]

Thanks for checking, what were the core temperatures? (I mean individual cores)?

Individual cores are covered in the range I listed...I didn't see more than a 6C difference in this short testing. If you let it run longer, do the cores balance out?

ASUS's "flagship" products. Pro art X670e-creator wifi.

Most would consider this a mid-range solution. It's made for Creators so it has good connectivity (TB4, 10 GbE, etc.) for that/high-end, but what matters in this case, power delivery, is average with 16 phases 70A. Still, it will do the job. In fact, one of the 17 boards I mentioned that was the ProArt you have. It scored 37,568 and 2,007, respectively (with DDR5-5600 CL30). What do you score when you put decent memory in and not use JEDEC speeds? All the scores I listed use 5600.

(of course it's a different cpu,

This. Not the rest. Can't compare unlike things (p/e cores are dramatically different than AMD, for example).

But the temp spread, I hear ya.......... I think we (Woomack) mentioned Friday to reach out to AMD to ask if it's a problem/see if you can RMA. But again, you seem to be in range with my results (the range of 17 mobos), no?

 

[User5566]

Individual cores are covered in the range I listed...I didn't see more than a 6C difference in this short testing. If you let it run longer, do the cores balance out?

Sorry, I somehow missed it. Indeed they are. Great.

Most would consider this a mid-range solution. It's made for Creators so it has good connectivity (TB4, 10 GbE, etc.) for that/high-end, but what matters in this case, power delivery, is average with 16 phases 70A. Still, it will do the job. In fact, one of the 17 boards I mentioned that was the ProArt you have. It scored 37,568 and 2,007, respectively (with DDR5-5600 CL30). What do you score when you put decent memory in and not use JEDEC speeds? All the scores I listed use 5600.

This is good to know (about the score).

Well, I've been using the same RAM all the time (I just got the new ddr5 2x48gb 6400 cl32 kit today, I haven't unpacked it yet). I was just running JEDEC to exclude the possibility of ram oc causing any issues.

The first few tries I run it at 6000MT cl30 and I didn't see much change in the Cinebench r23 score, (maybe a hundred points - I'd have written it down if it was more) . This probably shows even more where the bottleneck is.

This. Not the rest. Can't compare unlike things (p/e cores are dramatically different than AMD, for example).

But, your own test had an almost identical spread (5C vs 6C), but yours is even better because in the video I only considered 8 p cores (his e cores were more like 12C away) while in your test you had 5C amongst 16 cores!

But the temp spread, I hear ya.......... I think we (Woomack) mentioned Friday to reach out to AMD to ask if it's a problem/see if you can RMA. But again, you seem to be in range with my results (the range of 17 mobos), no?

I haven't reached out to amd yet, because I wanted to gather more data. I don't really care about RMA, because I still have about 7 days on the "return no questions asked" policy. The reason I'm trying to find how prevalent this is is because I wonder if I made a good choice going with AMD. I really could use the extra cache 7950x has over 13900k/14900k, and for my workloads 16 even (more or less) cores are much better than 8 fast, 16 slow, but if I have to buy 3 7950xs and return 2 to get one that performs as advertised my favourite shop will not want to do business with me anymore

Also If this is an actual failure I need to tell them when returning it so they don't sell it to some person who was saving for a year to get a sub par cpu.. Although, they'll probably put it in a MB, see it boots they'll sell it anyway.

About this testing with the 17 mobos, this is very interesting, do you by any chance remember what was the core to core temp variations in these MB that scored very low? It would be very interesting if the MB affected it. Sadly, that's the only am5 MB I know with 10gig, and 3 pcie slots 2 of which have the correct spacing to use the 3 space nvlink connection so unless I switch to Intel... I'm stuck on it.

Also, was all the difference basically between chipsets (670 vs b650), or was there large differences within the same chipset? I'm trying to identify if this was due to the different buses, chipsets, defaults, or all 3...

 

[EarthDog]

The reason I'm trying to find how prevalent this is is because I wonder if I made a good choice going with AMD.

You're really swimming upstream in the minutia here. I can't imagine this is some intrinsic problem with the chips. You can google ANYTHING and get hits. Honestly, I'd worry about my chip only.

About this testing with the 17 mobos, this is very interesting, do you by any chance remember what was the core to core temp variations in these MB that scored very low?

No. but I could pull data for how it ran during the stress test (I don't run Hwinfo when benchmarking as that throws the scores off, lower, too - have you always been running with it active?). But it's the same chip, so I'd imagine similar temps among the cores.

Also, was all the difference basically between chipsets (670 vs b650), or was there large differences within the same chipset? I'm trying to identify if this was due to the different buses, chipsets, defaults, or all 3...

It's the BIOS on the boards. Low end, cheap arse boards limit things more than high-end boards that let processors run free. AMD is different than Intel in that manner though. It's a mix, but obviously cheap arse A620 boards are towards the bottom/slowest versus X670E boards.... but there are exceptions! Mount matters, for example! Why would X670E Taichi be so low when it runs full tilt? It's going to be tough to pin it down. I'd just RMA the bad chip when you find a 'good one' whatever that means.

....try assigning the ST process to a core that doesn't overheat and see if those change.

 

[User5566]

You're really swimming upstream in the minutia here. I can't imagine this is some intrinsic problem with the chips. You can google ANYTHING and get hits. Honestly, I'd worry about my chip only.

if the second chip runs fine that will be the end of it. If not.... Well I need a plan for that.

No. but I could pull data for how it ran during the stress test (I don't run Hwinfo when benchmarking as that throws the scores off, lower, too - have you always been running with it active?). But it's the same chip, so I'd imagine similar temps among the cores.

If you're saying you can, please do I'd love to find out. The way these platforms fail to deliver is quite interesting. Also the wattage pulled is very important

It's the BIOS on the boards. Low end, cheap arse boards limit things more than high-end boards that let processors run free.

Can you elaborate on this? I suspect you mean things like more beefy vrm can provide bigger power spikes when requested rather than have voltage sag. I can definitely see how this can affect things a lot in very spikey workloads like gaming, but rendering where there is long term sustained workload? (what Cinebench r23 does)

But BIOS?

AMD is different than Intel in that manner though. It's a mix, but obviously cheap arse A620 boards are towards the bottom/slowest versus X670E boards.... but there are exceptions! Mount matters, for example! Why would X670E Taichi be so low when it runs full tilt? It's going to be tough to pin it down. I'd just RMA the bad chip when you find a 'good one' whatever that means.

"a good chip" in this context means all cores stay within 10C of each other and scores 38k in Cinebench r23 on stock settings without upping pbo limits(within +-500 points). Also one that actually responds to pbo limit increase by increasing power draw.

Also flatness of the cold plate. If you want to get into the woods a bit.. One of my hobbies is high precision mechanics. I have equipment to produce and measure surfaces flat down to 150nm across the width of an ihs as a matter of routine. If I really spend quality time on it I can get down to 50nm, but even to achieve 150nm one has to control heat expansion from one's own fingers affecting the shape of material when touching it. A couple of days ago I evaluated the flatness of my AIO cold plate. Long story short it was far from flat (the total deviation from flatness was 8 microns when I checked). The way one can visualise such deviation at an amateur workshop is with precision ground flat stones (that's a technical term not a description). The cold plate was contacting directly only on two small spots. Then I've used a manual lapping technique on a granite precision surface plate to flatten it with 800 grit sandpaper (one has to be skilled in using it to achieve flatness), then 1500 grit, then I restored the surface roughness to its previous reflectivity with precision ground flat stones. I didn't have much time so I finished at 1 micron deviation overall the entire plate, but the central area where the ihs touches is no worse than half a micron.

I cannot wait to measure my new waterblock when it arrives too (in a couple of weeks I'll be making a custom water loop for this pc). I can't lap that one as it's nickel plated and I don't want to loose the plating unless it's really bad in which case I might lap it and replate afterwards.

Does a difference between a couple of microns really affect things that much? Probably not, but it's a hobby right. So I'm going the extra mile.

....try assigning the ST process to a core that doesn't overheat and see if those change.

Of course I've done that a number of times. I believe I described it above. I run the same ycruncher single core task one any other core on ccd0 and completes in 71.4s (from memory) while the core is ~62C, try that on core 7, it is ~83C and completes in 72.8s.

There is no point doing that with Cinebench r23 single core because it will switch core affinity however it likes at the beginning of the test (at least that's what it did when I tried).

View attachment 363644

View attachment 363645

Thanks. It is an interesting chart. It is also most interesting none of the X670e boards tested were worse than 1k points away from my target of 38k. That's kind of reassuring.

Edit: regarding running hwinfo while benchmarking, I found about 150 points of difference when set at a polling rate of 2s,logging and having a large 2 pane window open.

Edit2: I wish boards like Asrock X670e taichi would come with 10gig. I could live without the third pcie slot, but not without 10gig (with 2 pcie slots otherwise occupied)

 

[EarthDog]

The way these platforms fail to deliver is quite interesting.

That's an interesting take I must admit. They're doing their job to me. To be clear, when I say limit, it's to the processors specified TDP and power/thermal limits.

These are great reads:

Why Intel Processors Draw More Power Than Expected: TDP and Turbo Explained

www.anandtech.com

AMD Zen 3 Ryzen Deep Dive Review: 5950X, 5900X, 5800X and 5600X Tested

www.anandtech.com

Reaching for Turbo: Aligning Perception with AMD’s Frequency Metrics

www.anandtech.com

(on the first link are a couple of other articles relating to these TDP and power limits/how the processors work)

In the Intel (and AMD) world, motherboard vendors take liberties with the specs of the processor. In a lot of cases, when you set the cooling type in the BIOS (if offered) that raises/lowers the limit. For example, if you choose 'air,' it may go to Intel defaults and follow that explicitly. Whereas if you choose 360mm AIO, the power limits are raised, so the limit is, essentially, going to be cooling on the CPU for any modern processor (whose board doesn't limit the chip - follow its own specs - by default).

As far as flatness...at one point, the IHS was concave, and coolers that were slightly convex worked best. If it's flat, then flat should work best for all other variables the same.

There is no point doing that with Cinebench r23 single core because it will switch core affinity

That's why I said to assign it a core. But yeah, it's the same difference in the y-cruncher testing you already did.

Edit: regarding running hwinfo while benchmarking, I found about 150 points of difference when set at a polling rate of 2s,logging and having a large 2 pane window open.

Mine is every second... typically worth a couple of hundred on average. Not your make or break...just making sure we're comparing apples to apples.

EDIT: @User5566 - After looking through a few of the data for stress testing (AIDA64, CPU/FPU/Cache/Memory), the cores averaged around 7-9C different. Power use varied. One board was 186W, another 165-170. The X670E Creator with w/e BIOS was used back in July and was around 180-185W. the cheapo A620 board........~155W. Core temp range was the same as the higher wattage boards as I would have expected.

 

[User5566]

Well, I just pulled the new cpu out if the box, I loaded "optimised defaults" in the bios which is stock settings and on the very first try (no curve optimizer, no pbo limits raised) I had the difference between cores under 10C, power consumption 230W not 165W like previously and ~36800 r23 points with hwinfo polling every 0.5s. The score is not my expected 38k,but at last I see potential in this For the comparison sake, my very first score on the previous cpu was 34800k.

More testing to come.

Edit: Without hwinfo running, 37200 r23 points, still on bios defaults.

 

[User5566]

So, I'm pretty happy with the new cpu. It's not a golden sample, but compared with the previous one it's a huge step up. Curve Optimizer undervolts are (negative) 10,10,8, 2,10,3,10, 6,20, 30,25,27,25, 27,30, 22. Upping the pbo limits doesn't make such a huge difference on this cpu as it did on 5950x, but between it and curve optimizer I get solid ~38300 Cinebench r32 points.

If I lock the multipliers I can get as far as 5.3Ghz @ 1.2V getting 39571 r23 points at that. Temperatures are 93 and 91.5C (for each ccd). Power is 209W. Unfortunately most my workloads use avx512 a lot so I can't just lock it like that, because prime95 would cook that cpu at these settings in 20s. I haven't found what are the fastest locked multpier settings usable with prime95 yet as I intend to leave it on pbo2 until I have better cooling.

BTW all the above benchmarks were done with memory on 5300MT/s so I'm likely to get slightly better scores (I hope) with it at 6000MT.

So thanks to everyone who helped me in this troubleshooting