Author Topic: New PC build p0rn photos.  (Read 6241 times)

0 Members and 1 Guest are viewing this topic.

Offline Howardlong

  • Super Contributor
  • ***
  • Posts: 5424
  • Country: gb
Re: New PC build p0rn photos.
« Reply #50 on: April 18, 2022, 01:03:23 pm »
Server != Workstation != Gaming Desktop.

Very true.

Here's the email machine I built over the past year, RTX 3090 and 5950x.

Useless for server stuff as it only has 32GB RAM. I've found that it does play Flight Simulator in VR reasonably well though.

For my day to day productivity, I run an 11th gen i7 NUC w/64GB RAM & 2TB Gen 4 NVMe SSD.

My laptop of choice is an LG Gram 17 2020 w/40GB RAM 10th Gen i7 with 4TB NVMe SSD, which, as well as having the best laptop screen I've ever encountered, for embedded dev it enjoys great I/O including 3 x USB A 3.1, so no dongle world for me.

For servers, I have two headless dual socket 24c/48t Ivy Bridge era boxes, one with 64GB and the other with 192GB, for spinning up multiple VMs to test out infrastructure projects.

Of interest is that if you spin up a VM in the cloud, there's still a fair bit of Ivy Bridge hanging around, although more modern power efficient stuff is coming along to replace it.

 
The following users thanked this post: paulca

Offline paulcaTopic starter

  • Super Contributor
  • ***
  • Posts: 4362
  • Country: gb
Re: New PC build p0rn photos.
« Reply #51 on: April 19, 2022, 08:59:38 am »
So I finally found the key to the cage they were keeping the CPU horsepower in. 

It's a 3.8Ghz base clock and max boost of 4.8Ghz.  Yes sure I've seen it hit 4.7+Ghz, but when I give it a full 16 threads of prime calculations and FFTs it heats up and drops, eventually to 3.8-3.9Ghz.  Most people would just shrug and accept that.

Nope.  There is the concept of the voltage curves.  Basically as the clock goes up the core needs more voltage to be stable.  So as clocks go up, die power goes up.  There is a hard limiter on 120W total die power.  However my motherboard has awesome VRMs with uber stable voltage at up to 200Amps... so I found the curve offset settings and dropped the whole curve by 30mV.  The furthest it will go without getting nasty with it.

Now when it was pushed hard it didn't reach that power limiter until it was happily boost clocking on all cores to 4.7Ghz.  So having maxed out the benchmark for multi-threading, leaving every other 8 core way, way behind and tripping on the heels of the 16 core threadripper ahead... I moved onto single threading.

Here the magic is in the AMD percision boost overdrive system and... primary max boost clock.  Straight away it was slamming it's two "golden cores" to max boost clock of 4850Mhz and alternating them to prevent the individual cores thermal throttling.  It shuts off all the other cores, leaving the 2 primary single thread cores and 1 core to run everything else.  Doing everything it can to keep that one thread at absolutely max performance.

Of course my cooling system with the -30mV under volt didn't seem all that stressed and the CPU was churning through stress tests and benchmarks, beating even the Intel chips and leading the leader board (on Cinebench).  But I still wanted to crash the thing.  So I went digging for what else I could turn up.

Found the Max boost clock override.  Did a quick bit of maths and said, "Will it do a full 5Ghz?" and set it to +200Mhz and ran the benchmark.  It actually worked with single cores boosting to 5050Mhz.

Then the screens went black and it rebooted.  I was able to complete benchmarks and get scores, but I had another 2 instant resets and another thing morning.  So, it's not stable.  I think I might just take that 200Mhz down to 150Mhz and see.

And, yes, placebo effect or not, I did notice the desktop feel more snappy and boot time come down the first time it booted with the 30mV under volt.  Over all it's about 25% faster over all.  Oddly it's not producing much more heat, the CPU socket power went from ~130W to ~150W which isn't exactly a big leap in thermals.
"What could possibly go wrong?"
Current Open Projects:  STM32F411RE+ESP32+TFT for home IoT (NoT) projects.  Child's advent xmas countdown toy.  Digital audio routing board.
 
The following users thanked this post: cdev

Offline paulcaTopic starter

  • Super Contributor
  • ***
  • Posts: 4362
  • Country: gb
Re: New PC build p0rn photos.
« Reply #52 on: April 19, 2022, 09:12:49 am »
Because it boots and runs single threaded and multi-threaded benchmarks and seems to only crash on the desktop, I suspect the 5050Mhz max boost which is applied to all cores is too much for one (or more) of the lesser cores.  The main "golden cores" handle it fine, but occasionally if one of the lesser cores gets boosted to 5050Mhz it might make an error, if that operation cannot be repipelined or if it create an unrecoverable CPU fault, then it's night night.

So, maybe I should look at per-core max boost clocks and take the lesser 6 cores down 20Mhz at a time until no more resets.

Maybe I could go even higher on the golden-cores.
"What could possibly go wrong?"
Current Open Projects:  STM32F411RE+ESP32+TFT for home IoT (NoT) projects.  Child's advent xmas countdown toy.  Digital audio routing board.
 

Offline Berni

  • Super Contributor
  • ***
  • Posts: 5050
  • Country: si
Re: New PC build p0rn photos.
« Reply #53 on: April 19, 2022, 10:37:08 am »
Yep power is the most important thing in overclocking.

If you can give it more power and cool that power then you can give it more voltage to get higher clocks. But this eventually gets into diminishing returns where it takes more and more power for smaller and smaller gains.

For this reason i don't tend to overclock any of my CPUs. The only thing i do is crank up the power limit to something like 150W in the bios to just let the CPU pull as much juice as it wants to, usually this makes the CPU run at max boost clock on all cores when under heavy load, but has no effect when under light load. It doesn't tend to cause any instability (unless the motherboard can't keep up with delivering this much power) Thing is not all benchmarks make the CPU produce the same amount of heat even tho they all hold it pegged at 100% utilization. Some instructions or combinations of instructions take more power to execute, this is especially apparent for the new AVX512 instructions that do a massive amount of math in parallel.

As for overclocks crashing during light load but stable during a benchmark, this is typically the fault of the CPU pulling too sharp of a power spike (when switching clock speed), making the core voltage regulation too slow to respond. This is why huge overclocks tend to lock the CPU freqency to be max all the time, but this is obviously a very bad idea for every day use since the CPU will be burning tons of power even when just sitting on a desktop. Not like you are going to notice the slight performance gain.
 

Offline PlainName

  • Super Contributor
  • ***
  • Posts: 7508
  • Country: va
Re: New PC build p0rn photos.
« Reply #54 on: April 19, 2022, 05:20:39 pm »
Quote
Useless for server stuff as it only has 32GB RAM

!

!!

My mail and DNS (and some other stuff) server has 3GB memory. Surely it depends on what the server's function is - 'server' covers as many variations as 'vehicle', and a 5-litre turbo-charged V12 isn't necessary for mowing the lawn (although it could be fun).
 
The following users thanked this post: Howardlong

Offline Howardlong

  • Super Contributor
  • ***
  • Posts: 5424
  • Country: gb
Re: New PC build p0rn photos.
« Reply #55 on: April 19, 2022, 06:24:22 pm »
Quote
Useless for server stuff as it only has 32GB RAM

!

!!

My mail and DNS (and some other stuff) server has 3GB memory. Surely it depends on what the server's function is - 'server' covers as many variations as 'vehicle', and a 5-litre turbo-charged V12 isn't necessary for mowing the lawn (although it could be fun).

As already explained, I use physical servers to run multiple VMs, a couple of dozen on a single physical machine is not unusual when you're testing a typical distributed application. Hope that explains it!
 

Offline PlainName

  • Super Contributor
  • ***
  • Posts: 7508
  • Country: va
Re: New PC build p0rn photos.
« Reply #56 on: April 19, 2022, 06:28:21 pm »
Quote
Hope that explains it!

Sure, sorry for missing the earlier bit.
 
The following users thanked this post: Howardlong

Offline free_electron

  • Super Contributor
  • ***
  • Posts: 8550
  • Country: us
    • SiliconValleyGarage
Re: New PC build p0rn photos.
« Reply #57 on: April 20, 2022, 01:08:04 am »
and a 5-litre turbo-charged V12 isn't necessary for mowing the lawn (although it could be fun).
Professional Electron Wrangler.
Any comments, or points of view expressed, are my own and not endorsed , induced or compensated by my employer(s).
 
The following users thanked this post: PlainName

Offline PlainName

  • Super Contributor
  • ***
  • Posts: 7508
  • Country: va
Re: New PC build p0rn photos.
« Reply #58 on: April 20, 2022, 01:47:46 am »
Gosh, don't they look young. And slim!
 

Offline paulcaTopic starter

  • Super Contributor
  • ***
  • Posts: 4362
  • Country: gb
Re: New PC build p0rn photos.
« Reply #59 on: April 20, 2022, 08:36:17 am »
I spent 2 hours trying to over clock the 3080 GPU.  I could get a little here and there, but nothing substantial.

It's a bit like riding a bull.  It has a mind of it's own.  Primarily the effective limiter is the power limiter. 

On default config, when loaded to 100%, it initially overshoots to 2000Mhz, slams into the power limiter and backs off to around 1860Mhz.  It fluctuates a lot as obviously slightly differing loads, different power requirements etc.  Maybe peaks at 1900Mhz and averages around 1880Mhz.

Now I can forcefully add 200Mhz to whatever it suggests and this is where it gets really odd.  It will initially spike, again to 2000Mhz and again it will fall back to 1860Mhz.  Yet if you let it run the full benchmark it produces a slightly, 2-3% higher score.

So digging a bit deeper the board (or driver) maintains a table of clocks per volt.  The driver basically runs up that table taking the clock and thus voltage higher and higher until it hits one of it's limiters.  Most likely the power limiter.  If it hits one of those limiters it steps back one step on the table and sees if that is within limits.

The software lets me lock the clock to a point on that curve, so I painstakingly tried to find it's maximum clock per watt.  Pushing the clocks higher at lower voltages until it crashed.  After an hour moving about 10 points on the graph a dozen times each, until it crashed and then back off and change to a different voltage....

I could get it to run stable at 1920Mhz at 925mV.  However, then it threw another twist at me.  Running a benchmark and it just undid my "clock lock".  Basically the 1920@925 hit the power limiter under benchmark, so it moved "left" in voltage and dropped the clocks back to around 1860Mhz again.  Grrr.

Another hour moving key voltage milestones up and down, I could find it's most efficient curve point with a slight under-volt and setting that as the TOP of the curve, resulting in slightly slower peak clocks, but much less power, only 260W and no limiters.  Which is nice and an option to keep as a profile I think for power/heat efficiency.

What I determined after these hours I won't get back... is the driver has done this all before and while I might squeeze a 100Mhz here or there on the curve, that damn power limiter will get me nearly every time.  Seeking peek clocks with that hard power limit, is a fools errand.    I even ran the MSI scanner which does similar, but much more thorough and it came back with it's best of +50Mhz Core and +100Mhz RAM... but said it was "unstable".

So why does it benchmark higher when you set a higher peak clock, even if it doesn't appear to use that peak clock?  I think this comes down to the slight under volting the +200Mhz actually relates to, while watched on a display with an update of 1 second or so, I expect I am not seeing the faster adjustments and with the given power limit, the voltage average per Mhz being a few mV lower is actually getting a few dozen more MHz here and there throughout the benchmark.

Went off to play one of the latest triple A titles - Far Cry 6.  With V-Sync set to my monitor refresh of 75Hz... the GPU, without any overclock, wasn't even hitting 100% usage.  The fans barely came on.
"What could possibly go wrong?"
Current Open Projects:  STM32F411RE+ESP32+TFT for home IoT (NoT) projects.  Child's advent xmas countdown toy.  Digital audio routing board.
 

Offline paulcaTopic starter

  • Super Contributor
  • ***
  • Posts: 4362
  • Country: gb
Re: New PC build p0rn photos.
« Reply #60 on: April 20, 2022, 08:50:21 am »
A thought provoking thing is that pulling 320W at 900mV is about 350Amps.  The CPU is 140A EDC on the dies and 160A total.  The motherboard can supply it with 200A.

But that GPU figure of 350 Amps is insane.  I've not seen a pin map of it, but I'm going to go out on a limb and suggest that as much as 50% of them are going to be 1.5V rail and ground.  There is no way they are push 350Amps through a bit of silicon on 7nm process.  I expect that power enters the die in a widely parallel fashion spread out across the whole die and isn't "transported" around the chip unless absolutely necessary, so each little 5mm square area is only pulling a few amps locally and only a few dozen amps on the nearest Vcore pin.

What do you think?

I know power is generated from multiphased VRMs, so each VRM is only generating a burst of high current on a duty cycle.  I assume again that those VRMs output along many parallel tracks which cover the shortest possible distance to a Vcore pin.

Of course the main issue with amps is heat.  But given these things have coolers capable of handling 500+W, certainly for the die and VRMs, there should be little concern of it melting.

The mobo has (IIRC) 12+4 VRM phases, which I believe is 12 CPU Core VRMs and 4 for the rest.  I didn't count the GPU VRMs when I had it apart, but there are two solid lines of Vcore VRMs either side of the chip socket.  The modded cooler has heat transfer pads for both rows, front and back side of the board, in addition to the VRAM modules and of course the GPU die.  It must be 8 each side?  10 each side?
« Last Edit: April 20, 2022, 08:52:34 am by paulca »
"What could possibly go wrong?"
Current Open Projects:  STM32F411RE+ESP32+TFT for home IoT (NoT) projects.  Child's advent xmas countdown toy.  Digital audio routing board.
 

Offline Howardlong

  • Super Contributor
  • ***
  • Posts: 5424
  • Country: gb
Re: New PC build p0rn photos.
« Reply #61 on: April 20, 2022, 11:48:31 am »
I spent 2 hours trying to over clock the 3080 GPU.  I could get a little here and there, but nothing substantial.


The overclocking thing is a sure fire way to lose much of your life!

I had a reasonable uptick in Ethereum hashrate on my 3090 using a static config, but it's a completely different config on the card for gaming, where I used a voltage/frequency curve.

Overclocking the 5950x took several weeks elapsed delving into the per core curve optimiser to get it 100% stable, tweaking each core until none crashed anymore.

I'm still not really sure I've ever benefitted that much from the tweaking.
 

Offline Ranayna

  • Frequent Contributor
  • **
  • Posts: 986
  • Country: de
Re: New PC build p0rn photos.
« Reply #62 on: April 20, 2022, 12:56:05 pm »
Modern hardware, both CPUs and GPUs, are already near their limits.
These beasts gobble up power like no tomorrow, for at best marginal performance increases while gaming.

You should look at undervolting to reduce power consumption for little to no actual performance loss, outside of artificial benchmarks.
Some cards can actually increase in performance when undervolted.
Or even actual underclocking, to further reduce power consumption with a little more performance loss, depending on the workload.

 

Offline hans

  • Super Contributor
  • ***
  • Posts: 1698
  • Country: nl
Re: New PC build p0rn photos.
« Reply #63 on: April 20, 2022, 02:22:53 pm »
Exactly. On my server I run a 5600G processor. In the interest of efficiency (it's a server - not a desktop), I've chosen to reduce the package power to 35W. These CPU dies are derivatives of their mobile SOCs, so the BIOS may support multiple TDP profiles. In my testing it yielded only 10-15% lower Cinebench scores, but *halved* power consumption (went from 130W system load to about 60-65W, around 12W idle). Since my server will transcode the odd movie for storage, this significantly cuts back on the energy cost for that job.

In my desktop I run a 3900X, and never gained anything by tweaking the settings. Out of the box it boosts 4GHz+ all core at around 80-85C under a big tower cooler. I've tried a slight undervolt, enabling PBO, etc. but I was not very impressed by what any of these did. Either the CPU became unstable for the same performance, or ran extremely hot (90C+ with PBO) for a 0.5% performance bump.

And I haven't even looked at the 3080 card I've got. The release with unstable BIOS's from some manufacturers (and the red herring that looked at a bunch of capacitors) didn't warrant me investigating any further.
 

Offline paulcaTopic starter

  • Super Contributor
  • ***
  • Posts: 4362
  • Country: gb
Re: New PC build p0rn photos.
« Reply #64 on: April 20, 2022, 05:37:08 pm »
It amusing how many people don't "like" PBO. Some people say it idles at 1.5V which they don't like so they don't use it.  Others say it runs the system hotter than it should.  Both non-sense according to AMD, the chips are fine to 105 and will shutdown at 95-100 anyway.  Also  reading into PBO it is a complicated beast doing a lot more than most overclockers can, but doing it 100 times a second.  I mean at very least there will be times when a core won't got to 4.xGhz and a static overclock will become unstable.  A PBO one won't, because PBO will know to back off a bit.

The thing about PBO is, it will save you power, it turns core OFF when it needs to.  It undervolts and underclocks them when they are no being used much.  The 3400X will boost higher than 4Ghz.  It's clock boost is 4.3Ghz and it will go higher than that.

The other thing about PBO is that motherboard manufacturers often ship with "their" tweaked/optimized versions of the control parameters, which may not be ideal or may be too unstable.
"What could possibly go wrong?"
Current Open Projects:  STM32F411RE+ESP32+TFT for home IoT (NoT) projects.  Child's advent xmas countdown toy.  Digital audio routing board.
 

Offline paulcaTopic starter

  • Super Contributor
  • ***
  • Posts: 4362
  • Country: gb
Re: New PC build p0rn photos.
« Reply #65 on: April 20, 2022, 05:44:23 pm »
I suppose it's a good question.  Does it actually do any harm.

Leaving aside the concerns that overclocking boards like the motherboard I have, allow you to play with the dangerous settings and could probably override the CPUs protections enough to cook it fairly quickly.

Running the chips hard on high power does apparently degrade the silicon faster.  So maybe the chip might last 12 years instead of 15.  Shrug.  Who will want a 5800X in 12 years?

The other issue gamers face is that different overclocks (especially manual ones) are prone to seasonal issues, warmer in summer etc.  That silicon degradation can also bite when the chip or some tiny part of it will not longer clock or volt as high as before.

I normally find the peak overclock, get a few benchmarks and then reduce it down to something much more sane and stable for daily use.

I have set an undervolt/underpower profile, which as suggested only loses about 1% performance for a reduction on 50W of power.
"What could possibly go wrong?"
Current Open Projects:  STM32F411RE+ESP32+TFT for home IoT (NoT) projects.  Child's advent xmas countdown toy.  Digital audio routing board.
 


Share me

Digg  Facebook  SlashDot  Delicious  Technorati  Twitter  Google  Yahoo
Smf