Log Walrus WS — 2.6L Ryzen workstation with RTX A2000 and optimized cooling

[Valantar]

Since we are both trying infill as mesh and trying to make it look as good as possible, just reporting I have had a bit better success by turn on Combing, Retraction, and being ready to pause the print on the first layer if there's any stray fuzz that gets layed down before the pattern is layed. I have had a few prints now that have no artifacts at all in the mesh. Also Lines or Zigzag is the the Safe option which won't have any chance of misprint, with Grid looking better, but having a higher chance to fail.

You guys inspired me with this stuff, I'm just getting into some light 3D printing, and among other projects I'm making a DIY portable monitor and needed a ventilated lid over its controller board. 3Dmesh turned out perfect for the job. Sliced in PrusaSclicer, "stars" infill at 60% (part is thin and needs a decent amount of rigidity). While the slicer view looks a lot more open than the finished part, I still think it turned out perfect for my needs. Got really lucky with the print too - a couple of holes have some very minor stringing or other defects, and there's that one odd void in one corner, but it's still pretty good. Only cleanup done on this currently is cutting away the brim - but I can imagine a larger, more open version of this pattern needing more invasive cleanup in the holes around the edges. Obviously this is a much smaller (and less airflow sensitive) part than what you guys are working on, but still, thanks for the inspiration!

 

[msystems]

Yep, that looks like a really clean result. The only stringing I see is along the edge where it isn't noticeable anyway.

What we should really do is put together a list of slicer patterns and the software that does it. There might be some patterns exclusive to certain slicers. Would love a Voronai or honeycomb

 

[Valantar]

PrusaSlicer has a honeycomb:

Or, technically two, there's also "3D honecomb", though it's much less useful as a mesh.

The first alternates the direction of the lines in the honeycomb each layer, so it's probably quite strong. The 3D one is ... weird, stacking layers in a strange way that ends up with very little open area, where none of the layers actually create a full hexagon shape, and some are just straight lines across. Doesn't really have the honeycomb look either IMO. The regular honeycomb looks like a decent option though.

Here's the full list of options for PrusaSlicer:

 

[morj]

BTW I realized that if I put the new HDPLEX 250W GAN on the bottom (in vertical orientation) adding 25mm to the height, I will still have a sub 3L package at 265.46 * 64 * 176.15 ~= 2,993L. Debating whether I should go brickless...

The problem, of course, is that the wiring needs to be creative with that amount of space available for cable routing:

 

[morj]

A bit of a preview. Used Kryonaut Extreme TIM, preliminary results are very good (at 90W PPT).

 

[timginter]

great project, thanks for detailed posts! can't wait to see the final product and potentially steal the 3D printed case idea for my lapdesk

Potential issues: can't seem to find a suitable PCI 4.0 riser so far, using short 3.0 riser I had lying around.

My go-to is ADT-Link, they have some listings on aliexpress: https://www.aliexpress.com/store/gr...2g0o.store_pc_home.pcShopHead_393151816.1_1_5, I usually send them a message to get custom length and spec. Shipping time is quite long, though

 

[morj]

go-to is ADT-Link, they have some listings on aliexpress: https://www.aliexpress.com/store/gr...2g0o.store_pc_home.pcShopHead_393151816.1_1_5, I usually send them a message to get custom length and spec. Shipping time is quite long, though

Thanks! The riser you can see in the most recent pictures is already PCI 4.0 by LinkUp. Unfortunately I could find any suitable ADT risers.

 

[timginter]

Thanks! The riser you can see in the most recent pictures is already PCI 4.0 by LinkUp. Unfortunately I could find any suitable ADT risers.

No problem. That's what I meant - you can send a message to ADT-Link on aliexpress and they can make you a custom cable (custom length and connector orientation if possible). The white ones are the flexible cables if I remember correctly. Delivery time may be long, though

 

[morj]

No problem. That's what I meant - you can send a message to ADT-Link on aliexpress and they can make you a custom cable (custom length and connector orientation if possible). The white ones are the flexible cables if I remember correctly. Delivery time may be long, though

Yeah, thanks, but the point is I already bought an expensive PCI 4.0 cable by LinkUp, so I don't need another one

 

[morj]

I'm still evaluating the new 5900X, so far I really don't like the undervolting capability on this particular unit. Seems that it can't even handle -6 all core PBO undervolt without sporadic WHEA errors (Cache hierarchy error) that occur, of course, not during benchmarks, but during some light work with changing load. I'm debating whether it's because of the power supply quality or because of the CPU itself being a bad sample.

Hope to write a proper post about the case and future plans soon-ish. Feel really demotivated by the undervolting issues though.

EDIT: for now I will just evaluate in "Eco mode" without any PBO undervolting, if stability issues will persist, I will RMA the CPU.

 

[Valantar]

I'm still evaluating the new 5900X, so far I really don't like the undervolting capability on this particular unit. Seems that it can't even handle -6 all core PBO undervolt without sporadic WHEA errors (Cache hierarchy error) that occur, of course, not during benchmarks, but during some light work with changing load. I'm debating whether it's because of the power supply quality or because of the CPU itself being a bad sample.

Hope to write a proper post about the case and future plans soon-ish. Feel really demotivated by the undervolting issues though.

EDIT: for now I will just evaluate in "Eco mode" without any PBO undervolting, if stability issues will persist, I will RMA the CPU.

Check event viewer to see which core is causing the instability, then switch to per-core offsets. Most likely it's one or two cores holding you back. My 5800X is at -7/-11/-20 for the rest (which could possibly do more, I just haven't tried it). AFAIK it's typically the highest boosting cores that need the lowest offset, though YMMV. Took me maybe a week of testing to arrive at that, but at least the PC can be used normally in the meantime (just make sure to save anything important frequently).

 

[morj]

Check event viewer to see which core is causing the instability, then switch to per-core offsets. Most likely it's one or two cores holding you back. My 5800X is at -7/-11/-20 for the rest (which could possibly do more, I just haven't tried it). AFAIK it's typically the highest boosting cores that need the lowest offset, though YMMV. Took me maybe a week of testing to arrive at that, but at least the PC can be used normally in the meantime (just make sure to save anything important frequently).

Thanks! I'm planning on benchmarking with clock tuner for ryzen (CTR) on the weekend. The sad part is that my other Ryzen 5900X that I have in my watercooled rig (the one with external rad) has -18 offset all core without any stability issues and I haven't invested a lot of time into tuning it.

 

[Valantar]

Thanks! I'm planning on benchmarking with clock tuner for ryzen (CTR) on the weekend. The sad part is that my other Ryzen 5900X that I have in my watercooled rig (the one with external rad) has -18 offset all core without any stability issues and I haven't invested a lot of time into tuning it.

Yeah, that sucks. There's a lot of core-to-core variability though, so it might not be that bad in the end. Most likely it's just one or two cores holding you back, and the lower tiered cores can likely handle much more of an offset.

 

[morj]

Yeah, that sucks. There's a lot of core-to-core variability though, so it might not be that bad in the end. Most likely it's just one or two cores holding you back, and the lower tiered cores can likely handle much more of an offset.

True. Unfortunately, I already discovered that CTR for Ryzen doesn't support newer B2 stepping processors, so I'll need to figure out something on my own. On the other hand, I also borrowed a thermal camera from a colleague and now I can confirm some of the things that seem to need further improvements in terms of the ventilation of the case itself. There is definitely a hot spot in the Pico PSU and RAM (might be also causes for instability).

 

[BaK]

On the other hand, I also borrowed a thermal camera from a colleague and now I can confirm some of the things that seem to need further improvements in terms of the ventilation of the case itself. There is definitely a hot spot in the Pico PSU and RAM (might be also causes for instability).

Had the same issue with my B01T3.
Solved it by adding a larger fan on top of the CPU heatsink, so that it overhangs the RAM sticks and the Pico PSU.
Here and following posts.

Went with two exhaust fans



Not sure but it seems like your RAM is low enough for such a mod!

 

[morj]

Not sure but it seems like your RAM is low enough for such a mod!

You are right! Not sure if I have space for a 120mm fan because of the riser below, but a slightly offset Thermalright TY-100 fan might work (of course that's just a mockup because I need to make some mounting hardware for the fan itself, now it's just wedged in place).

 

[morj]

Starting with thermal imaging eye-candy.

That's a relative test, but it seems changing the fan might do the trick. I need to figure out proper mounting to conduct a prolonged test, but so far the pico seems to feel better with some fresh air. To check it, I was testing which part of the inside was the hottest. Previously it was the pico, now it's the heatsink of the stressed component.

Temperature before: 84° on the pico and 68° on the GPU (no case mesh lid).

After (100mm CPU fan, thanks @BaK for the idea); 15m of Furmark, pico seems to be always cooler than the GPU heatsink/backplate:

Actual temperature a bit higher than pictured because of the mesh, but the pico is not the hottest component (less red in infra) and the test doesn't crash after 13 minutes.

Furmark screenshot (click to zoom):

 

[morj]

(of course the test is not apples to oranges since I can't run the 100mm fan without the lid, but it gives me some confidence in this tweak)

 

[Valantar]

True. Unfortunately, I already discovered that CTR for Ryzen doesn't support newer B2 stepping processors, so I'll need to figure out something on my own. On the other hand, I also borrowed a thermal camera from a colleague and now I can confirm some of the things that seem to need further improvements in terms of the ventilation of the case itself. There is definitely a hot spot in the Pico PSU and RAM (might be also causes for instability).

You don't actually need CTR though. I made do with the manual method I described above and it's worked perfectly fine. Event viewer will tell you which cores are unstable, so set a negative offset, use the PC until it crashes, decrease the offset for that core by 5, repeat as necessary until stable, then (if you really want to) fine-tune offsets with 1-2 step offset adjustments. It takes a while simply because of the instability mainly being at or near idle, but other than that it's easy and requires very little effort.

 

[morj]

You don't actually need CTR though. I made do with the manual method I described above and it's worked perfectly fine. Event viewer will tell you which cores are unstable, so set a negative offset, use the PC until it crashes, decrease the offset for that core by 5, repeat as necessary until stable, then (if you really want to) fine-tune offsets with 1-2 step offset adjustments. It takes a while simply because of the instability mainly being at or near idle, but other than that it's easy and requires very little effort.

I've updated the BIOS, looks like the stability has improved, but I'm getting crashes seemingly due to transient current spikes when I load the GPU. Without the GPU load the system pulls 130W in Windows if the CPU is loaded. Maybe you were right about 192W being not enough.