Description

The goal was to build a small, near silent, powerful machine learning system using known high quality parts. I chose the Fractal Design Define Mini C MicroATX Mid Tower Case for its size and quietness. An Intel 8700K was preferred over an AMD Threadripper for its power/performance trade-off. The ASROCK Z370M Pro4 motherboard has two 16x PCI 3.0 slots, possibility for M.2 Raid, and four DDR4 slots supporting 64GB but no WiFi. The Seasonic power supply was a fit for the case, quiet, on sale, albeit with 850W is over-powered for the build. The MSI GTX 1080 Ti Gaming X Trio is a top of the line GPU for machine learning needs and is particularly quiet. Despite a warning from PcPartPicker, the GPU really fits once the fan is put in the alternate location (see photos). I got 2x16GB DDR4-3200 to leave room for an additional 32GB, if needed.

The newly available Intel Optane SSD 900P 480GB is reviewed to outperform other solutions such as NVMe M.2 cards. I got the PCIe version but might have opted for the U.2 version thus freeing up a PCIe slot. Unfortunately a 480GB U.2 was not yet released.

Update 18-Nov-2017: The idle cores are around 45C. Under Prime95, overclocking at 4300MHz to 4500MHz, cores show 80C. I'm mostly interested in overclocking up to the point that the fans increase in loudness which is a very modest overclock.

** Using either Prime95 or Intel Extreme Tuning Utility Stress Test benchmarks.

Update 19-Nov-2017: How quiet can it be? I removed side panels and disabled the two chassis fans. Overclocked temps soared to above 90C while the CPU fan roared. Next, I tried lowering the CPU maximum clock until the rig became near silent. At maximum CPU clock of 3900MHz, I can't hear any fan noise. Running Prime95 on all CPUs, temps stayed between 48C and 68C. Idle temps were 26C to 32C. This is a very good baseline for silence. I'll experiment with increasing performance while keeping it as quiet.

Update 23-Nov-2017: Replaced stock Fractal fans with Artic Silent. They are indeed quieter than stock Fractal. Front 120mm changed to 140mm correcting direction. With max CPU clock of 3900MHz temps dropped 5C to 48C to 63C. With max CPU clock of 4200MHz temps were 55C to 70C. With max CPU clock of 4400MHz temps were 58C to 76C (quiet but not as silent). With max CPU clock of 4600MHz temps were 60C to 80C (borderline quiet, definitely quieter than my notebook fans). With max CPU clock of 4800MHz temps were 60C to 83C (notebook fan loud). With max CPU clock of 5000MHz -- crash.

Update 7-Dec-2017: Running day-to-day with ASROCK A-Tuning software settings at 48x CPU, 46x CPU cache, -.05 Vcore voltage (offset). Temps are 31C idle, 78C max (100% CPU and 3DMark Extreme stress tests). Temps are very dependent on A-Tuning FAN-Tastic settings. I optimize RPMs for silence instead of cooling/performance.

Update 23-Dec-2017: I've returned to running the CPU clocks with modest overclocking (CPU 4400MHz (nominal 3700MHz), CPU cache 3700MHz). The system runs 24x7. The system has awesome power which easily handles day-to-day tasks (idle is 26C, average is 30C-36C, max is 65C-80C). So far, the machine learning tasks are GPU bound so little need for overclocking. I know 4800MHz is available whenever I need it by simply by starting up ASRock's A-Tuning software. Ambient sound level in the room is 44db. The system adds no measurable noise unless really stressed in which case, measuring at the fans, 47db, 54db and sometimes as high as 64db (overclocked to 4800MHz,100% all cores plus 100% GPU).

Comments

  • 22 months ago
  • 4 points

You might want to consider turning the heatsink so the exhaust is going out the back of the case. It might improve temps.

  • 22 months ago
  • 2 points

Wow. Thanks so much for that tip. I'll correct it immediately. I can only imagine it will improve circulation. If I can lower temps by a few degrees, it will make a difference.

  • 22 months ago
  • 1 point

Check out my build using this cooler. My idle temps are 27 and under load is 50 when you have the heatsink oriented correctly. If you are into overclocking this will help out a bunch.

  • 22 months ago
  • 1 point

Sweet. Thanks again for your catch. I've reoriented the cooler. My idle temps are similar to yours but you're way better under load. I haven't experimented enough to determine why my temps are higher but there's some important component and goal differences between our builds. I've really favored achieving near silence over top performance.

  • 22 months ago
  • 1 point

My underload temperate readings was when I was running Destiny 2 which is more GPU heavy. At 100% CPU usage it could be a few degrees hotter.

  • 8 months ago
  • 1 point

Hi. Were you able to rotate your cooler? I have the same motherboard and I was thinking to purchase DH15. However, based on Noctua website, this cooler doesn't fit Arock Z370m vertically.

  • 22 months ago
  • 1 point

How cool does the CPU stay under that cooler?

  • 22 months ago
  • 1 point

The idle cores are around 45C. Under Prime95, overclocking at 4300MHz to 4500MHz, cores show 80C. I'm mostly interested in overclocking up to the point that the fans increase in loudness which is a very modest overclock.

  • 22 months ago
  • 1 point

Ok, good to know, just wanna know about how good that cooler is. Thanks!

  • 22 months ago
  • 1 point

I've updated Description with latest results.

  • 22 months ago
  • 2 points

Thanks!

  • 21 months ago
  • 1 point

Does you card have a little bit of coil whine under heavy load? I have the same card and I can hear coil whine when the room is dead slient and the card is under full load. It's not super loud but I paid $800 and want to be 100% satisfied. Debating on whether I should ask for a replacement.

  • 21 months ago
  • 1 point

I haven't explored the graphics card at all. My focus is exploring maximum CPU performance while remaining quiet. Can you recommend a quick GPU test?

  • 21 months ago
  • 1 point

https://m.newegg.com/products/N82E16833980093 Get this asap, it's on sale.

  • 21 months ago
  • 1 point

Thanks for the tip on 3DMark. I'm running Time Spy Extreme Stress Test (98.5%). I don't hear any coil whine but that might just be my old ears.

The stress test froze my system because it was under volted -.100. I mistakenly thought it was stable. Good that I discovered that.

  • 21 months ago
  • 1 point

Hey have you tried delidding the chip? You have already have a great cooler and delidding should definitely help reduce temps and give more overclock headroom. Also adding another Noctua fan to the cpu cooler should help with temps. It’ll be similar looking to Benjamin1996’s build.

  • 21 months ago
  • 1 point

I'm satisfied with performance and temps. I'll leave the remaining optimizations to those who wish to follow that path. I'm really pleased by the build. Is really is small, silent, powerful. I've been running the system at default clocks. It's a really capable system. When I need an overclock boost, I know I can reach 4800MHz CPU and 4600MHz CPU cache. I'm not even sure that's needed as my machine learning tasks, so far, are GPU constrained.

  • 19 months ago
  • 1 point

Hey bsalita, quick question. I want to get a 8700k with a intel 900p and a 1080ti, but I realize that the 8700k only has 16 pci-e lanes. Are you doing 8x to the 900p and 8x to the 1080ti? Is there a bottleneck with the 1080ti or the 900p because of the pci-e limitation? I am thinking of just doing the U.2 version of the 900p because i think the u.2/m.2 connections route through the PCH and dont take any PCI-e lanes

  • 18 months ago
  • 1 point

The GPU is pci 16x. IIRC the 900p may only be 8x or 4x which, might be full speed.

  • 16 months ago
  • 1 point

Is the Optane 900P helping at all with your machine learning work?

  • 10 months ago
  • 1 point

Sorry for not seeing this earlier. No, I don't feel the 900p is a good value. It may be somewhat faster but not the best place to put money in. It might be better to use the money to get a better GPU or more CPU cores.

  • 9 months ago
  • 1 point

Did you find anything on using both optane and a 960 SSD? Seems like Optane helps the slower drives, but I was curious about having both. The cost isn't that high anymore!

  • 22 months ago
  • 0 points

This doesn't seem like a machine learning system, FP16 performance is severely gimped on GP102 (*Note - Geforce GPUs) based parts, 1/64 of FP32 performance without library translation. 4x INT8 works, but there is nothing that uses that, as most of the algorithms require FP16 performance. If you are serious about doing machine learning get either Vega or GP100 based parts, they boast 2x FP32 perf in FP16 operations which are crucial to machine learning.

Great gaming PC though

  • 22 months ago
  • 4 points

You're giving advice without providing perspective. You should mention a GP100 costs $5,000. I'm guessing that's an important detail to readers. You aren't mentioning that Vega has no OFFICIAL support from Tensorflow, Python 3.6, PyTorch, CNTK, MXNet. I'm guessing those are important details too.

  • 22 months ago
  • 0 points

Well I was thinking that if you had money to spend on 8700K + 1080Ti, then you would have money to spend on Quadro GP100 which is 4000-4500 currently (not 5000). Although I guess I assumed too much for what it is so I apologize for that. Fair enough for on GPU choices.

  • 22 months ago
  • 2 points

Cheers.

  • 22 months ago
  • 0 points

Hey! I think the thermals will be better if you turn the front fan to intake instead of exhaust.

  • 21 months ago
  • 1 point

Thanks for the catch. I changed the orientation and swapped in a 140mm Artic Silent fan. The Artic Silent fans are quieter than stock Fractal. Appears to make a difference but I haven't measured yet.

[comment deleted]