Need PC Hardware advice SFF / ITX Build for Roon & HQPlayer

heyguyslol · July 30, 2025, 8:14pm

I’m seeking PC hardware & OS software advice for a dedicated PC to run both Roon Server & HQPlayer Embedded.

While my current PC Setup (listed below) is functional & meets performance expectations; I’d like to downsize to an ITX / SSF build to reduce both physical size & power consumption.

Current PC Setup:

Windows 11 Pro (24H2) / Roon Server / HQPlayer Desktop

i7-14700K + RTX 3090 + 64GB 6200 MT/s DDR5 + M.2 SSD STORAGE + ETHERNET (no wifi)

Audio Chain:

PC ► ROON ► HQP ► HOLO-RED ► SPRING-3 ► GOLDPT-PREAMP ► AMP ► CANS

My Use Case:

2-Channel Headphones (no plans to expand beyond headphones)

DACS: Spring 3 KTE & Cyan 2

ROON: preferred source 44.1 & 48 kHz (able to source 192 kHz when needed but not required)

HQP: PCM & DSD, I have HQP Advanced setup to CPU Multicore DSP (Auto), E-core allocation (Default) & GPU CUDA offload (all) Default CUDA & Convolution CUDA (RTX 3090) & Dac Correction Enabled.

I’m fairly new to HQPlayer so I’m still experimenting w/ different modulator & filter combinations. With that being said my current setup can do all ASDM7EC DSD256 & 512 but not 1024. However, I can do AHM7EC8B DSD1024.

Filters I currently use:

poly-sinc-gauss-xla

poly-sinc-gauss-hires-lp

sinc-M

sinc-MGa

sinc-L

sinc-Lh

With all that info in mind; now comes the questions on whether or not the hardware & software I’m considering is a viable option to maintain current performance expectations while also reducing power consumption & system size.

For the CPU & MOBO & RAM combo, I’m considering either of the two MINISFORUM barebones:

BD795i SE (AMD 7945HX)

BD795i SE X3D (AMD 7945HX3D)

w/ 48GB 5200MHz DDR5

For the GPU, I’m considering:

RTX 5070 (cuda core downgrade from my 3090 = 10496 vs 5070 = 6144)

NOTE: Ampere RTX 3090 CUDA 8.6 vs Blackwell 2.0 RTX 5070 CUDA 12 can provide 20–50% or higher performance increases over Ampere for compute-intensive workloads (e.g., matrix multiplications, FFTs) due to better FP32 throughput and Tensor Core utilization.

For OS Software:

I’m looking to make the jump to Linux. Most likely starting w/ a GUI Ubuntu and eventually moving into a CLI version so I can run HQPlayer Embedded (Linux distro recommendations are welcomed).

In my attempts to research the recommended hardware for Roon & HQPlayer; I’ve noticed a lot of ppl are using Intel CPU’s. While I have found more often than not that Intel generally fits my use case; I’m considering AMD bc of the lower power consumption which is a much better fit for an ITX / SFF build rather than an Intel equivalent cpu even tho I’m going to be offloading to the GPU.

My understanding (plz correct me if I’m wrong) is that Roon Server is very light which doesn’t require much as far as cpu horse power and while the same could be applied to HQPlayer when using PCM that isn’t the case for DSD.

HQPlayer will utilize cpu cores for modulators and when paired w/ a Nvidia GPU will handle filters & convolution. One of the concerns based on what I’ve researched is Intel vs AMD L1, L2, & L3 Cache.

HQPlayer greatly benefits from L1 Cache which Intel chips tend to have more of (14700K L1 = 80 KB vs 7945HX = 64KB).

I’m also questioning possible performance differences between the 7945HX vs 7945HX3D if the extra L3 cache = 128 MB & 3D V-Cache: 1x64MB is of any real benefit for my use case even tho the system will not be used for gaming. I’ve heard that the extra Cache helps compensate for the slower ram latency speeds.

Any advice and/or corrections to my research is greatly appreciated.

Kyle_Oliver · July 31, 2025, 6:18am

Nice setup. I can only get dsd512 (although best my dac can do is dsd512 anyway) but no sinc-L at dsd512 and dsd256 is also a stretch with a decent modulator.

My setup and depending on what I am listening to these are my preferred combinations:

Headphone setup

Legion-s7-15ach6
AMD Ryzen 7 5800H 24gb ram
Nvidia RTX 3060
OS Linux Ubuntu
HQPlayer desktop
Ifi lan isilincer
Rapsberry Pi 4 NAA
Ifi ipurifier3
Geshelli JNOG 3 w/ sparkos 2590 opamps
Topping A90D

Vocal, Acoustic, 70s rock
HEDDphone two (EQ’d for vocals)
poly-sinc-gauss-hires-mp & ASDM7ECv2 DSD512

Rock, Metal
Focal Clear MG (heavy eq)
Poly-sinc-xtr-mp & ASDM7ECv2 DSD512

Movie scores, soundscapes
Hifiman HE1000 Stealth (slight EQ)
Poly-Sinc-Gauss-long & ASDM7EC-Super 512+fs DSD512

Pop / electronic
Hifiman HE1000 Stealth (some EQ)
Poly-sinc-short-mp ASDM7ECv2 dsd512

I am really liking the minimum phase filters to my ears sound more natural on headphones YMMV. I am also still experimenting so above is what I like at the moment.

I have read AMD CPUs and Nvidia combo work really well with HQPlayer but I don’t have another reference than what I have although considering getting a desktop with latest gen AMD cpu and a 5070 Ti or better.

jussi_laako · July 31, 2025, 8:18pm

It depends on the Roon configuration and what it happens to be doing. When it is performing background scans, it can consume quite a bit of CPU power. Also if you have any DSP or such enabled at Roon side, it can also consume a lot of CPU.

Overall, especially L3 cache sizes matter more. But please note that you are comparing a powerful desktop CPU to a low power laptop CPU. So your CPU performance may suffer with 7945HX compared to the 14700K.

Larger caches do help. But these caches are always a bit of black magic, so it is difficult to predict how much it ends up making difference in the end. It depends on the cache’s perdiction hit/miss ratio…

heyguyslol · July 31, 2025, 11:41pm

Tyvm for the info.

It appears that the MINISFORUM BD795i SE X3D 7945HX3D out performs the non-3D version while also using less power & lower temps. While it is true both cpu’s are mobile chips the performance gap based on benchmarks closes quite a bit being that it’s in a ITX FF w/ a better cooling solution.

However, the 14700K has a fairly large lead in single core performance. The other alternative that I was originally considering before discovering the 7945HX3D is to go w/ a 9900X3D which would drop 4 cores but boost that single core performance. While the power and heat are very close the price more than doubles. The 9950x3D would go way out of bounds to the point that I might as well stick w/ the 14700K setup.

I think the only real way to know for sure is to buy it & try it. I’m not in any rush to do so and Amazon typically has a larger return window during the holidays so I may take advantage of that to fully test it.

jussi_laako · August 1, 2025, 10:14pm

If you use GPU offload, then the single core performance becomes stronger decision factor, since most of the algorithms that can be run in parallel on multiple cores are offloaded to the GPU.

heyguyslol · August 1, 2025, 11:40pm

Turns out I might be wrong about the 9950x & x3D from my previous assesment comparing power & heat to 14700K. Apparently the 9950x3D can be tuned to power limit w/o loosing much in performance while allowing the use of a low profile air cooler for an ITX/SFF build.

Here’s a vid testing the 9950x3d power limits using a low profile air cooler (ID-COOLING IS-67-XT)

That inspired me to try some tests of my own on the 14700K.

I set the power limit down from 253 to 160w. I first tested Cinebench R23; temps were averaging about 65C (90C is the norm for 253w) using Thermalright Peerless Assassin 140 running at 1100 RPM speed doing 10 min = 29500, while single core stayed the same 2150.

Next I tested the 14700K + 3090 HQPlayer 2-channels, dac correction on, no eq/dsp (I generally don’t use it), 1x & Nx filters set to sinc-Mx & ASDM7ECv2 @ 48K x512 which from what I’ve read is quite a heavy load to use as a benchmark test. The playback performance showed no degredation (no static or drop outs) smooth clean playback.

I’m not so sure trying to power limit the 14700K down further to get it to fit into a ITX/SFF build would be the best solution given how good the 9950X3D performs in the test video.

I’m very skeptical about the lower single core performance of the 7945HX3D and the 9900x3D only has 6 cores per ccd which from what I’ve read can cause issues of latency communication between the two. Considering HQPlayer fully utilizes all 8 cores of my 14700K w/ the test I ran the 9950x or x3D might be the best solution.

However, the 9950x or x3D might be over kill for what I need but it appears it would work very well for a ITX/SFF and allow for headroom for other configurations should I choose to expand.

EDIT: Another plus by going 9950x3D is that the mobo will allow for Zen 6 upgrade which AMD claims potentially hitting 6ghz clock speeds and of coarse more cores, maybe even lower power requirements and less heat.

jussi_laako · August 2, 2025, 12:19am

I have i9-13900T in such setup, with RTX2080 Founders Edition GPU. Running Fedora Server 41 and HQPlayer Embedded. It can do DSD1024 without issues. I didn’t do anything special about the power consumption. This is Fractal Design Ridge case.

I also have i9-14900K + RTX A4500 on a Mini-ITX main board and case, with 280mm AIO cooler. Two 140mm PWM fans and three 120mm PWM fans. No power limiting there too, but this is not as much “SFF”, since it has regular Seasonic hybrid 1000W ATX 3.0 PSU, etc. This in Fractal Design Define Nano case.

But if you are interested on HQPlayer performance, only benchmark that matters is HQPlayer itself. Anything else is very likely misleading at best.

heyguyslol · August 2, 2025, 8:32pm

With my current 14700K + 3090 setup based on your system specs I should be able to do DSD1024 but I can’t (even w/ default power settings). Atm I can only do DSD1024 if use AHM7EC8B. When I test ASDM7ECv2 or v3, I get drop outs.

I am using Win11P which could be the weak link vs Linux real time. I do have Windows debloated w/ HQPlayer set to high priority while only running Roon Server & HQPlayer. I also monitored system apps and Roon Server is barely hitting the cpu so I don’t believe that is the weak link. If interested I can post my full cfg screens to help find the problem. I do have my system setup as a dual boot w/ Ubuntu 24 so I will also try and test on Linux to see if that allows DSD1024.

Furthermore, I did research the 13900T & 14900T and I must say those cpu’s are rather interesting but very niche. So niche that I was unable to find benchmark scores not to mention that the 14900T is unavailable to purchase along w/ the fact that the 13900T is about $200 more than a 14900K.

I’m guessing the increase cost (13900T) is bc the chip is able to deliver very good performance w/ low power consumption which I would agree w/ you would not need any tinkering w/ power.

However, I did discover the following article which I found to be quite interesting and I will be doing some further power limit testing w/ my 14700K based on their reports. It’s long read so I’ll understand if you skip it but I got the gist of what they were saying.

In short as an example lets say you were using your 14700K or 14900K and the most amount of cpu power used was 130w you could power limit the cpu to 140w, maintain the same performance which can reduce power & heat. Knowing that I did a couple of tests w/ the following:

1x sinc-Mx + Nx sinc-Mx - ASDM7ECv2 @ 48Kx512; still power limited to 160w. The 14700K only pulled about 60w & gpu 200w and the wall outlet monitor full system load was right at about 300watt.

In the article they tested various power limits w/ the 14900K even going as far down as 35w. Now why is this important? well theoretically one could power limit a desktop i7 & i9 down to only what’s needed for HQPlayer which will reduce power & heat therefore making it more ITX/SFF friendly.

Power limiting mainly effects multi-core performance but single core remains the same unless you go below the required threshold. If one were offloading to the GPU then HQPlayer is still going to benefit from the fast single-core performance.

Another question is why bother power limiting 14700K or 14900K? Just go buy the 13900T or AMD 9950x3D; well for one a $200-250 price increase over a 14900K which can be power limited up & down (more flexible) and the jump to the AMD platform is stupid expensive compared to Intel, which is why I was considering the MINISFORUM BD795i SE X3D. Then there’s the questionable performance difference between the Intel vs AMD architecture on whether or not HQPlayer has a preference for optimal performance.

There’s also the issue that some ppl over look and that’s AMD’s idle power consumption and heat output. I live in south Texas so things are a bit toasty down here. Running an Intel chip is going to be much cooler both idle and under load; Intel idle power 5-10w 30-35C vs AMD 20-40w 40-50C.

heyguyslol · August 2, 2025, 10:09pm

Ideally I’ve been hoping to get a powerful sys build in ITX/SFF at 10L or less but cooling either a 13/14 i7/9 Intel or AMD 7/9000 chip while maintain performance is rather challenging.

I did discover the CRYORIG C5 COPPER low profile fan cooler (not yet released) that utilizes a vapor chamber and is expected to handle up to 185w TDP at a height of 55mm. Still not sure if and when this will actually see a launch as it was suppose to launch in June.

If the Cryorig C5 does launch and it performs close to what they claim then one could tune either an Intel or AMD chip w/o the need for a large air cooler, aio, or worst a custom loop and down size while maintaining performance.

My Current ITX Case line up from most to least preferred w/ a stopping point at 10L (I had to make an exception for the Lian Li 11L)

Sparrow MQ6 (6.8L) AIR Cooling only

Archer X Mini AF4 (7.0L) AIR Cooling only

Ghost S1 MK3 (8.3L) AIO & AIR Cooling but requires top hat for aio (too jank for my taste, I’d rather go FormD T1 or Lian Li)

FormD T1 Customize (10.0L) AIO & AIR Cooling

Fractal Terra (10.4L) AIR Cooling only

Lian Li A4-H2O (11.0L) AIO & AIR Cooling

jussi_laako · August 2, 2025, 10:35pm

I have not tested those older modulators. How about for example ASDM7EC-light? or ASDM7EC-fast?

I don’t have any special realtime stuff. Stock Fedora Server 41 with Nvidia’s open source driver. Output to a NAA.

You should never touch HQPlayer priority, as HQPlayer manages it’s own priority assignments. It has from few tens to couple of hundred threads executing different kind of tasks, each with carefully designed priority levels vs each other. Setting some blanket priority to everything will break the performance…

T-series are low TDP parts. Not available as boxed, only as tray. So you need to buy from some reseller who is OK to stock bulk parts instead of consumer boxed ones.

I have also an old i5-7600T running on Streacom FC10 case with their ZeroFlex PSU, so completely passive cooled. And it’s doing DSD256 upsampling with default filters and ASDM7EC-fast to the Marantz SA-12SE, at about 75% load (no GPU). And it barely warms up. Running on Ubuntu Server 24.04 LTS and Intel build of HQPlayer Embedded.

So for low power builds, I prefer T-series CPUs. It has not been particularly difficult for me to obtain such from regular computer part stores.

Which are irrelevant for HQPlayer case anyway…

jussi_laako · August 2, 2025, 10:46pm

On the i9-13900T in the Fractal Design Ridge case, I have either this:
https://noctua.at/en/nh-l12s
this:
https://noctua.at/en/nh-l12-ghost-s1-edition
or this:
https://noctua.at/en/nh-l9x65
cooler. I cannot remember right now which one of these… I can probably dig it out if there is particular interest knowing the exact model.

jussi_laako · August 2, 2025, 11:14pm

And just for reference, I’ve been using this thermal paste for the recent builds:

heyguyslol · August 3, 2025, 12:00am

DSD1024 Tried both, did not work w/ either.

Setup Refresh:

I disabled running HQP high priority as you recommended (I only recently started that a few days ago as a recommendation when testing the power limit stuff).
Win11P (24H2), 14700K + 3090 64gb DDR5 ram, m.2 ssd, Ethernet (no wifi)
2-channel headphones, Qobuz > Roon > HQP no DSP/EQ w/ 44.1-48 source & Dac Correction enabled, Spring 3 KTE.
Screens of my settings in case you see something that may need changing.

1-INPUTS1156×754 16.9 KB

2-OUTPUTS1156×754 25.8 KB

3-PCM1156×754 19.7 KB

4-SDM1156×754 19.3 KB

5-ADVANCED1156×754 20.3 KB

6-MATRIX1192×892 30.4 KB

FYI, w/ this setup I’ve never been able to do DSD1024 until I updated HQP to the latest vs and tried the new mod AHM7EC8B, which works flawless.

jussi_laako · August 3, 2025, 9:40pm

Set “Adaptive rate” to checked here.

1x = poly-sinc-gauss-long
Nx = poly-sinc-gauss-hires-lp

Try Multicore DSP set to checked.

Not related, but with Roon, “Idle time” should be set to something. Like 10 seconds or 30 seconds for example.

You can also leave GPU setting under “Default CUDA” and “Convolution CUDA” set to “Default”, since you probably have just one GPU. Not that it makes any practical difference.

heyguyslol · August 3, 2025, 11:57pm

Applied all the changes but still having drop outs. The time between drop outs has decrease quite a bit.

On a positive note the recommended change for “idle time” when using Roon fixes another issue that I was having which was the need to refresh the hqp app on my tablet while monitoring the buffer bar and spectrogram graph.

Btw, I loaded hqp on my workstation which is a Win11P (24H2) 14700K + 5070 Ti 64gb DDR5 and applied the changes and got the same results.

The one thing to note is that the 3090 usage load spikes up & down from 50-100% during playback w/ a 50C temp while the 5070 Ti stays at a steady 15% 38C temp. Slightly better performance on drop outs but they still occur.

Is it possible that the use of Windows is the weak link vs Linux?

jussi_laako · August 4, 2025, 12:11am

Are you on latest Nvidia Studio or Game Ready drivers? Some people have got to set power profile from the Nvidia control panel to maximum performance or something similar (I don’t recall right now how it is called).

On Windows, please set the power profile to “High Performance”.

heyguyslol · August 4, 2025, 1:29am

The PC w/ the 3090 has the latest Game Ready drivers while the 5070 Ti PC has the latest Studio Drivers.

I always DDU when loading new drivers. I’m also not using that garbage Nvidia app, just load the base drivers & Nvidia Control Panel.

I have both PC’s Nvidia CP set to “Prefer Maximum Performance” and I have both PC’s Power plan set to high performance w/ no power limiting, just default bios power 253w.

Same results.

jussi_laako · August 4, 2025, 10:27pm

OK, seems to be a mystery. You have some AIO cooler for the CPU? And you are certain it doesn’t thermal throttle?

Is it DAC correction which is the bottleneck? It works if you turn it off?

Just in case, did you try Qobuz with standalone HQPlayer?

heyguyslol · August 4, 2025, 11:19pm

Neither PC thermal throttles. I have a couple of options to check for thermal throttling:
Aida64
HWiNFO64

PC 14700K + 3090 is using AIR cooling - Thermalright Peerless Assassin 140mm

PC 14700K + 5070 Ti is using AIO cooling - CORSAIR iCUE Link Titan RX 360mm

EDIT: CPU load is 15% on both pc’s w/ 50-55C temps depending on the cooler type.
GPU 3090 usage load spikes up & down from 50-100% @ 50C temp.
5070 Ti stays at a steady 15% @ 38C temp.

My ambient temps 24-25C.

HQPlayer Desktop v.5.14.0

I tried both PC’s w/ DAC correction on & off.

No, I’ve always run Qubuz directly thru Roon. Roon Server is running on the 3090 system but obviously not on the 5070 Ti system as I’m just running HQPlayer so technically the more powerful system not hosting Roon Server should have more resources to work with.

Just as a reminder pretty much every modulator & filter combo I’ve tried at DSD512 and below works. It’s just DSD1024 that does not unless I run that new modulator AHM7EC8B.

jussi_laako · August 5, 2025, 12:49am

OK, something is limiting the performance, but I don’t know what. I have no trouble running DSD1024 without DAC correction on i9-14900K or i9-13900T, without GPU. Although neither one is running Windows (former is Ubuntu 24.04 LTS and latter is Fedora 41).

My Windows 11 server with i9-10900K and RTX 2080Ti is usually running DSD512 output with DAC correction.