1 Topic by manasseh (edited by manasseh 2026-02-02 19:01:34)

  • From: Lisbon
  • Registered: 2009-04-03
  • Posts: 70

Topic: streamed ASIO channels reduction over USB technical question

Hi all,

I wanted to clear a technical question, in case someone can help.

I've been frequently reading here than rme USB interfaces simultaneously transmit every channel it has over wire, given that a UFX III with 188 channels requires a disturbance-free USB endpoint to guarantee error-free streaming.
In optional USB 2.0 mode, the channel count is reduced to 60, excluding the 64+64 MADI channels.

When we deactivate ASIO channels in the DAW, isn't it effectively reducing the driver-streamed ASIO channels and, consequently, the throughput requirement?
Or are all channels always being transmitted regardless of whether they are active in the DAW?

https://i.imgur.com/9EN2xnf.png

Thank you

manasseh's Website

2 Reply by ramses (edited by ramses 2026-01-27 13:44:22)

  • ramses
  • ramses
  • DAW Mastermind
  • Offline
  • From: UTC+1
  • Registered: 2012-10-14
  • Posts: 15,575

Re: streamed ASIO channels reduction over USB technical question

Hi manasseh.

manasseh wrote:

[...]
1- When we deactivate ASIO channels in the DAW, is it not effectively reducing the driver streamed ASIO channels, and consequently the throughput requirement?
2- Or are all channels always being transmitted regardless of whether they are active in the DAW?
[...]

to 1- no
to 2 - yes

The only few exceptions to it:
- Firewire driver option for some older interfaces to deactivate digital ports like ADAT, then only analog ports were transferred,
- HDSPe MADI FX driver because of the high number of channels (3 MADI buses, AES, phones).

See my review in the chapter "RME DRIVER OPTIMIZATION".
The PDF document with the complete review for download: https://www.tonstudio-forum.de/attachme … -v1-1-pdf/

The blog article: https://www.tonstudio-forum.de/blog/Ent … Pro-FS-BE/

BR Ramses - HDSPe MADI FX, M-1620 Pro D, 12Mic, UFX III, ADI-2 Pro FS R BE, Nuendo 15, Win10 IoT Ent

3 Reply by manasseh (edited by manasseh 2026-01-28 23:21:40)

  • From: Lisbon
  • Registered: 2009-04-03
  • Posts: 70

Re: streamed ASIO channels reduction over USB technical question

Many thanks for clarifying, Ramses.

As a tangent to a subject of common interest :
I've been doing some load tests between AIO Pro and Madi FX PCIe using 32 buffer size, and >90% cpu vst load, using in W11, Cubase 15 and Ableton 12 - Despite the optimized driver of the Madi FX, I'm still getting less dropouts under certain circumstanes with AIO Pro (30-Channel) under load when tested against Madi FX (only headphone jack connected - driver should optimize to transmit an 8 channel packet in theory)i - I've tried different configurations (turned on redundancy/MADI mirror etc), swap the PCIe slot position and it seems to be a consistent behaviour.

manasseh's Website

4 Reply by MC

  • MC
  • MC
  • Administrator
  • Offline
  • From: Germany/Thailand
  • Registered: 2007-04-06
  • Posts: 26,989

Re: streamed ASIO channels reduction over USB technical question

You might have overlooked the option 'Optimize Multi-Client Mixing' on the About tab. If that one is checked it adds a bit of CPU load that could cause small differences as in your test. If it was off (default) then we need to find a different explanation.

Regards
Matthias Carstens
RME

MC's Website

5 Reply by ramses (edited by ramses 2026-01-28 06:31:33)

  • ramses
  • ramses
  • DAW Mastermind
  • Offline
  • From: UTC+1
  • Registered: 2012-10-14
  • Posts: 15,575

Re: streamed ASIO channels reduction over USB technical question

manasseh wrote:

I've been doing some load tests between AIO Pro and Madi FX PCIe using 32 buffer size, and >90% cpu vst load, using in W11, Cubase 15 and Ableton 12 - Despite the optimized driver of the Madi FX, I'm still getting less dropouts under certain circumstanes with AIO Pro (30-Channel) under load when tested against Madi FX (only headphone jack connected - driver should optimize to transmit an 8 channel packet in theory)i - I've tried different configurations (turned on redundancy/MADI mirror etc), swap the PCIe slot position and it seems to be a consistent behaviour.

How are your HDSPe FX driver settings?
Is the driver option "Enable MMCSS for ASIO" activated or not?

BR Ramses - HDSPe MADI FX, M-1620 Pro D, 12Mic, UFX III, ADI-2 Pro FS R BE, Nuendo 15, Win10 IoT Ent

6 Reply by manasseh (edited by manasseh 2026-01-28 23:18:47)

  • From: Lisbon
  • Registered: 2009-04-03
  • Posts: 70

Re: streamed ASIO channels reduction over USB technical question

Hello to both, thanks for the interest in the thread.

Optimize Multi-Client Mixing was off, as I know it can contribute to a slightly higher CPU load, according to the documentation.

With the MCCSS flag, I took a deep dive (also attempting to optimize HLM\SOFTWARE\Microsoft\Windows NT\CurrentVersion\Multimedia\SystemProfile specific parameters through numerous iterations) and extensively benchmarked it, enabled and disabled.

With Ableton 12, my conclusion is that MMCSS active is consistently better at very high loads (for both Babyface Pro FS/AIO Pro /Madi FX - with dropouts occurring later, at higher plug-in count.

Across all interfaces, the corresponding loaded ASIO .dll driver thread gets a priority boost to 26 when the MMCSS service is subscribed (as seen in Process Explorer's thread tab), although it can temporarily deprioritize it to lower values.
In Ableton 12.3, the MMCSS service is subscribed only if the flag is active; otherwise, the ASIO thread gets the default priority of 15. Cubase always seems to subscribe to MMCSS, overriding the RME control panel setting.

Coming back to the topic, it did surprise me that MADI FX had earlier dropouts than AIO Pro. Also, the sound of the dropouts had a different character, and sometimes the PCIe interfaces went straight to skipping/glitching instead of audible fast clicks (with Babyface, audible clicks always happen at some point).
Obviously, the PCIe performance differences between the two were quite narrow, often with a 2~4 plug-in offset, but the difference between the two was effectively consistent across several tests.

manasseh's Website

7 Reply by ramses (edited by ramses 2026-01-28 14:28:39)

  • ramses
  • ramses
  • DAW Mastermind
  • Offline
  • From: UTC+1
  • Registered: 2012-10-14
  • Posts: 15,575

Re: streamed ASIO channels reduction over USB technical question

manasseh wrote:

Coming back to the topic, it did surprise me that MADI FX had earlier dropouts than AIO Pro. Also, the sound of the dropouts had a different character, and sometimes the PCIe interfaces went straight to skipping/glitching instead of audible fast clicks (with Babyface, audible clicks always happen at some point).
Obviously the PCIe performance differences between both were quite narrow, often with a 2~4 plug-in offset, but indeed they were consistent across several tests.

My findings...

System:
- CPU: Intel Xeon E5-1680 v4
- Mainboard: Supermicro X10SRi-F
- Network card: Intel X710-DA2 (2x 10Gbit, SFP+)
- DRAM: 64 GB
- Windows 10 Pro
- Application / Music Player: MusicBee with RME ASIO driver
Using CPU-Z to stress the CPU (all cores) to 100% load.

I have connected HDSPe MADI FX (HDSPe FX driver) and UFX III (MADIface driver).
[x] MMCSS for ASIO enabled in both drivers.

I started playback with MusicBee, CPU at 100% (CPU-Z multicore stress test).

With MMCSS enabled
- HDSPe MADI FX needed an ASIO buffer size of 128 samples
- UFX III needed an ASIO buffer size of 64 samples
to be able to play without any clicks.

When there were clicks, they were short, not many but noticeable every few seconds.
With HDSPe and only a 32-sample buffer size there were some more clicks compared to a 64-sample buffer size.

With MMCSS disabled in the HDSPe MADI FX driver, the clicks were completely gone, and I could use the smallest ASIO buffer size of 32 samples without any issues. The UFX III I didn't retest without MMCSS for time reasons, and as the HDSPe MADI FX is my main interface, it was of higher interest.

This was even running click-free when my Macrium Reflect and FreeFileSync backup jobs started
in addition to the CPU-Z CPU Stress Test. Macrium I had set to the lowest priority (configurable in the application).

Additionally, I was able to open in Firefox a bookmark folder tab of 85 web pages regarding photography.
Because many of these sites contain plenty of images. By that I wanted to additionally create a little network load.

I think when all of this was running, Bitsums Process Lasso Pro feature, ProBalance, helped a bit, that all these applications did not use too many system resources.

Therefore, I asked whether you disabled MMCSS; for me, this helped a lot.

If I remember right, there was an older posting from Timur Born on RME forum, who did some investigations on the topic whether or not to activate MMCSS and it could be, that this is application/DAW dependend, maybe even OS dependend.

After disabling MMCSS for Music Playback with MusicBee I have to say that I am deeply impressed about the stability, no audio loss at even very high system loads and CPU utilization of instantly 100%. Thats simply great.

I updated my blog article: https://www.tonstudio-forum.de/blog/ent … tup-en-de/
with a new chapter "HDSPe Driver Settings".

BR Ramses - HDSPe MADI FX, M-1620 Pro D, 12Mic, UFX III, ADI-2 Pro FS R BE, Nuendo 15, Win10 IoT Ent

8 Reply by manasseh (edited by manasseh 2026-01-29 23:54:44)

  • From: Lisbon
  • Registered: 2009-04-03
  • Posts: 70

Re: streamed ASIO channels reduction over USB technical question

Thanks for sharing - interesting findings.

For further insights into the benchmark procedure, I opted for a more Daw-centric methodology, progressively loading Ableton with increasing plug-in instance count (Waves L4 VST3 and Ableton built-in Echo, independently), with no other programs running.

I'm convinced that testing MMCSS through other means can yield different observations, since adding an independent background process load with CPU-Z may steer the test into a different context, likely with other repercussions. Restricting to Ableton (or most DAWs), things are prioritized differently, being a foreground process (including the ASIO .dll driver) with different handling by the OS scheduler. Even the driver can get quite a noble CPU usage under high load.

https://i.imgur.com/ci4f3x8.png

The tested system is somewhat aged but reliable and performant - Coffee Lake i7 8700K/16 GB Ram/140 mm Noctua Cooler installed with Windows 11 25H2 and very low DPC latency. Also simplified the system setting config. for easier analysis (hitting the system ceiling earlier without having a super high plug-in count), restricting CPU to 4 physical cores (No HT). Power plan set to ultimate performance, with a very lean system: 1 M.2 NVMe SSD and no extra peripherals connected. Other tweaks include C-States off and VBS disabled.

Buffer sizes:
AIO Pro / Madi FX : 32 samples (Redundancy/mirror active in Madi FX for theoretical decreased resource usage)
Babyface Pro FS : 48 samples
I also tested everything at a more level playing field of 64 samples, but found that the differences are more pronounced at the lowest buffer sizes.

For extra coverage, AIO Pro/Madi FX cards were installed in different configurations - swapping PCIe slots with each other (one x16 lanes slot is direct to CPU, the other x1 slot to PCH) and just each one by itself in the two available slots - no differences observed.

Interesting to note that AIO Pro RTL @ 64 samples is quite closer to Babyface RTL @ 48 samples than 64 :

https://i.imgur.com/EDIo6UL.png

To test single-core exhaustion, I assembled a project with a single track and increased the plug-in count. The other 2 test iterations used 4 and 8 track counts to evenly distribute the load across the cores.

Allow me to reiterate, my test results with Ableton 12 - MMCSS active is consistently better at very high loads in all 3 interfaces. This could be verified audibly and visually using Live’s real-time meter (set to current, to observe peaks and dips). In the vast majority of test cases, when setting the flag off and reloading the ASIO driver, audible dropouts occurred immediately, whereas with MMCSS enabled, there were no playback stutter/dropouts. I have recorded most of the test results' audio clips using a laptop/external interface, recording each interface's headphone output, in case anyone is curious to hear.

Babyface Pro FS USB 2.0 performance hit earlier ceilings for stutter/dropout free playback, as expected (not by huge margins, to be clear) - more prominent at 48 buffer size (vs. AIO/Madi FX @ 32). The last one to hold up consistently was AIO Pro, with MADI FX a very close second.

Some audio clips from the 4 track load test @ 32 samples - at the verge of dropout occurrence :

AIO MMCSS       https://od.lk/s/NDlfMzg5MjkzMjVf/AIO%20MMCSS.mp3
AIO                      https://od.lk/s/NDlfMzg5MjkzMjZf/AIO.mp3
MadiFX MMCSS https://od.lk/s/NDlfMzg5MjkzMjNf/MadiFX%20MMCSS.mp3
MadiFX               https://od.lk/s/NDlfMzg5MjkzMjRf/MadiFX.mp3
  *note that Madi FX presents audible dropout artifacts while AIO Pro is still clean (with MMCSS)

For Babyface Pro FS, here's another load test example @ 48 samples, with less plug-in instances than above :

MMCSS enabled
https://od.lk/s/NDlfMzg5MjkzMjFf/babyface%20MMCSS.mp3
https://i.imgur.com/X5Bz2Ou.gif


MMCSS disabled
https://od.lk/s/NDlfMzg5MjkzMjJf/babyface.mp3
https://i.imgur.com/2hOghnd.gif


By the way, Ramses, I follow your postings and blogs with great interest. Your contributions to the community are highly valued.

manasseh's Website