Why is USB latency an afterthought in modelers?

[EOengineer]

Sorry this one is kinda long and probably pointless.

I’ll start by saying I LOVE tracking guitars through Helix hardware. The guitar input section in that unit sounds SO GOOD direct, even when driving other modeling software or reamping.

HOWEVER, as a live performer in an all direct/modeling band, I’m unable to use Helix’s USB interface because of its poor USB audio performance, with latency ~10ms in a modern Mac. I’m of the opinion that putting in the work to drop USB latency would open up some really cool workflows for tech savvy bands. Let me elaborate.

Our bass and both guitars use Helix and/or Tonex. We run XLR or 1/4 cables from each hardware modeler to a Quantum 2626 into Logic/Mainstage and monitor back out through an Aviom system. We do this weird hurdle because the Quantum is extremely low latency (< 3ms) and helps us feel more connected as a band. We can’t introduce the Helix into our OSX Aggregate device pool without tripling the latency. OSX’s aggregate devices always slow to the speed of the slowest device in the pool.

If Helix Stadium and other modelers had the same kind of USB performance as the Quantum, each band member could just run USB, cut out the extra DA conversion stage both coming in through the quantum, and likely get their monitor mix from the modeler as well, this also eliminating the need for the separate monitoring system and THAT DA stage too.

It’s probably too niche a case for mass appeal, most people aren’t gigging with a computer, but I love the idea of each band member basically bringing their own interface and monitoring setup.

 

[MirrorProfiles]

I KIND OF feel like aggregate devices are too flaky to rely on as a serious/robust option. I almost never use the built in audio interface features of guitar modellers (so I agree that they're not really up to par). Usually the limited I/O means there's some amount of routing juggling to do and I think that can lead to more issues than using the I/O in a caveman way. It's a bit of a waste of the USB features but I guess they're handy as a fallback in some situations. For band use, I'd either be doing SPDIF or AES, or just suck up the digital conversions (as they're generally not causing many issues).

If Helix Stadium and other modelers had the same kind of USB performance as the Quantum, each band member could just run USB, cut out the extra DA conversion stage both coming in through the quantum, and likely get their monitor mix from the modeler as well, this also eliminating the need for the separate monitoring system and THAT DA stage too.

I think this would be a great thing to achieve, but even if it was REALLY solid, I'd still be apprehensive about doing aggregate devices vs having a solid "all in" rig.

 

[Lysander]

The tl;dr is that there are a lot of moving parts involved when it comes to USB audio, to the point that guaranteeing low latency is but impossible
And this is the case even with dedicated recording gear, where getting low single-digit latencies involves battling with ASIO drivers and hardware settings.

I love using the HX Stomp for recording - it's a fantastic audio interface for guitar - but you're right, there's no way to track realtime audio through USB with it.

 

[EOengineer]

The tl;dr is that there are a lot of moving parts involved when it comes to USB audio, to the point that guaranteeing low latency is but impossible
And this is the case even with dedicated recording gear, where getting low single-digit latencies involves battling with ASIO drivers and hardware settings.

Yup, I totally agree with you. It’s challenging.

FWIW I have a couple very stable aggregate configurations with various devices that have worked well for me for a couple years and across OS updates. It could be total luck or anecdotal but I’ve felt like aggregate stability is better than it used to be.

None the less, get 5 devices together and who knows. It would be a crapshoot.

 

[LP59]

Sorry this one is kinda long and probably pointless.

I’ll start by saying I LOVE tracking guitars through Helix hardware. The guitar input section in that unit sounds SO GOOD direct, even when driving other modeling software or reamping.

HOWEVER, as a live performer in an all direct/modeling band, I’m unable to use Helix’s USB interface because of its poor USB audio performance, with latency ~10ms in a modern Mac. I’m of the opinion that putting in the work to drop USB latency would open up some really cool workflows for tech savvy bands. Let me elaborate.

Our bass and both guitars use Helix and/or Tonex. We run XLR or 1/4 cables from each hardware modeler to a Quantum 2626 into Logic/Mainstage and monitor back out through an Aviom system. We do this weird hurdle because the Quantum is extremely low latency (< 3ms) and helps us feel more connected as a band. We can’t introduce the Helix into our OSX Aggregate device pool without tripling the latency. OSX’s aggregate devices always slow to the speed of the slowest device in the pool.

If Helix Stadium and other modelers had the same kind of USB performance as the Quantum, each band member could just run USB, cut out the extra DA conversion stage both coming in through the quantum, and likely get their monitor mix from the modeler as well, this also eliminating the need for the separate monitoring system and THAT DA stage too.

It’s probably too niche a case for mass appeal, most people aren’t gigging with a computer, but I love the idea of each band member basically bringing their own interface and monitoring setup.

It's not an afterthought. The problem is modelers do more processing than an audio interface and must use larger internal buffers because of that. That means higher latency. Your Quantum does even less than the minimal processing that most audio interfaces do and that's why it uses extra small buffers and has such low latency.

I'm assuming you know you can use the Helix driver to lower the latency compared to class-compliant, but it will still be higher than what you get with an audio interface.

And I'm assuming you know you can turn up the sample rate and turn down the computer audio buffer size to minimize the latency. In other words, spend computer CPU to lower your latency.

SPDIF might be worth checking out, because it's not a buffered protocol, but you'd have a hard time finding an audio interface that can manage more that 2 SPDIF connections.

 

[Elric]

So you’re asking why modelers are not state of the art audio interfaces? Is it not obvious? That’s not their primary design objective.

 

[Whizzinby]

SPDIF

 

[Sascha Franck]

Late reply, but anyway...

Re: Aggregate devices.
I think that overall latency is possibly higher than what you get from the slowest device, simply because aggregate devices need to enable inter-device audio routing. I have once measured the overall latency of an aggregate device using the output of one and the input of the other device, it wasn't as low as the slower of the two. But that's been a while ago, so things might have changed ever since.
Performance seemed rather stable, though.

So you’re asking why modelers are not state of the art audio interfaces? Is it not obvious? That’s not their primary design objective.

Defenitely not. Yet, given there's interfaces such as the Motu M-series, allowing for very low latencies at super affordable prices, it does make you wonder that some top tier modelers not only perform somewhat sub par but downright horribly bad. The vast amount of the prices you pay for audio interfaces is a result of hardware costs - and that's covered with modelers already. Heck, a Blablaringer U-Phoria interface for 60 bucks is performing VASTLY better than pretty much all high end modelers. Strip all the hardware out and you end up with - what? - 10-20 bucks for the crucial interface part, namely the chipset. Plus a little bit of programming work in case you want to deliver dedicated drivers and not rely on CoreAudio under macOS and that adapted Thesycon BS under Windows.
But what you get instead of even the cheapest interface's performance is a whooping 16+ milliseconds at 32 samples buffersize with the HX series. Seriously, that's embarrassingly bad, especially as it's roughly twice the latency as when running it without installing the dedicated driver (which unfortunately is a requirement for most things you might plan to do).

I absolutely don't expect RME or Lynx alike performance, but performing worse than the cheapest interface you can find is just embarrassing.

 

[EOengineer]

Late reply, but anyway...

Re: Aggregate devices.
I think that overall latency is possibly higher than what you get from the slowest device, simply because aggregate devices need to enable inter-device audio routing. I have once measured the overall latency of an aggregate device using the output of one and the input of the other device, it wasn't as low as the slower of the two. But that's been a while ago, so things might have changed ever since.
Performance seemed rather stable, though.



Defenitely not. Yet, given there's interfaces such as the Motu M-series, allowing for very low latencies at super affordable prices, it does make you wonder that some top tier modelers not only perform somewhat sub par but downright horribly bad. The vast amount of the prices you pay for audio interfaces is a result of hardware costs - and that's covered with modelers already. Heck, a Blablaringer U-Phoria interface for 60 bucks is performing VASTLY better than pretty much all high end modelers. Strip all the hardware out and you end up with - what? - 10-20 bucks for the crucial interface part, namely the chipset. Plus a little bit of programming work in case you want to deliver dedicated drivers and not rely on CoreAudio under macOS and that adapted Thesycon BS under Windows.
But what you get instead of even the cheapest interface's performance is a whooping 16+ milliseconds at 32 samples buffersize with the HX series. Seriously, that's embarrassingly bad, especially as it's roughly twice the latency as when running it without installing the dedicated driver (which unfortunately is a requirement for most things you might plan to do).

I absolutely don't expect RME or Lynx alike performance, but performing worse than the cheapest interface you can find is just embarrassing.

I’m running an aggregate setup pretty consistently and haven’t noticed additional overhead RTL - the figures seem to exactly or nearly match that of the slowest device in the chain. If there is overhead, it’s certainly a much smaller piece of the latency pie. I’d have to take some measurements on my specific hardware to see what I’m getting.

 

[LP59]

Yet, given there's interfaces such as the Motu M-series, allowing for very low latencies at super affordable prices, it does make you wonder that some top tier modelers not only perform somewhat sub par but downright horribly bad.

There's no free lunch. It's all a matter of what you're willing to give up for lower latency. See the explanation above.

 

[Sascha Franck]

See the explanation above.

No, that's irrelevant.

The problem is modelers do more processing than an audio interface and must use larger internal buffers because of that.

This isn't true. The modeling engine and the audio interface could be completely separated. Also, if you can get device latency (that is the modeler under full load) of around 1-2ms (or even <1ms as with the Boss stuff), then you could just slap the cheapest interface on top and would still end up with RTL numbers very well under 10ms. If I run my GT-1000 with both loops engaged serially through my Motu M2 in software monitoring mode, overall latency for 4 complete ADDA cycles and plenty of processing on two devices is still just around 6ms, go figure.

 

[deathbyguitar]

I feel like I'm missing something obvious here. If I'm understanding you correctly, it looks like you're using the Quantum as a live mixer for your Helix and Tonex devices? If that's the case, why not hook these modelers up with analog cables to the Quantum and avoid the aggregate device BS? Do you have 75 other things also sharing the interface when playing live?

 

[LP59]

No, that's irrelevant.



This isn't true. The modeling engine and the audio interface could be completely separated. Also, if you can get device latency (that is the modeler under full load) of around 1-2ms (or even <1ms as with the Boss stuff), then you could just slap the cheapest interface on top and would still end up with RTL numbers very well under 10ms. If I run my GT-1000 with both loops engaged serially through my Motu M2 in software monitoring mode, overall latency for 4 complete ADDA cycles and plenty of processing on two devices is still just around 6ms, go figure.

There is only one buffer system for all the audio, and that includes the safety buffers. That's why the latency is proportional to the buffer size that is required to do whatever processing the device does.

It's not that all the modeler makers are incompetent or are intentionally trying to aggravate you by making the latency unnecessarily high...the latency is simply a consequence of the physics of the situation.

P.S. If all modelers ran at 96K like Boss, they would have lower latency, but less cpu available for processing. Again, there's no free lunch.

 

[sleewell]

im probably doing it wrong but ive owned modelers for years now and barely ever plug into the USB

 

[Sascha Franck]

There is only one buffer system for all the audio, and that includes the safety buffers. That's why the latency is proportional to the buffer size that is required to do whatever processing the device does.

Again, that's irrelevant. See my example. The interface could be completely separated from the modeling circuit. It likely even is, it's just using the same analog connections.

P.S. If all modelers ran at 96K like Boss, they would have lower latency, but less cpu available for processing.

The GX series are running at 48, from all I know. Same latency numbers.
And fwiw, the GT also performs sort of decently as an interface.

 

[JiveTurkey]

I used my SY1000 as an interface AND tone generator and it was awesome. (Well; insert Boss modeling quality comment here obviously)

 

[deathbyguitar]

I have to wonder if OP is making this more complicated than it needs to be. Why not run bass through the Helix alongside guitar, for example?

FWIW USB latency is a non-issue with Axe-FX. I've been using them for years as low latency audio interfaces for e-drums into Superior Drummer 3 while running 2 guitars and bass at the same time. No dropouts or anything. I'd imagine Helix could do something similar with the ammount of I/O on there?

 

[Sascha Franck]

FWIW USB latency is a non-issue with Axe-FX. I've been using them for years as low latency audio interfaces for e-drums into Superior Drummer 3 while running 2 guitars and bass at the same time.

But you're not using software monitoring I assume, but rather the processing of the Axe FX.

 

[deathbyguitar]

But you're not using software monitoring I assume, but rather the processing of the Axe FX.

Superior Drummer uses the Output 1 on the Fractal and I regularly use it with like a 96 sample buffer or lower, I think giving me maybe around 2 to 3 milliseconds round trip? I'll have to double check when I get home.

I'm just really confused about the op's problem in general. What is the ultimate goal, here? Again, I feel like I'm missing some really obvious context.

 

[LP59]

Superior Drummer uses the Output 1 on the Fractal and I regularly use it with like a 96 sample buffer or lower, I think giving me maybe around 2 to 3 milliseconds round trip? I'll have to double check when I get home.

It's higher than that, but this illustrates the issue. Compare that to an FM3, which has less CPU power and so must use larger buffers. Consequently the FM3 has a higher RTL than the AxeFX III.

I'm just really confused about the op's problem in general. What is the ultimate goal, here? Again, I feel like I'm missing some really obvious context.

I believe the goal is to monitor all 3 modeling devices through Logic while avoiding D/A/D conversion. As I mentioned above, an interface with 3 SPDIF inputs would solve his problem, but I don't think such a thing exists, which is why he mentioned using an aggregate.