Bricasti M7 Native Plugin

[Andrew Souter]

Even Intel implements audio DSPs functions into their next CPUs - Broadwell.

http://vr-zone.com/articles/meet-intels ... 60323.html

Specifically designed for speech recognition, yes? Maybe it has something interesting for us... Who knows until we try, right?

The basic problem is a business one: pro audio is too small of a market for big money to care about. Sure it would be AWESOME if Intel spent a year to develop a cutting edge audio potato/chocolate/wood/pick-your-fav "chip" that gave us pro-audio guys exactly what we wanted. It is simply very unlikely to happen.

So IMHO, we are best to piggy-back off of larger industries and adapt general hardware to our needs as best as possible... that's my opinion at least...

Maybe we get lucky and "speech recognition" siri/star-trek-computer becomes the killer app, that makes truly powerful DSP chips mainstream and it is easy to adopt for pro-audio use. I would NOT complain. But I would be looking for some large market like this to support it before gambling on it...

[valhallasound]

Even Intel implements audio DSPs functions into their next CPUs - Broadwell.

http://vr-zone.com/articles/meet-intels ... 60323.html

Specifically designed for speech recognition, yes? Maybe it has something interesting for us... Who knows until we try, right?

The basic problem is a business one: pro audio is too small of a market for big money to care about. Sure it would be AWESOME if Intel spent a year to develop a cutting edge audio potato/chocolate/wood/pick-your-fav "chip" that gave us pro-audio guys exactly what we wanted. It is simply very unlikely to happen.

So IMHO, we are best to piggy-back off of larger industries and adapt general hardware to our needs as best as possible... that's my opinion at least...

This has been the way it has been for a while. The Blackfins in the M7 (to get back to the nominal topic) were originally designed as laser printer controller chips. People inside of ADI didn't think they were suited for audio, as the SHARC was thought of as the proper DSP for pro audio. Some VP at ADI came up with a mandate to show that the Blackfin could work for audio, so the small team I was on worked on this for awhile. I'm not sure why this would be a big deal internally; probably something to do with various executives and what chip lines they had "ownership" of.

Right now, the ARMs are coming up as a powerful choice for embedded audio. This is totally piggybacking off of the mobile phone space. In theory, the latest ARM processors would be MONSTERS if entirely dedicated to a single audio process, and fairly cheap to boot.

Sometimes piggybacking off of larger industries backfires. Around 2004, Analog Devices responded to customer demands for a fast SHARC with a smaller pincount by removing the fast external memory pins off of the entire ADSP-2126x line. These were the "latest and greatest" SHARCs, but the external memory access was limited to DMA, which was about 50 times slower than the external memory access of the 2116x SHARCs. A certain reverb developer at Lexicon (who is probably tired of hearing this story ) let me know about how this sucked for his work, in no uncertain terms, at the 2004 AES. I agreed with him on this issue, as I had struggled with writing reverb examples for the 2126x. What I wasn't able to tell him was that the removal of the external memory from the 2126x series was largely driven by demands from the car audio division of Harman, the parent company of Lexicon. ADI eventually released the 2136x DSPs, some of which restored the fast external memory access, but this arrived too late for the PCM96.

Of course, the entire plugin business is piggybacking off of larger industries as well. Plugin developers are subject to the whims of Apple, Avid, Steinberg and Microsoft, and the entire audio industry is a VERY small business consideration compared to the global PC and mobile market.

Sean Costello

[Andrew Souter]

This has been the way it has been for a while.

So notwithstanding the aforementioned anecdote, I guess you are agreeing with me then?

I would go even further and say, it's ALWAYS been this way for the vast majority of the market...

[valhallasound]

This has been the way it has been for a while.

So notwithstanding the aforementioned anecdote, I guess you are agreeing with me then?

Well, I consider off-the-shelf DSPs like the Blackfin "general hardware." There are very few DSPs left that are specific to audio. The ones that come to mind (AL3201, FV-1) I would put in the "fun, but underpowered" category.

As far as the recent Intel and AMD announcements about embedding audio DSP into their chipsets, I don't claim to understand the reasoning behind that. It doesn't sound like it is "generic" audio processing capability. Instead, it is VERY application specific (3D audio for AMD, speech recognition for Intel).

The AMD stuff is probably a side effect of having their processors used in the next-gen consoles. I doubt that they will get much traction in the PC space with their proprietary spatial audio hardware, but it will get utilized in the console space. It reminds me of the Nvidia custom audio hardware for the Xbox, that was available on some PC motherboards in the early 2000s. The hardware was definitely utilized on the Xbox, and died a quick and quiet death in the PC space.

Sean Costello

[Warp69]

It is more difficult these days IMHO. Things simply move faster every day.

But generally, yes, my tendency would be to say it is quite difficult to make such long range predictions at any point in past/preset/future...

I'm actually not so sure that cpu performance will increase with the same rate as now in the future, since focus have shifted to other metrics. We now have 3 generations of Core architecture - Sandy Bridge, Ivy Bridge and Haswell - and the overall performance difference between them are minimal at best. The next generation, Broadwell, won't even be available in a desktop version.

Software is moving way slower than hardware - for every hardware generation we rely more and more on the abilities (automatic vectorizations etc.) of the compilers and high level languages. As Intel and others have hit the law of diminishing returns regarding IPC and start increasing the number of cores and making wider SIMD units, we need to focus more on parallelism in our algorithms and we will eventually hit Amdahl's law. Not many audio products take advantage of multiple cores, AVX and AVX2.

My point is more that "most accurate match" to an exact physical space is not always the goal or even most desirable. Much of the DG lit talks more about psychoacoustics than matching spaces exactly. The ultimate goal IMHO is "whatever sounds best" for the particular musical source. Not trying to match Abbey Roads or Royal Albert Hall perfectly with 100% accuracy.

I agree. It's not always necessary or desirable to match any real space, but my primary goal when designing reverbs is to emulate the different phenomenons happening in real world reverberation. There might be way too much low frequency reverberation, but you have the posibility to eliminate some of it if you desire. The same goes for spaciousness/envelopment and other phenomenons. Those are impossible to add later if the algorithm isn't capable of emulating those abilities.

Regarding idolized hardware of yester-years and the question of matching it in software, take for example 480L. Your own work proves it can be done easily (computationally speaking) in software these days. The alg itself is relatively simple compared to current designs AFAIK. You know MUCH more than me regarding this device, but as I understand it is not computationally expensive, correct? What perhaps IS computationally expensive is going to all the great lengths that you went to model all the EE stuff in the actual hardware box in effort to get as close as absolutely possible to the exact sound as the original.

You're completely right, the algorithm is quite simple and I believe I use 65-75% of the cpu usage on stuff not specifically related to the actual algorithm. Even though we have so much more processing power compared to 1978 (Lexicon 224) and 1986 (Lexicon 480L) the art of reverberation haven't improved much (if at all) - the algorithms can still compete with any plugin reverb and win, except noise floor and perhaps density.

[Andrew Souter]

Software is moving way slower than hardware - for every hardware generation we rely more and more on the abilities (automatic vectorizations etc.) of the compilers and high level languages. As Intel and others have hit the law of diminishing returns regarding IPC and start increasing the number of cores and making wider SIMD units, we need to focus more on parallelism in our algorithms and we will eventually hit Amdahl's law. Not many audio products take advantage of multiple cores, AVX and AVX2.

Yes, agree. Not too many. As it happens, we do...

I agree. It's not always necessary or desirable to match any real space, but my primary goal when designing reverbs is to emulate the different phenomenons happening in real world reverberation. There might be way too much low frequency reverberation, but you have the posibility to eliminate some of it if you desire. The same goes for spaciousness/envelopment and other phenomenons. Those are impossible to add later if the algorithm isn't capable of emulating those abilities.

Me too.

At least for the "normal use" presets. I am a composer/pianist and also enjoy writing (neo)classical music, and generally love purely classical music, and for these applications, these are exactly the type of things I focus on for our verbs/presets.

But I also love making FX-oriented presets too.... I guess simply b/c I also love listening to and producing ambient music, EDM, film scores, game scores, experimental academic computer music, and other adventurous things. And in these things there are pretty much ZERO rules...

You're completely right, the algorithm is quite simple and I believe I use 65-75% of the cpu usage on stuff not specifically related to the actual algorithm. Even though we have so much more processing power compared to 1978 (Lexicon 224) and 1986 (Lexicon 480L) the art of reverberation haven't improved much (if at all) - the algorithms can still compete with any plugin reverb and win, except noise floor and perhaps density.

Well... I might disagree here some, and have lots of customers who have tons more experience than I do with all of these devices who tell me the same, but I am admittedly not an expert on classic devices. I'm a futurist for better or for worse...

I certainly DO have great admiration and respect for all of these great things that came before plug-ins. It's almost unfathomable how some of these things were achieved. I just don't believe that innovation stopped 20-30 years ago...

[Shy]

I certainly DO have great admiration and respect for all of these great things that came before plug-ins. It's almost unfathomable how some of these things were achieved. I just don't believe that innovation stopped 20-30 years ago...

Until very recently, there has been no significant innovation, only variations or minor improvements of existing ideas. The next true innovation will be an implementation of research and various ideas that have never been put together before in a complete package, partially because processing power limitations would have prevented real-time operation of an end-user product. Its capability to simulate the sound of real reverberation with a level of realism far beyond anything prior to it as well as the fact that it's the first reverb simulation ever to really take advantage of the power of current, general purpose CPUs, will be the things that are the most obviously innovative about it.

[Burillo]

We now have 3 generations of Core architecture - Sandy Bridge, Ivy Bridge and Haswell - and the overall performance difference between them are minimal at best. The next generation, Broadwell, won't even be available in a desktop version.

and you got that information where? last time i heard Intel didn't have any plans to scrap desktop chips.

[Andrew Souter]

Until very recently, there has been no significant innovation, only variations or minor improvements of existing ideas. The next true innovation will be an implementation of research and various ideas that have never been put together before in a complete package, partially because processing power limitations would have prevented real-time operation of an end-user product. Its capability to simulate the sound of real reverberation with a level of realism far beyond anything prior to it as well as the fact that it's the first reverb simulation ever to really take advantage of the power of current, general purpose CPUs, will be the things that are the most obviously innovative about it.

You are speaking about something specific? What? How do you know this?

Do you have some background/experience in reverb development or affiliation with another developer? (I don't mean this in any hostile/rude way. Just an honest question, so I understand who I am talking to... so please don't take offense.)

[Shy]

No problem. I'm not a developer, I just have experience with various reverbs/types of reverbs, and have used hardware reverbs (from high to low-end, Lexicon, Quantec, Sony, AMS, Kurzweil, Yamaha, Alesis, Roland and more) and all reverb plugins, heard various examples from Bricasti, EMT and some others I haven't used, and have read all kinds of info online and in manuals about what makes up various reverbs. I'm not affiliated with any developer, but I've been at least a little involved in the development of some products. I couldn't mention anything that hasn't been announced, though.

[Warp69]

and you got that information where? last time i heard Intel didn't have any plans to scrap desktop chips.

Intel road maps. Broadwell (14nm) will be released for portable only - no LGA. Desktop, Enthusiast and servers will get Haswell Refresh (22nm).

The next real upgrade for Desktop will be Skylake

[stian]

Hi everybody!

I have enjoyed following this discussion. I'm just working on a new algorithmic reverb plug-in and I'm looking for some constructive feedback regarding the quality. I thought I'd post here, since there are obviously many good ears around here... The goal is to achieve highly realistic impulse responses while still having the flexibility and modulation capabilities of algorithmic reverbs.

Here is the dry guitar sample processed with settings corresponding to a hall:

http://acondigital.com/audio/Guitar_Acon_Hall1.wav

The left and right impulse responses are available here (modulation effects are lost here, obviously...)

Left:
http://acondigital.com/audio/Acon_Hall1_IR_L.wav

Right:
http://acondigital.com/audio/Acon_Hall1_IR_R.wav

This is still a work in progress and I'll experiment with further modulation techniques to achieve the subtle variations in the dispersion patterns of real acoustic spaces.

Stian

[kvaca]

Hi everybody!I have enjoyed following this discussion. I'm just working on a new algorithmic reverb plug-in and I'm looking for some constructive feedback regarding the quality. I thought I'd post here, since there are obviously many good ears around here... The goal is to achieve highly realistic impulse responses while still having the flexibility and modulation capabilities of algorithmic reverbs.

hi Stian, to me it sounds very close to a convo reverb using real hall IRs, with all pros and cons associated with this type of reverb...
and its hardly believable that your IRs are created with fully algorithmic reverb-when compared to best realworld IRs it still lacks some color and clarity, but who cares if your reverb is as flexible as average algo-based reverb

[stian]

hi Stian, to me it sounds very close to a convo reverb using real hall IRs, with all pros and cons associated with this type of reverb...
and its hardly believable that your IRs are created with fully algorithmic reverb-when compared to best realworld IRs it still lacks some color and clarity, but who cares if your reverb is as flexible as average algo-based reverb

Thanks a lot for your nice feedback! That is what I was hoping for as the plug-in aims primarily at realism. The slight lack of clarity you describe might be related to the modulation I'm using at the moment, so I hope that I'll be able to improve that. Anyway, it is of course possible to modify all these parameters such as room size, levels, modulation settings, early reflection programs, feedback & post filtering and so on.

Stian

[v1o]

Wrong counter-argument. GPUs are there because gaming industry "made them necessary". Computers were running just fine with non-3D accelerated graphics cards long before nVidia, ATi, Matrox, 3dfx, etc.

That's not true at all. GPU's far outperform CPUs at handling 3D graphics. And Computers were not running just fine without 3D non-accelerated graphics, in those days people used highly expensive SGI Workstations for heavy graphics work because PC's were just not up to the task of complex 3D graphics. That includes special effects work for the movie industry, scientific work and the games industry. It wasn't until modern GPU's came out that performance levelled out, to the point that PC's were good enough to use and there was no need for exotic hardware such as the likes made by SGI. Ironically it was mainly former SGI engineers designing the silicon at Nvidida and ATI.

Additionally you could run early 3D games like Quake and Unreal with software-only rendering on the host CPU, but the performance and feature set was always severely reduced compared to running on the GPU. That's because GPUs are massively parallel (an order of magnitude more than CPUs) and have dedicated silicon dedicated specific to the task. They're not just designed for general purpose code. The gulf in performance can be proven by running benchmarks or even writing your own code and trying to do intensive tasks like global illumination, tessellation and self shadowing on a host CPU.