HPC384 Spec. I 384kHz over Infiniband in realtime - Alpha Stage

Serve as catalyst for next-generation discrete-time systems

HDMI 2.0 is here and soon audiences worldwide will have the ability to enjoy 1536kHz bandwidth. This magic number can be translated to 8x192kHz channels or 4x384kHz channels. Using HPC technologies, we are, for the first time, able to produce music for these specifications using commodity virtual instruments (VST) and effects. 

 

Analogue tools, synthesizers, acoustic spaces and real instruments consist of complex frequency-dependent nonlinearities. 

To implement those features in a descrete-time system it is necessary to deploy advanced nonlinear differential equations and mathematical transfer functions. For accurate emulation of significant nonlinearities in a virtual environment, the output of those equations may have significantly more bandwidth than the input signal. Until today, the fidelity of those techniques was limited by computational resources. It is time to open the doors of unlimited creativity.

 

Supercomputing technologies along with proprietary advanced algorithms give us the necessary bandwidth to work with thousands of tracks, effect and unlimited polyphony in real time. High Performance Computing eliminates the major bottleneck of how much processing power can be assigned to a track or channel. VST developers can now implement highly advanced quality features that this bottleneck would not permit.

 

Using HPC we can work with advanced physics to model plate reverbs, create evolving non-linear auditorium acoustics and emulate multi-microphone positions that will give sound endless possibilities. It is no longer necessary to work with oversampled peak detection in order to estimate the peak samples on a signal. We have overcome those barriers of conventional underpowered discrete-time systems. We can process the actual audio and not the estimation of it without fighting with conventional CPU or DSP constraints.  There is no way we can overload an HPC music production system when we work with 88.2kHz, 96kHz or 192kHz. Moreover, HPC allows us to have different sound qualities in the same project so we can push hard when you want to emulate analog synths or luscious reverbs. Our lab tests indicate that we are now ready for accurate solid-state circuitry emulation that needs advanced resolution at a microsecond's time domain. 

 

At this critical juncture of entertainment evolution, with 3D & HDR, IMAX Cinema, Dolby® Atmos, DTS® Headphone X, 6K Cinema , 4K TV with HDMI 2 and In-Vehicle Infotainment, the industry creates a roadmap for a quality aware audience.  A true quality upgrade of the overall cinematic experience is on-going. HPC384 Spec. is here to keep music production on par with those innovations and it will provide the necessary tools, specifications and revolutionary techniques so that music professionals will be able to produce and deliver high quality content to meet the demands and expectations of their audience.

 

Thank you for visiting us. You can find interesting Bibliography at the end of the page.

Below you can find some tests results for initial proof-of-concept validation: 

Personal computing bottlenecks VS HPC

High Performance Computing at 384kHz

Dr. Fritz Riehle wrote, "Of all measurements quantities, frequency respesents the one that can be determined with by far the higher degree of accuracy (Riehle, Fritz. Frequency Standards. Die Deutsche Bibliothek, 2004).

HPC384_Partners_HPCmusic.jpg

Tests and Measurements - The sound differences to date

On another test, we render the first ever reverb at 1536kHz using U-He Zebra 2 (HS Eternal Chimes) clocked at 384kHz as our sound generator. This sound is quite likely the most mathematicly complex and harmonically rich single sound ever created in the digital domain. More importantly, it was created using Native VST instrument and effect technologies. 

 

We also carefully measure latency and network overload. 1536kHz is completely experimental and it represents the potential that high performance computing brings into music production. Currently, we conduct tests of digital sound creation up to 6144kHZ/24bit PCM using AAS Ultra Analog  Virtual Synthesizer, Ircam Verb 3DMG Audio CompassionSoftube Pultec equalizer and 2CAudio reverbs. 

 

Software developers will gradually take advantage of the massive computational resources of HPC systems and the sound quality gain of those systems will become even larger.

44.jpg

384a.jpg

44.1kHz

384kHz downsampled to 44.1kHz

Spectral Analysis of 384kHz at 192kHz inside Adobe Audition CS6

384 to 192.jpg

Download uncompressed audio file
What do high performance computing for music production and high quality processing algorithms actually mean for music and the audience?


1. Composers and songwriters have a new 3D sound and multi-texture palette to create with. In the past, we discovered great new instruments like the electric guitar and the analogue synthesizer, resulting in the emergence of new genres of music. Today, HPC for music is a hyper-instrument that facilitates the creation of new sounds, timbers, textures, virtual acoustic spaces and eventually new music.


1a. An unlimited or hyperscaled amount of computational resources translate to an infinite amount of tracks/instruments for         professionals to shape music as well as an infinite amount of effects to shape sound.

 

1b. Offline Rendering speed gain more than an order of magnitude on complex orchestral or commercial pop productions.

 

1c. Advanced physics modelling, analogue circuit simulators, detailed simulating of non-linear / non stationary characteristics.


2. It could put an end to the well-known loudness war, which degrades the sound quality of modern music. For more than a decade, mastering engineers have been 'forced' to compete in the 'average loudness' of music due to lack of other sound ways of differentiating their products. As a result every new hit record has less dynamic range than the previous one. By pushing the sound quality to the next level, engineers and producers will care less about loudness and more about spatial sound information, texture, timbre and sound quality. Competing on different sound aspects can lessen the industry's focus on loudness.


3. Extensive digital sound processing would not result in poor sound quality; instead, it would be elegantly refined. This is akin to photo processing, in which the processing of the large raw file looks better than processing the jpg. Low quality processing creates digital artifacts like aliasing.

4. High quality multichannel audio can allow mixing engineers and producers to utilize advanced surround mixing and give algorithms like DTS Headphone X and Dolby Atmos a true value.

5. In a movie world that innovates with 3D and HDR picture technologies, sound and music will eventually find its place on par with those high quality developments.

Low Latency Supercomputing for Music
 

Mellanox RDMA protocols and offload engines are fully interoperable with standard TCP/UDP/IP stacks, which in common parlance means that even when higher bandwidth create huge demands on the network, Infiniband is able to maintain realtime performance of virtual instruments and effects at 384kHz. State of the art interconnect enables us to implement ideal load balancing so that we can make use of advanced software algorithms like 2CAudio Band-Limited Interpolation in real-time. HPC384 specifications have now determined most of the CPU, network and power usage for producing music at both 192kHz and 384Khz.

Cost per GFLOPS

 

HPC for music cost per GFLOPS is roughly 35x better than AVID (NASDAQ:AVID) Protools HDX DSP-accelerated solution.

 

Steinberg GmbH underlines that "When choosing a processor, please be aware that floating point operations are crusial for audio performance". Here is updated look of FLOPS (FLoating-point Operations Per Second) on processors and systems.

Click here to download the official Intel Performance PDF

The degree to which the processing can add capacity without disruption and without incurring excessive overhead (nonproductive processing) is largely determined by the scalability of the particular computing platform. There are many solutions that try to deliver scalable computational power by using dedicated DSP. The two most popular ones come from Universal Audio and Avid Technologies. One Universal Audio UAD2 SHARC DSP (Analog Devices ADSP-2136) delivers 2.4 GFLOPS and their flagship PCIe DSP card puts together 8 of them for $1500 (19 GFLOPS).  An Avid Protools PCIe HDX system delivers 38 GFLOPS and the entry level HDX DSP PCIe card (18 Ti C672x Chips running at 350MHz ) costs approximately $7500.

 

An Intel i7-4770K based node using Mellanox interconnect can deliver 177 GFLOPS (Floating point performance, Linpack benchmark using the Intel MKL library) and cost $1000 (12/2/2013).

 

 

Even by rough calculations and without taking into account the economies of scale that HPC systems can deliver, value for money of HPC for music is at least 14x better than Universal Audio card and 35x better than Avid Protools HDX card.

 

Dynamic resource allocation and load balancing allow more than 80% CPU utilization on HPC systems.

Form Factors

 

HPC384 clusters for music production can scale up to a huge supercomputer but they do not necessarily have to be huge systems inside machine rooms. More info about form factors soon... :)

HPC+Advisory+Council_Cluster_Center_HPCmusic_Antonis+Karalis+(1).jpg

 
A brief Bibliography

 

Technion - Israel Institute of Technology

Performance analysis of dual source transfer-function generalized sidelobe canceller, Jan 2006

 

KTH University

A. Askenfelt, ed., Five Lectures on the Acoustics of the Piano, Stockholm: Royal Swedish
Academy of Music, lectures by H. A. Conklin, Anders Askenfelt and E. Jansson, D.
E. Hall, G.Weinreich, and K.Wogram, Oct 1990

 

Stanford University

S. Bilbao, Wave and Scattering Methods for the Numerical Integration of Partial Differential Equations,
PhD thesis, Stanford University, June 2001

 

Technion - Israel Institute of Technology

In-kernel Intergration of Operating Systems and InfiniBand Primitives for High-Performance Computing Clusters, Sep 2005

 

IEEE

S. Bilbao, “Time-varying generalizations of all-pass filters,” IEEE Signal Processing Letters, May 2005

 

AES - Audio Engineering Society

O. J. Bonello, “Modular parametric equalizer-filter, a new way to synthesize the frequency
response,” Audio Engineering Society Convention, Preprint 1170, Oct. 1976

 

Electronics Letters Vol27

P. Bowron, M. R. J. Motlagh, and A. A. Muhieddine, “Harmonic characterisation of feedback
systems incorporating saturation nonlinearities, Oct 1991

 

Tokyo Institute of Technology

PHoM — a polyhedral homotopy continuation method for polynomial systems, Dec 2002

 

IEEE Cluster 2013 Conference

Influence of InfiniBand FDR on the Performance of Remote GPU Virtualization, Sep 2013

 

IEEE 18th International Conference on Parallel and Distributed Systems

Comparing the Performance of Blue Gene/Q with Leading Cray XE6 and InfiniBand Systems, Dec 2012

 

IEEE Cluster 2013 Conference

Distributed Resource Exchange: Virtualized Resource Management for SR-IOV InfiniBand Clusters, Sep 2013

Press Resources - SC13 Supercomputing Conference Poster as seen at HPC Advisory Council Booth