Loading…
Schedule as of Oct 11, 2022 - subject to change

Default Time Zone is EDT - Eastern Daylight Time


Back To Schedule
Thursday, October 20 • 10:15am - 11:15am
Real-Time Parallel DSP Processing on GPUs: An Introduction

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!

GPUs have long been considered something of a unicorn in the pro audio industry. The power of their parallel architecture, especially in recent years, is simultaneously exciting and incredibly frustrating; to many it would seem that the ease of which they handle massive amounts of tasks is rivaled only by the difficulty of understanding their architecture, in particular the average dsp developer. Until now, the state of research has always concluded that because of heavy latency and a myriad of computer science issues, DSP on GPUs was just not possible nor preferable. This is no longer the case.

the implications and use-cases are great: from ultra fast plugins and software performance (10x-100x), to cloud rendering infrastructure, GPUs offer a bright future for DSP.

In this talk, we will
Give a basic intro to GPU based audio processing, how it is fundamentally different from CPU, how we made parallelization work in real-time
Unpack our “Scheduler” the “magic” behind enabling low latency performance of DSP algorithms
Break down the parallelization of our Early Access FIR Convolver
Introduce the GPU Audio SDK (pre-release possible)
offer a vision for why this is a viable standard, to power the audio industry of today, and bridge it with the computing platforms of tomorrow.

The State of Research: Why GPU AUDIO ‘Doesn’t Work’
The CPU vs the GPU: architectural differences between SIM and MIM devices, sequential processing versus parallel, and specific problems related to DSP
Graphic depictions of the differences
Mathematics


The Attempts, using old APIs and existing CUDA libraries
Reverb projects
Largely offline rendering projects
Very few instances loaded
Heavy compatibility issues
The original UAD hardware
Previous Conclusion: Fab Filter guy quote “a cow pregnant with triplets still only has x months to give birth” or whatever….

Disruption (Solution):
“we solved it - and we learned some things along the way. The results blew us away and gave us a real sense of what is possible”
A brief overview of “scheduler” the brains behind the GPU Audio engine
Graphic Visualization and Description of how the Scheduler works

A Practical Application: Parallelizing the FIR Convolution Reverb
In May of 2022, we decided to release an “ugly” Early Access plugin in order to test driver and hardware compatibility in order to ensure that our Open Beta will run smoothly. For this test, we chose an algorithm that is particularly friendly to parallelism, partially a nod to earlier attempts at reverb: the FIR Convolver Reverb. With it we demonstrate the power and scalability of GPU Audio processing, both with and without VST3 formats. Early Access also gives us the ability to test upcoming features, such as AMD GPU support, Apple Silicon, Intel GPU support, networked and cloud based DSP, etc.

First, a few bullet points about performance:

- The plugin runs consistently within 1ms latency, regardless of instance count
- No latency increase with live tracking
- Stress test nearly 256 stereo instances of the FIRC plugin in Reaper, with audio playback at 1ms latency.
- Standalone, runs as low as 150 microseconds.

Tear-Down of the FIRC (Alexander Prokopchuk and Markus Steingberger)
[code/component visuals of the DSP algorithm parallelization and data flow]
3-5 point breakdown and comparison of a sequential FIRC versus the parallelized FIRC
A diagram and description of the data flow from CPU<>GPU
A diagram and description of the Scheduler in thread management and instance cascading, handling chains, and multiple tracks

Overview of the SDK and Invitation to the “Workshop”

Value Proposition: From Music to Metaverse.

Finally to conclude our presentation, we wanted to share a vision for what a world of Pro Audio that shares interconnected highways with High Performance Computing (HPC) can look like now, or in the very near future.

The Industry Nexus:

Slide showing NVIDIA’s performance increase over the last decade

HPC existing IT infrastructure is ripe for the picking
GPUs are ubiquitous, and are appearing on most major PC laptops companies
Despite hype over M1/M2 cpus, Apple does indeed have GPUs and we intend to support them

The Potential:

Real-Time DSP is already a hot topic, but with the way content and its delivery is scaling, GPUs offer the perfect low level solution for building the future of Pro Audio infrastructure upon - from Music to Metaverse.
Plugin and DAW software
Local and Remote DSP
AI and Machine Learning Acceleration (run-time!)
Cloud based Processing
Cloud based Collaboration
Interoperability in other industries
Bespoke hardware
Live Audio and FOH redesigns

Speakers
avatar for Andres Ezequiel Viso

Andres Ezequiel Viso

Product Manager, Braingines SA / GPU Audio Inc
I studied Computer Science at the University of Buenos Aires and received my PhD on semantics for functional programming languages. I did a posdoc at Inria, France, in the context of the Software Heritage project, developing the provenance index for the SWH Archive. My interest vary... Read More →


Thursday October 20, 2022 10:15am - 11:15am EDT
3D09