Help Drive the Evolution of Machine Learning Acceleration Standards
Khronos opens Machine Learning Forum for anyone to join. Initiative seeks to enable system developers to deploy ML algorithms and applications efficiently across multiple platforms.
The speed of innovation within the machine learning ecosystem is faster than ever before, and shows no signs of slowing with new architectures regularly hitting the headlines. While such innovations are fantastic in driving performance and functionality, the sheer diversity of hardware and software solutions poses a challenge for developers hoping to support the latest advancements, while also reaching the widest audience.
At Khronos, we know that open and royalty-free standards can play an essential role in reducing fragmentation, removing industry friction, reducing costs, and providing everyone the opportunity to grow. And this chimes with the growing feedback from ML developers who lament the growing fragmentation in the ML ecosystem.
One of the challenges for any standardization effort addressing ML acceleration is the sheer breadth of silicon solutions and applications available. But the whole ecosystem can become a winner if standards can evolve that enable ML system developers to deploy algorithms and applications efficiently across multiple platforms.
We’ve Been Here Before
Khronos has been working as an industry consortium for over twenty years, and understands how important it is to provide a safe space for a broad spectrum of organizations to come together to discuss the challenges, define the requirements and work to evolve standards based solutions to address fragmentation. This safe space approach has most recently resulted in the successful formation of the Khronos Camera and 3D Commerce working groups, both of which bring together a diverse range of often competing voices to solve common problems. So, can something similar be achieved for the benefit of the machine learning community?
Leading ML Experts Share Their Thoughts
On October 20, 2021 the Khronos Group invited a group of machine learning experts to an online Machine Learning Summit to kick start discussions on how ML acceleration standards could evolve. During the meeting participants were invited to share the key issues they face and identify the opportunities for ML acceleration and interoperability. Over 30 organizations took part with six sharing their experiences. The feedback received was unvarnished; and those attending were not afraid to express their pain points, and what would make development easier.
During the summit several of the presenters shared the challenges they face accelerating ML algorithms and applications when using existing Khronos APIs such as OpenCL, SYCL, Vulkan, NNEF and OpenVX.
|Low-level, low-overhead cross-platform GPU API.||Low-level parallel programming of heterogeneous processors.||Single source C++ abstraction layer for heterogeneous processors.||Computer vision acceleration API with ML extension and tensor objects.||Exchange Format for connecting trained networks with inference engines.|
|Learn More||Learn More||Learn More||Learn More||Learn More|
In response to this feedback a follow-up summit was held in January 2022 where experts from the various Khronos Working Groups responded to the issues raised. Providing solutions were available, or taking the feedback to guide future updates to the standards.
The video recording and presentations from the summit are available online.
- October 2021 Summit: Session #1 Video - Session #2 Video
- January 2022 Summit Response: Presentation Slides
Based on feedback from the summit and subsequent discussions there appears to be a clear demand for an industry-wide effort to guide the next steps in the use and evolution of open standards that support ML-centric acceleration.
“Judging by the significant industry interest, the time seems right to organize an effort around identifying the requirements and steering the evolution of interoperability APIs that meet the industries needs for cross platform ML acceleration. This is a topic that is very relevant to Khronos as our acceleration APIs such as OpenCL, SYCL, OpenVX and Vulkan are often used to accelerate machine learning.” said Pierre Boudier, chair of the Khronos Machine Learning Task Sub Group.
We don’t expect there to be one API to rule them all, but we do see clear benefit in creating a coalition of interested parties (a Khronos ML forum) that will be focused on meeting the needs of the ML community for hardware acceleration.
Step One - Making better use of what is available now
It was clear from ML summit discussions that developers are often unaware or unclear on the best way to use the ML-acceleration features that are already available in the current generation of Khronos APIs and associated extensions. To address this issue we are planning to run a series of tutorials where a panel of experts will share the latest information and best practice, and answer questions from the audience. Our first webinars will be:
- May 5, 2022 - Accelerating ML with Vulkan
- May 11, 2022 - Accelerating ML with OpenCL
- To be announced - Accelerating ML with SYCL
- To be announced - Accelerating ML with OpenVX and NNEF
Step Two - Khronos Launches ML Forum to Help Drive Machine Learning Acceleration
As a next step the Khronos Group now invites interested companies and individuals to join the Khronos ML Forum. The forum will provide the opportunity for participants to discuss the challenges they face with existing APIs and cross-platform development, and have Khronos members respond with solutions. These solutions could include, but are not limited to; advice and guidance on using the APIs, evolution of the existing standards and extensions, bug fixing of the standard, project to address gaps in the ecosystem, or feedback to implementers to encourage consistent cross-platform support of ML-acceleration features.
The Khronos ML Forum
- Open to any company, university, consortium, open-source participant, or industry expert
- Free to join
- All discussions are in the open (no NDA required)
- All participants have an equal voice
- A dedicated Slack workspace for discussions
- Email distribution list for announcements and meeting invites
- Quarterly ML Forum Summit open to all members, starting July 2022.
Get Involved - Join the ML Acceleration Forum
We invite all interested parties to join the Khronos ML Forum today and help guide the evolution of the next generation of machine learning acceleration standards.