Practitioner Context
For teams building ML inference pipelines on heterogeneous hardware, or spatial-AI applications that need to run across headset platforms and GPU vendors, open standards convergence is the difference between targeting stable shared APIs and writing bespoke adapters. The Khronos SIGGRAPH Asia 2025 BOF program covered five active areas where that convergence is progressing.
What Khronos Presented (per the Khronos Group blog, December 2025)
At SIGGRAPH ASIA 2025 in Hong Kong, December 15-18, Khronos ran two days of BOF sessions. December 17 opened with an extended "Fast Forward" covering recent progress on OpenXR, ANARI, WebGL, and glTF, followed by a Vulkan and Slang deep-dive. Vulkan remains the primary cross-platform GPU compute API; Slang is Khronos's shading language for neural and differentiable rendering, relevant for teams building real-time ML models on the GPU. Immediately after, the Khronos Machine Learning Council presented "When Graphics and AI Collide: Accelerating AI with Open Standards" - an interactive session soliciting community input on cross-platform inferencing acceleration.
December 18 BOFs covered three more areas. "XR and Spatial Computing: Transforming the Future with OpenXR" highlighted the Spatial Entities extension and spatial perception contributions from PICO. "glTF and Gaussian Splats: The Future of Real Time 3D on the Web" gathered practitioners working on Gaussian Splatting and Volumetric Media in the glTF ecosystem; Khronos announced KHR_gaussian_splatting and KHR_gaussian_splatting_compression_spz experimental extensions at SIGGRAPH 2025 in August in collaboration with OGC, Niantic Spatial, Cesium, and Esri. Finally, the Metaverse Standards Forum BOF explored how Khronos's coalition of 2,600 members is working toward open, interoperable spatial computing across standards organizations.
The ML Inference Thread
The Machine Learning Council session is the highest-stakes item for AI/DS/ML practitioners. A cross-platform standard for inferencing acceleration - analogous to what Vulkan did for GPU graphics compute - would let teams building ML inference on heterogeneous hardware target a common acceleration API, reducing lock-in to CUDA or vendor-specific SDKs. The SIGGRAPH Asia session was a community input phase; published guidance, reference implementations, or extension proposals from the working group are the next deliverables to track.
Gaussian Splats in glTF
For practitioners working on 3D capture pipelines, scene reconstruction, or spatial AI that includes geometric representations, glTF now has an experimental path to interoperable Gaussian Splat assets across web and native runtimes via KHR_gaussian_splatting. Teams building on this format can track extension status in the Khronos registry as it moves from experimental toward ratified.
What to Watch
- •Published extension specifications for Spatial Entities, inferencing acceleration, and glTF Gaussian Splats
- •Reference implementations and conformance suites from Khronos working groups
- •Metaverse Standards Forum outputs on open spatial computing interoperability
Key Points #
- 1Khronos's ML Council is working toward a cross-platform inferencing acceleration API modeled on Vulkan's GPU compute model, which would reduce ML runtime vendor lock-in across heterogeneous hardware.
- 2Khronos and OGC added Gaussian Splats to glTF via KHR_gaussian_splatting experimental extensions, giving 3D capture and spatial-AI pipelines an interoperable web-native asset format.
- 3OpenXR's Spatial Entities extension and PICO contributions signal continued maturation of portable XR APIs for spatial-computing and AR/VR application teams.
Scoring Rationale #
Khronos BOF sessions at SIGGRAPH Asia 2025 cover active standards areas - cross-platform ML inferencing acceleration, glTF+Gaussian Splats, and OpenXR Spatial Entities - making this a solid infrastructure story for spatial-AI and real-time graphics practitioners. Not a frontier-model release or market-moving event, but directly relevant to cross-platform tooling decisions for teams building heterogeneous inference pipelines.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.