HAMi Beijing Meetup: Optimize Efficiency, Not Scale

The recent HAMi Meetup Shanghai left us with memorable technical discussions and candid feedback. The message from the community and enterprise practitioners was clear:

As domestic heterogeneous compute evolves rapidly, the real challenge isn't "do we have compute," but "can we use it efficiently, reliably, and operationally."

Building on this consensus, the HAMi Meetup Beijing is coming up—another opportunity to explore how compute can be used more effectively.

Beijing, a city that brings together research institutions, industry clusters, and cutting-edge technology, is becoming a key hub for compute efficiency innovation. HAMi will also enter its next phase here—one that's more aligned with real-world enterprise deployment.

As deep contributors and long-term participants in the HAMi community, RiseUnion will again join the Beijing meetup, working alongside community members, chip vendors, and enterprise users to tackle real-world deployment challenges for domestic heterogeneous compute. This will be a "no fluff, just practice" deep dive.

From Community to Production: What Real Problems Is HAMi Solving?

HAMi is the only open-source project in the CNCF ecosystem focused on heterogeneous GPU resource sharing. Maintained by 350+ contributors from 15+ countries, it's been validated in production at 200+ enterprises and institutions.

In real deployments, we're seeing more and more organizations face similar challenges:

• Multiple domestic GPU / NPU / DCU / XPU architectures coexist, dramatically increasing scheduling complexity;
• Inference workloads are highly fragmented, leading to persistently low memory and compute utilization;
• Chip characteristics vary widely, creating a huge gap between "works" and "works well";
• Scheduling decisions aren't observable, making problems hard to diagnose and operations costly.

What HAMi does isn't replace existing scheduling systems—it's about making heterogeneous compute truly visible, divisible, schedulable, and reusable through lightweight, non-invasive, pluggable GPU virtualization capabilities.

This is exactly why RiseUnion has been deeply involved in the HAMi community for the long term.

RiseUnion Session Preview

Accelerating Domestic Compute Compatibility in HAMi v2.7.0

At this HAMi Meetup Beijing, RiseUnion R&D Engineer & HAMi Reviewer Ouyang Luwei will share insights from real engineering practice.

What We'll Cover:

1️⃣ Kunlunxin P800 vXPU Dynamic Partitioning Deep Dive

We'll break down vXPU's fine-grained resource partitioning capabilities, systematically explaining how it works in 1/4-card, 1/2-card, and other virtualization scenarios, along with scheduling coordination mechanisms. We'll show the real benefits of running multiple tasks in parallel on a single card for lightweight inference, development, and testing.

2️⃣ The Engineering Value of XPU Topology-Aware Scheduling

Drawing on Kunlunxin P800's physical topology characteristics, we'll share how to balance performance stability and resource utilization through scheduling strategies like topology optimization and health assessment in multi-card training and complex heterogeneous environments—avoiding the "fine-grained but slow" trap.

3️⃣ Visualizing Scheduling Decisions for Faster Troubleshooting

In complex scenarios combining heterogeneous compute and fine-grained partitioning, scheduling issues often happen fast but take forever to diagnose. We'll share how to visualize scheduling decision paths for rapid problem identification and analysis, helping organizations build truly observable and operational compute systems.

This isn't lab theory—it's real-world experience from RiseUnion's deep integration work with Kunlunxin and other domestic chips, directly applicable to enterprise practice.

Why Attend?

The HAMi Meetup Beijing will cover multiple dimensions—from community standards and chip capabilities to platform engineering and business practice:

• HAMi's new features and capability matrix standardization;
• DCU / XPU / GCU software virtualization and Kubernetes practices;
• Production experience scaling vGPU inference clusters;
• Optimizing AI software stacks with virtualization scheduling;
• Next-generation resource abstraction with HAMi-Core × DRA.

Engineers from CNCF, Migua Intelligence, Hygon, Qingcheng Jizhi, Ke.com, 4Paradigm, and RiseUnion will deliver high-density, reusable engineering insights.

This is a serious technical gathering focused on getting the most out of compute.

Event Details

• Date: Saturday, December 27, 2025, 1:30 PM–6:00 PM
• Location: 5th Floor, Building 12, Zhongguancun Innovation Street, Haidian District, Beijing
• Registration: Scan the QR code below for free registration (limited seats)
  

Closing Thoughts

Domestic compute is evolving from "available" to "usable, reliable, and sustainable." This journey depends on open-source community collaboration and continuous feedback from real-world engineering practice.

RiseUnion will continue our deep involvement in the HAMi community, working with more partners to bring heterogeneous compute scheduling—this "hard but right thing"—into production environments.

See you in Beijing on December 27.

Let's talk efficiency, not compute wars.