On a dual-GPU machine, after HAMi divided each GPU into 10 partitions, eight instances sharing the compute resources showed a performance difference of only 0.2%. But once GPU core limits were enabled, throughput fell by 48% and tail latency rose from 1.34 seconds to 4.77 seconds. Our judgment is that the technical challenge in GPU sharing is no longer whether partitioning is possible, but how to design core-limiting policies and ensure scheduler visibility.

What this is

HAMi is one of China’s more mature open-source GPU orchestration projects. It can partition a GPU’s memory and compute cores among multiple containers and bring accelerator cards from different vendors—including Ascend, Cambricon and Moore Threads—into Kubernetes’ unified resource model. GPU orchestration, which means allocating, reclaiming and isolating multiple GPUs according to workload needs, was previously associated mainly with NVIDIA infrastructure. It is now moving toward an environment where multiple chip architectures coexist.

The author installed HAMi 2.11.0 on a dual-GPU machine with an RTX 4090 and RTX 5090, divided each card into 10 partitions, and then ran eight real-world workloads alongside nine acceptance tests. Memory quotas could be specified down to the byte, and containers could not request memory beyond their quotas, showing that basic isolation is already usable.

Industry view

This test supports a clear conclusion: as enterprises adopt heterogeneous compute, the scheduling layer is becoming a critical part of infrastructure. A GPU is no longer an indivisible device; it must be measured, allocated and reclaimed like memory and CPU. HAMi’s open-source, multi-vendor approach can also help enterprises reduce their dependence on a single GPU ecosystem.

But the risks are equally clear. First, core limiting reduced throughput by 48%, demonstrating that successful sharing does not necessarily mean worthwhile performance. Enterprises must distinguish memory sharing from compute-core restrictions and benchmark each workload type accordingly. Second, the scheduler cannot see memory consumption outside the scope it manages. It may dispatch work to a GPU that is already close to full capacity, ultimately triggering an out-of-memory failure. HAMi is currently more of a resource-management foundation than a complete solution that makes capacity planning and performance testing unnecessary.

Impact on regular people

For enterprise IT: Once GPUs from multiple vendors enter the data center, resource orchestration will directly affect utilization and fault isolation. Enterprises should first inventory their actual chips and workloads, then decide whether to introduce a sharing layer.

For careers: Roles in algorithm development, operations and platform engineering will encounter GPU resource quotas, container scheduling and heterogeneous compatibility more often. These skills are shifting from specialized expertise to foundational infrastructure knowledge.

For the consumer market: The near-term impact will be limited. GPU cloud service prices may fall as sharing efficiency improves, but performance fluctuations caused by core limiting could create a more pronounced experience gap between low-cost shared instances and dedicated instances.