SoftBank, AMD to test GPUs for AI infrastructure

SoftBank and AMD are collaborating to validate the use of AMD Instinct GPUs for next-generation AI infrastructure.

The joint effort focuses on developing functionality to partition and allocate GPU resources based on AI model size and concurrent usage.

SoftBank is enhancing its resource management system, called Orchestrator, to better control multiple AI applications running on a single GPU.

The validation includes testing AMD’s GPU partitioning technology, which allows a GPU to function as multiple logical devices.

A demonstration of this work will be showcased at MWC Barcelona 2026. Both companies plan to continue evaluating the technology’s potential for scalable AI deployment.

SoftBank’s vice president Ryuji Wakikawa stated that the collaboration aims to improve resource efficiency, while AMD’s Kumaran Siva highlighted the goal of supporting diverse AI services with flexible GPU allocation.

The project aims to help meet growing demands for AI processing infrastructure.

🔗 Source: SoftBank

🧠 Food for thought

Implications, context, and why it matters.

SoftBank’s orchestrator was built for a dual AI and telecom role

Orchestrator sits at the center of the AITRAS (AI-RAN) solution and shifts compute resources between AI workloads and virtualized radio access network (vRAN) functions 1, 2.
SoftBank first built and ran the system on NVIDIA accelerated platforms, including NVIDIA GH200 Grace Hopper Superchip-based infrastructure, so one server can run AI plus vRAN workloads 3, 2.
The work with AMD tests another GPU option for the dual-use AI-RAN design, widening hardware choices beyond NVIDIA-based platforms.

This partnership moves toward flexible, multi-tenant AI infrastructure

The collaboration uses AMD hardware partitioning where one Instinct MI300X GPU can split into up to eight independent logical devices through CPX mode 4.
This setup lets cloud providers and large enterprises rent fractional GPUs for smaller inference or development jobs, instead of leasing a full accelerator.
AMD can use the feature to push GPU sharing and virtualization. MLOps (machine learning operations) platforms like ClearML can run several workloads on one physical GPU, and Red Hat positions OpenStack Services on OpenShift for multi-tenant setups that support NVIDIA and AMD GPU accelerators 5, 6.