SoftBank and AMD have begun validating GPU partitioning technology for AI workloads using AMD Instinct GPUs, allowing a single GPU to handle multiple AI tasks simultaneously. SoftBank developed an Orchestrator system that divides AMD Instinct GPU resources based on workload requirements, including model size, concurrent executions, and memory needs.

The system splits compute workloads across multiple GPU instances running on individual Accelerator Complex Dies (XCDs), ranging from single instance mode (SPX) to eight instances (CPX mode). HBM memory pools are divided into individual regions for each GPU instance to prevent latency spikes. The goal is to eliminate the inefficiency of uniform GPU resource allocation, which can lead to GPU shortages or waste depending on workload demands.

SoftBank claims the enhanced Orchestrator runs multiple AI applications on a single GPU with minimal resource strain, though no performance figures have been shared yet. The company plans to explore similar orchestration for other AI accelerators beyond AMD and will demonstrate the technology at the AMD booth during MWC Barcelona 2026 in March. This development comes as AMD faces manufacturing delays with its next-gen Instinct MI455X accelerators, with only limited production expected this year and mass production pushed to Q2 2027.