
MangoBoost and Preferred Networks to test heterogeneous AI inference architecture
The AMW Read
The partnership proposes a specific heterogeneous hardware architecture, but its feasibility and performance remain unverified.
MangoBoost and Preferred Networks to test heterogeneous AI inference architecture
MangoBoost and Japan's Preferred Networks plan to collaborate on infrastructure that divides AI inference work across different processors. Under a memorandum of understanding announced for exchange at SC26 in Chicago, the companies intend to combine GPUs and Preferred Networks' MN-Core processors with MangoBoost's data processing units and memory and storage optimization technology. The proposed design assigns prompt processing, or prefill, to GPUs and token generation, or decode, to MN-Core processors.
The collaboration adds a proposed hardware configuration to the AI infrastructure market, where inference systems must coordinate compute, networking, memory, and storage. Separating prefill and decode gives each stage its own processor role, making the connections between those components central to the design. The companies intend to test whether this architecture is technically feasible; the announcement does not establish that it improves performance or lowers costs.
For builders and investors, the next evidence will come from system integration tests in a simulated environment. The partners plan to expand performance testing to physical equipment once samples of Preferred Networks' MN-Core L1000 are available. They may also publish joint research papers and technical materials, depending on the evaluation results. Until those tests are complete, the partnership is a research commitment rather than evidence of a production-ready system.
#AIInfrastructure #AIInference #Semiconductors #HeterogeneousComputing #MangoBoost