Moreh claims 2x Nvidia B200 throughput on AMD MI355X
Moreh presented its Token Factory strategy at AMD Advancing AI World Tour Seoul 2026, with its CEO joining an AMD executive on stage for the keynote.
Moreh unveiled a Token Factory software strategy built on AMD GPUs at AMD Advancing AI World Tour Seoul 2026, claiming that its framework delivers more than double the throughput of Nvidia's B200 when running the latest open-source models on AMD Instinct MI355X, according to datanet.co.kr.
Moreh CEO Jo Gang-won appeared in a keynote session alongside Ramine Roane, AMD's vice president of AI product management, to present the approach.
The company said its MoAI Inference Framework combines new-model optimization, generation-specific AMD GPU optimization and datacenter-scale heterogeneous distributed inference in a single software stack.
Moreh also reported results for a mixed configuration in which Nvidia GPUs handle prompt processing and AMD GPUs handle response generation, saying it reached about 1.7x the throughput of a homogeneous GPU cluster.
Earlier AMD work
In 2021, Moreh commercialized an AMD GPU-based AI cloud service with KT, operating more than 2,000 AMD GPUs. This year, the company said it has supplied heterogeneous Token Factory operations software to multiple customers at token factories in Korea and abroad.
Quick answers
What did Moreh present at AMD Advancing AI World Tour Seoul 2026?
Moreh presented its Token Factory software solution for optimizing AI inference costs on AMD GPUs, built around its MoAI Inference Framework.
What performance did Moreh claim on AMD Instinct MI355X?
Moreh said its framework achieved more than double the throughput of Nvidia B200 under typical service conditions when inferencing the latest open-source models on AMD Instinct MI355X.
What did Moreh report for mixed Nvidia and AMD GPU setups?
In a configuration where Nvidia GPUs handle prompt processing and AMD GPUs handle response generation, Moreh reported about 1.7x the throughput of a homogeneous GPU cluster.