Mirendil Conducts Pre- and Post-Training Using TPU and GPU on AI Hypercomputer

Mirendil is adopting Google Cloud's AI Hypercomputer to accelerate AI development. It plans to perform model pre-training and post-training by jointly leveraging Google's TPU AI accelerators and NVIDIA's full-stack AI infrastructure running on Google Cloud.

Mirendil is building an end-to-end research workflow spanning from initial model pre-training to post-training and large-scale reinforcement learning. The company explains that it can select TPU and NVIDIA infrastructure based on workload characteristics, allowing it to quickly secure necessary computing resources and apply architecture suited to workloads in the long term.

Google Cloud and Mirendil have designed integrated infrastructure encompassing compute, storage, network, and control plane. Provisioning and management of TPU and GPU environments have been simplified through the managed training clusters of the Gemini Enterprise Agent Platform, and Mirendil is currently operating a TPU v5P cluster. NVIDIA AI accelerated computing systems are expected to be added soon.