Overview
Central Java carries the frontier tier: AMD Instinct MI300X with 192 GB of HBM3 per GPU, coupled by Infinity Fabric. It is the first MI300X deployment in the AGATE GREEN estate outside Malaysia, and at 1.3 MW it is deliberately a focused block rather than a campus.
The tier exists for a different workload class than the RTX fleet. Large-scale pre-training, tensor-parallel inference and HPC are constrained by inter-device bandwidth rather than memory per card, and that is what the site is designed around.
Power & site
1.3 MW is what the local interconnection supports today, and the site is built to that number rather than to a wish. A smaller block with a real feed beats a larger one waiting on reinforcement.
Air cooling is engineered into the module envelope from the start — at this power density it is part of the design, not a retrofit.
What runs here
- Pre-training runs
- Tensor-parallel inference
- HPC and simulation
- Overflow training for Cyberjaya tenants