Pegatron introduced 128-GPU AMD Instinct MI350X-based system — up to 1177 PFLOPS and 36,8 TB of HBM3E memory
The exhibition Computex 2025 Pegatron Shows Unique Rackmount Solution, which uses 128 AMD Instinct MI350X accelerators, designed for resource-intensive AI inference and HPC computing tasks. This large-scale system is ahead of AMD's proprietary rack-scale platforms by a generation and will form the basis for future solutions based on the MI450X, which are expected next year.
Model Pegatron AS501-4A1/AS500-4A1 built on eight 5U modules, each equipped with a processor AMD EPYC 9005 and four MI350X accelerators. All computing components cooled by liquid, providing stable performance under high loads. The chassis is made in the format 51OU ORV3, making it compatible with data centers built to standards OCP (including Meta).
The overall system performance reaches 1177 PFLOPS in FP4 mode, offering near-linear scalability. Each MI350X accelerator is equipped with 288GB HBM3E memory, which in total gives 36,8TB of ultra-fast storage — more than current Nvidia Blackwell-based solutions, allowing it to process AI models that don’t fit in other memory GPU- system.
However, the platform does not use AMD's own Infinity Fabric switches, as it is implemented by Nvidia with NVLink, but relies on communication via 400 GbE Ethernet between cases. This limits the maximum scale-up architecture to 8 GPU, reducing efficiency in highly synchronized tasks such as training large LLM models.
However, Pegatron's solution shows how you can build a productive and memory-rich GPU-cluster based on AMD Instinct, suitable for parallel inference, scalable tasks and architectural prototyping. The system is already used within the OCP community and serves as a basis for developing future solutions based on AMD Instinct MI450X, including the upcoming IF64 and IF128 configurations.
