AMD CDNA 4 Architecture · 3nm Chiplet · 8 TB/s Bandwidth · AI Inference & Training
Instinct MI350X
AMD CDNA 4
TSMC 3nm / 6nm
288 GB HBM3e
8 TB/s
Up to 20 PFLOPS FP4/FP6 (series)
Up to 9.2 PFLOPS
~2.3 PFLOPS
~750W (air-cooled)
OAM Module
PCIe 5.0 / Infinity Fabric
2025
The AMD Instinct MI350X is the air-cooled member of the MI350 series built on AMD's fourth-generation CDNA 4 architecture. Fabricated on TSMC's 3nm/6nm chiplet process, it pairs 288 GB of HBM3e at 8 TB/s with expanded FP4/FP6 datatype support for efficient large-model inference and training.
The MI350 series delivers up to a 4× generation-on-generation uplift over MI300, reaching up to 20 PFLOPS of FP4/FP6 compute. The MI350X positions as a competitive alternative to NVIDIA's B200 for inference-heavy and mixed AI workloads, running within the open ROCm 6 software ecosystem.
QS Compute can source AMD Instinct MI350X accelerators for enterprise and cloud AI deployments. Contact us for cluster pricing, reference designs, and availability.
Need AMD Instinct MI350X?
Enterprise hardware — configured, tested, deployed. Contact us for pricing and availability.
Request Quote