AMD CDNA 4 Architecture · 16,384 Stream Processors · 1,400W TDP · PCIe 5.0 · Infinity Fabric · 8 TB/s Bandwidth
Instinct MI355X
AMD CDNA 4
TSMC 3nm / 6nm
185 Billion
288 GB HBM3E
8 TB/s
256 MB
16,384
256
1,024
2,400 MHz
10.1 PFLOPs
5 PFLOPs (10.1 w/ Sparsity)
2.5 PFLOPs (5 w/ Sparsity)
157.3 TFLOPs
78.6 TFLOPs
1,400W
OAM Module
Passive & Active Cooling
PCIe 5.0 x16
Infinity Fabric
7 Links · 153 GB/s
1 per GPU · 128 GB/s
Full-chip
Yes
The AMD Instinct MI355X is AMD's flagship data center GPU built on the 4th Gen CDNA 4 architecture. Fabricated on TSMC's advanced 3nm/6nm process with 185 billion transistors, the MI355X delivers breakthrough AI and HPC performance with 288 GB of HBM3E memory — the highest single-GPU memory capacity in the industry — and 8 TB/s of memory bandwidth.
With 16,384 stream processors, 256 compute units, and 1,024 matrix cores, the MI355X achieves 10.1 PFLOPs of MXFP4/MXFP6 AI performance and 2.5 PFLOPs of FP16. The expanded MXFP6 and MXFP4 datatype support enables efficient large language model inference and training at scale. At 1,400W TBP, the MI355X is designed for direct liquid cooling (DLC) in high-density AI clusters.
AMD Infinity Fabric with 7 links per GPU provides 153 GB/s of scale-up bandwidth, while the scale-out link delivers 128 GB/s per GPU for cluster-wide connectivity. The MI355X launches alongside the ROCm 6 software ecosystem with full support for PyTorch, TensorFlow, and JAX. Launched June 2025, the MI355X competes directly with NVIDIA B200 in the frontier AI training market.
QS Compute can source AMD Instinct MI355X for enterprise and cloud deployments across APAC. Contact us for cluster pricing and availability.
Need AMD Instinct MI355X?
Enterprise hardware — configured, tested, deployed. Contact us for pricing and availability.
Request Quote