AMD has officially unveiled its groundbreaking Instinct MI350 series, featuring a next-generation architecture built on TSMC’s 3nm process. The new chips come with up to 185 billion transistors and come in two variants: the MI350X and the MI355X. These powerful GPUs are engineered for AI and HPC workloads, offering outstanding performance and cutting-edge features. The MI350X and MI355X incorporate AMD’s CDNA 4 architecture, featuring a massive 288GB HBM3e memory and support for FP4/FP6 AI data types. 
With these innovations, the MI350 series promises to disrupt the high-performance GPU market.
In terms of performance, the MI350 series chips offer up to 20 PFLOPs of FP4/FP6 compute, marking a 4x improvement over the previous generation. The MI355X variant, the flagship model, boasts 8TB/s of aggregate memory bandwidth, 79 TFLOPs of FP64, and 20 PFLOPs of FP6/FP4 compute, all in a 1400W package. These new GPUs also feature 256 MB of Infinity Cache, enhancing data transfer speeds and improving overall performance. The chips are designed with 256 compute units across 8 zones, each housing 32 compute units, and come with dual IO dies based on TSMC’s N6 node.
When comparing the MI355X to NVIDIA’s B200, the MI350 series outperforms the competition in several metrics, including memory bandwidth, FP64, FP8, and FP6. The MI355X offers 1.6x higher memory capacity and significantly higher FP64 performance. AMD’s new chips also show impressive gains compared to the MI300 series, with up to 35x improvements in inference performance using the Llama 3.1 405B workload.
AMD also demonstrated the impressive scalability of the MI350 series, with up to 8 GPUs capable of delivering 2.3TB of HBM3e memory and 161 PFLOPs of FP6/FP4 compute performance. The full rack system, featuring 128-96 MI350 GPUs, will offer up to 36TB of memory and 2.6 Exaflops of FP4 compute performance. The Instinct MI350 series is expected to be available through AMD’s partners starting in Q3 2025, with the MI400 series already in the works for a 2026 launch.