arzh-CNenfrdejakoplptesuk
Search find 4120  disqus socia  tg2 f2 lin2 in2 X icon 3 y2  p2 tik steam2

AMD Introduces Instinct MI350: 20 PFLOPs, 185 Billion Transistors, and Up to 36TB of HBM3e in a Rack

AMD has officially announced Instinct MI350 - new generation GPU for AI and HPC, based on architecture CDNA 3 and TSMC 3nm process technology. The flagship MI355X contains up to 185 billion transistors and supports FP6/FP4 computing with peak performance up to 20 PFLOPs.

instinct

MI350X and MI355X are equipped with HBM3e with a capacity of up to 288 GB and provide throughput of up to 8 TB/s. A single node can use to 8 GPU, and in the server rack - up to 128 accelerators and 36 TB HBM3e. AMD also introduced the new UBB8 standard for quick integration of liquid and air cooling.

MI355X FP64 performance up to 79 TFLOPs, and in FP16 - 5 PFLOPs. Compared to MI300X, the gain in AI tasks (for example, Llama 3.1) reaches 35× by throughput. Also recorded Up to 2.2× better performance than B200 and GB200 in FP6 computing, despite the lower price.

MI355X is 1.2–1.3× faster than B200/GB200 in DeepSeek, Llama and token generation tasks, and a series MI400 is already in development and will be released in 2026.