AMD Unveils MI350 AI Chips, Claiming Edge Over Nvidia Blackwell

Advanced Micro Devices (AMD) has launched its latest AI chip lineup, the Instinct MI350 series, aiming to rival Nvidia’s Blackwell B200.

The new series includes two models: the high-performance MI355X, which uses liquid cooling, and the more thermally efficient air-cooled MI350X.

Designed for scalable deployment, the MI350 supports up to 128 GPUs in liquid-cooled racks, delivering up to 2.6 exaFLOPS of FP4 performance.

Built using a 3D, 10-chiplet design, eight compute chiplets are manufactured with TSMC’s 3nm process and stacked over two 6nm I/O chiplets.

Both MI355X and MI350X offer 288GB of high-speed HBM3E memory, surpassing Nvidia’s B200 by 60% in memory capacity.

AMD claims its chips outperform Blackwell in specific tasks—processing 8-bit floating-point numbers 10% faster and 4-bit FP numbers more than twice as fast.

Each MI350 chip can support models with up to 520 billion parameters.

AMD also introduced new 8-chip AI servers, delivering up to 160 petaflops of FP4 performance.

Looking ahead, the company will launch Helios rack systems, combining next-gen MI400 chips with CPUs and Pensando DPUs.

The updated ROCm 7.0 software platform promises over 3.5x inference and 3x training speedups, emphasizing efficiency and distributed processing.

Earlier this month, AMD acquired stealth-mode AI startup Brium in a move aimed at bolstering its position in the competitive AI hardware market dominated by Nvidia. Announced Wednesday, the deal’s terms were not disclosed.

Brium specialises in optimising machine learning applications for inference—the process of running trained AI models—across diverse hardware platforms.

This flexibility allows AI workloads traditionally tuned for Nvidia chips to be adapted for AMD’s hardware, such as its Instinct GPUs.