- MI355x leads the new AMD MI350 series with 288 GB memory and complete liquid refrigerated performance
- AMD DROPS APTRATION, focusing on the flexibility of GPU at rack scale
- FP6 and FP4 data types high
AMD has presented its new MI350X and MI355X GPU for AI workloads in its 2025 advanced AI event, offering two options built in its last Anc 4 architecture.
While both share a common platform, the MI355X is distinguished as the variant of higher performance and liquid refrigerated designed for large -scale demanding implementations.
The MI355X admits up to 128 GPU per Rack and offers high performance for training and inference work loads. It has 288 GB of HBM3E memory and 8TB/s memory bandwidth.
GPU only design
AMD states that the MI355X offers up to 4 times the AI compute and 35 times the inference performance of its previous generation, thanks to architectural improvements and a movement to the N3P process of TSMC.
Inside, the chip includes eight calculation ratings with 256 active computing units and a total of 185 billion transistors, marking an increase of 21% over the previous model. Each dice is connected through the redesigned I/O mosaics, reduced from four to two, to the double internal bandwidth by reducing energy consumption.
The MI355X is a only design design, which eliminates the APU CPU-GPU approach used in the MI300A. AMD says this decision better supports modular implementation and rack scale flexibility.
It connects to the host through a PCIE 5.0 x16 interface and communicates with pairs GPU using seven infinity fabric links, reaching more than 1TB/s in the GPU GPU bandwidth to GPU.
Each HBM battery is combined with 32 MB of infinity cache, and architecture admits newer formats of lower precision such as FP4 and FP6.
The MI355X executes the FP6 operations to FP4 rates, an AMD feature stands out as beneficial for inference workloads. It also offers 1.6 times the NVIDIA GB200 and B200 HBM3E memory capacity, although memory bandwidth remains similar. AMD claims an inference performance of 1.2xa 1.3x on the main Nvidia products.
The GPU attracts 1,400 W in its liquid refrigerated form, which offers a higher performance density for shelf. AMD says that this improves TCO by allowing users to calculate without expanding the physical footprint.
The chip adjusts to the standard OAM modules and is compatible with the UBB platform servers, which accelerates the implementation.
“The world of AI is not slowing down, and we either,” said Vamsi Boppana, SVP, Ai Group. “In AMD, we are not only keeping the rhythm, we are establishing the bar. Our clients are demanding real and drop -down solutions that are climbing, and that is exactly what we are offering with the AMD Instinct Mi350 series. With the performance of the cutting tip, the massive bandwidth and the open infrastructure and the open and flexible infrastructure, we are promoting the innovators through the industries to the industries to the industries to the industries to the industries. Fusion of scale and the construction of what is below. ” “
AMD plans to launch its Instinct Mi400 series in 2026.
