AMD is preparing to release the AI accelerator Instinct MI355X with a power consumption of up to 1.5 kW and 288 GB of memory

AMD will soon introduce a new graphics accelerator for artificial intelligence tasks — Instinct MI355X, built on the CDNA 4 architecture. The announcement of the device is expected during the day, but now a number of key characteristics have been revealed, indicating the transition to an extreme level of performance and power consumption in data centers.

According to preliminary information, the MI355X will develop up to 5.03 petaflops (PFLOPS) in FP16 format, while the maximum power consumption will be up to 1400 watts, which makes this accelerator one of the most energy-intensive solutions in the history of the GPU.

One of the main features of the device will be the memory configuration: 288 GB HBM3E with a bandwidth of 8 TB/s. This volume and speed of data exchange are aimed at servicing heavy AI models, including generative transformers, machine learning systems, and highly loaded neural network frameworks.

The Instinct MI355X will be offered both as a single GPU module and as part of a high-density rack that includes eight accelerators simultaneously, which will allow customers to create configurations for next-generation supercomputers and AI clusters.

The parallel release of the Instinct MI350X model is also expected, which will presumably be based on the same chip, but with reduced clock speeds, which will limit its peak power consumption to 1000 watts. This can provide more flexible configuration choices for data centers with power and cooling constraints.

The new Instinct accelerator line will be part of AMD's strategy to strengthen its position in the AI hardware segment, which is dominated by Nvidia solutions. Taking into account the amount of memory and processing power, the MI355X is aimed at direct competition with the flagship H100 and B100 from Nvidia, as well as promising solutions from Intel Gaudi and new developers in the field of specialized AI chips.