CFM has updated the domain of its english website from en.chinaflashmarket.com to www.memorymarket.com, Please be informed.

AMD MI400 GPU Product Matrix Expanded: MI455X Targets Inference, MI450X Lowers Barrier, MI430X Preserves Precision

By: M 9 hours ago

AMD has released detailed information on the cost-effective MI450X and the MI430X dedicated to scientific computing, as well as inference benchmark results for the MI455X based on its rack-scale solution "Helios."

According to AMD's testing, the MI455X GPU integrated into the Helios solution achieved a single-GPU processing speed of 21,333 tokens per second in an offline MLPerf inference environment based on the "DeepSeek-R1" model. This is 30% higher than the offline MLPerf test results of NVIDIA's recently released next-generation platform Vera Rubin. AMD stated that compared with Vera Rubin, its Helios architecture demonstrates superior performance, with 20% higher HBM bandwidth and superior scale-out bandwidth.

In addition to the MI455X, which focuses on AI inference, AMD has refined its Instinct MI400 GPU product matrix for different scenarios.

According to the company, the newly added MI450X GPU is equipped with 288GB of next-generation HBM4 memory and maintains the same core computing performance, memory bandwidth, and scalability (scale-up and scale-out) as the high-end MI455X model. It is designed to provide customers with a more cost-effective option for large-scale AI deployment and help Helios customers optimize their infrastructure based on workload characteristics and operating costs. The MI455X is based on the same open rack-scale architecture and "ROCm" software ecosystem and is suitable for ultra-large models or applications requiring long-context processing, while the MI450X is suitable for efficient operation within memory capacity limits.

The MI430X GPU is equipped with 432GB of HBM4 memory, with memory bandwidth of up to 23.3TB/s, and is optimized for processing large datasets. Its core advantage is native support for FP64 double-precision computing; in precision simulation and scientific computing scenarios, hardware-level FP64 performance is decisive.