AMD expands Instinct MI400 Series with new GPU options - Engineering.com

AMD has added new details about the AMD Instinct MI400 Series GPU portfolio, including the AMD Instinct MI430X GPU for high-performance computing and AI for science, measured inference results for the AMD Instinct MI455X GPU on the AMD Helios rack-scale solution and the AMD Instinct MI450X GPU for customers that want another memory-capacity option.

Scientific simulation, frontier AI and interactive inference have different infrastructure requirements. The MI400 Series GPUs are intended to address those needs with native double-precision compute for scientific workloads, higher memory capacity for large AI models and additional configuration options.

MI430X for scientific computing

The AMD Instinct MI430X GPU is designed for workloads that combine high-precision simulation and AI. It supports FP64 compute and includes 432 GB of HBM4 memory and up to 23.3 TB/s of memory bandwidth for larger simulations and datasets.

AMD also shared research on the role of native double precision in scientific computing. The work examines 47 workload configurations across 11 applications and benchmarks, looking at where FP64 emulation is sufficient and where scientific codes require native double-precision compute.

MI455X inference results

AMD testing on the MI455X GPU in the AMD Helios system showed 21,333 tokens per second per GPU on DeepSeek-R1 using an MLPerf Inference Offline configuration. AMD reported that this result was up to 30% higher than Nvidia Vera Rubin’s latest published MLPerf offline result.

AMD Helios also includes up to 20% higher HBM bandwidth and higher scale-up bandwidth than Nvidia Vera Rubin, according to AMD. Those characteristics support the inference results reported for the system.

MI450X option

AMD has added the AMD Instinct MI450X GPU to the AMD Instinct MI400 Series portfolio to give AMD Helios customers another configuration choice for their workloads. The MI450X includes 288 GB of HBM4 memory per GPU and the same core compute, memory bandwidth and scale-up and scale-out connectivity as the MI455X.

Built on AMD’s flexible compute architecture and open foundation, the MI450X is intended to support changes in customer requirements and reduce integration work for ecosystem partners.

For workloads that fit within its memory capacity, the MI450X provides another option for throughput and response-time targets while helping manage operating expense. In AMD Helios deployments, customers can choose either the MI450X or the MI455X GPU, with both supported by the same rack-scale architecture and AMD ROCm software foundation.

The MI455X provides more memory headroom for larger models, longer contexts or higher concurrency.