
AMD Says MI455X Delivers 30% Higher Token Throughput Than Nvidia - www.thelec.net
AMD has unveiled its latest Instinct GPU lineup, with the MI455X delivering 30% higher token throughput than Nvidia's Vera Rubin NVL72 server rack in MLPerf's offline AI inference benchmark.
AMD has disclosed the artificial intelligence (AI) computing performance of its latest Instinct server graphics processing units (GPUs) and unveiled additional products in the lineup. On Nov. 12 local time, the company announced that the MI455X, used in its Helios system, achieved throughput of 21,333 tokens per second per GPU when running the DeepSeek-R1 model in MLPerf's offline AI inference benchmark.
This result was 30% higher than Nvidia's Vera Rubin NVL72 server rack in the same test, whose results were released on Nov. 11. The inference performance advantage stems from differences in system data bandwidth, with the MI455X offering up to 20% higher high-bandwidth memory (HBM) bandwidth than Nvidia's Rubin GPU.
Superior Architecture
The MI455X's GPU-to-GPU scale-up bandwidth is also higher than that of Nvidia Vera Rubin. Scale-up refers to an architecture that connects multiple GPUs through ultra-high-speed communication within a rack. AMD Helios houses 72 Instinct MI455X GPUs, with a combined 31 terabytes (TB) of HBM4 capacity and memory bandwidth of up to 1.7 petabytes per second (PB/s). Its scale-up bandwidth reaches 260 terabytes per second (TB/s).
Scientific Computing Option
AMD also shared detailed specifications for the MI430X, designed for high-performance computing (HPC) in scientific research. The GPU features 432 gigabytes (GB) of HBM4 memory and maximum memory bandwidth of 23.3 TB/s. The MI430X natively supports highly precise 64-bit double-precision floating-point (FP64) calculations.
AMD analyzed 47 workloads across 11 scientific computing applications and benchmarks and found that the accuracy of many precision simulations and scientific calculations depended on native hardware FP64 computing performance. The company also introduced the Instinct MI450X, a cost-optimized GPU equipped with 288GB of HBM4 memory per GPU.