AMD Says MI455X Delivers 30% Higher Token Throughput Than Nvidia
What happened
On Nov. 12 local time, AMD disclosed the AI computing performance of its latest Instinct server GPUs and unveiled additional products in the lineup. 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, 30% higher than Nvidia's Vera Rubin NVL72 server rack in the same test, whose results were released on Nov. 11. According to the source, the inference performance advantage stems from differences in system data bandwidth: the MI455X offers up to 20% higher high-bandwidth memory (HBM) bandwidth than Nvidia's Rubin GPU, and its GPU-to-GPU scale-up bandwidth is also higher than that of Nvidia Vera Rubin. AMD Helios houses 72 Instinct MI455X GPUs, with a combined 31 terabytes (TB) of HBM4 capacity, memory bandwidth of up to 1.7 petabytes per second (PB/s), and scale-up bandwidth reaching 260 terabytes per second (TB/s). AMD also shared detailed specifications for the MI430X, designed for high-performance computing (HPC) in scientific research: 432GB of HBM4 memory, maximum memory bandwidth of 23.3 TB/s, and native support for 64-bit double-precision floating-point (FP64) calculations; AMD analyzed 47 workloads across 11 scientific computing applications and benchmarks. AMD additionally introduced the Instinct MI450X, a cost-optimized GPU equipped with 288GB of HBM4 memory per GPU, whose core computing performance, memory bandwidth, scale-up capabilities and inter-server scale-out connectivity are identical to the MI455X, available as a GPU option for AMD Helios.
Impact and signal
Affected components and suppliers
Source and time
Source: The Elec · 11/17/2026 · View original source
Real-time pricing, supply signals and AI analysis for semiconductor traders