AMD

Advanced Micro Devices, Inc. (AMD) is an American multinational semiconductor company headquartered in Santa Clara, California, and the leading challenger to NVIDIA in the market for artificial intelligence accelerators. Founded on May 1, 1969, AMD designs central processing units (CPUs), graphics processing units (GPUs), field-programmable gate arrays (FPGAs), and adaptive System-on-Chip products, competing across data centre, client, gaming, and embedded computing markets.^[1]

History

AMD was founded in 1969 by a group of former Fairchild Semiconductor executives led by Jerry Sanders, initially producing logic chips and memory products before entering the microprocessor market in the 1970s. The company spent decades as a second-source supplier and competitor to Intel in x86 processors, establishing itself through the K6 and Athlon lines and later the Ryzen CPU family that began in 2017. Under chief executive Lisa Su, who took the role in 2014, AMD repositioned itself as a high-performance computing company, completing the acquisition of Xilinx in 2022 to add FPGA and adaptive computing capabilities.^[1][2]

AMD's modern AI strategy is built around the Instinct line of data centre accelerators. The Instinct MI300X, launched in December 2023, was the company's first GPU designed explicitly for large language model training and inference, featuring 192 GB of HBM3 memory and an architecture that aimed to compete directly with NVIDIA's H100 and subsequent products. The MI300 series was followed by the MI325X and the MI350 series, with AMD publicly committing to an annual cadence of accelerator releases.^[2][3]

Key Technology

AMD's AI accelerators are based on the CDNA architecture, a GPU design lineage specialised for compute workloads rather than graphics rendering. The MI300X combines eight compute dies with eight input/output dies on a single package using advanced chiplet packaging, an approach that allows AMD to scale performance while managing manufacturing yields. The accelerator's 192 GB of high-bandwidth memory was a deliberate differentiator, allowing larger models to be served on a single GPU without memory pooling.^[3]

Complementing the hardware is ROCm (Radeon Open Compute), AMD's open-source software stack for GPU computing. ROCm provides a programming model, compiler toolchain, and libraries comparable in purpose to NVIDIA's CUDA platform, supporting popular machine learning frameworks such as PyTorch and TensorFlow. The openness of ROCm, and AMD's willingness to publish benchmark results and partner with software vendors, has been central to its efforts to erode CUDA's dominance in AI development.^[4]

AMD also produces the EPYC server CPU line, whose high core counts are widely used in AI inference servers and as host processors in GPU clusters, and the Ryzen AI line of client processors that integrate neural processing units for on-device AI workloads.^[2]

Applications and Impact

AMD Instinct accelerators have been adopted by major cloud providers and AI companies seeking an alternative to NVIDIA. Microsoft Azure, Meta, and Oracle Cloud Infrastructure have all offered Instinct-based instances, and the open-source AI community has increasingly targeted ROCm as a viable deployment path for models such as Llama. In the high-performance computing sector, AMD accelerators power some of the world's largest supercomputers, including Frontier at Oak Ridge National Laboratory, the first exascale system in the United States.^[5]

The competitive pressure from AMD has been widely credited with accelerating NVIDIA's product roadmap and influencing pricing in the AI accelerator market. AMD's stated strategy is to win share by combining aggressive memory capacity, open software, and a total cost of ownership argument against NVIDIA's established ecosystem.^[4]

>See Also

• NVIDIA
• Tensor Processing Unit
• GPU Cluster
• CUDA
• National AI Office Malaysia

References