AIWiki
Malaysia
Back to all articles
Companies & Toolsamdinstinctrocm

AMD

5 min readUpdated September 2026
AMD
Founded
1969
Founders
Jerry Sanders and others
Headquarters
Santa Clara, California, USA
Key AI products
Instinct MI300X, MI325X, MI350X/MI355X accelerators; ROCm software; EPYC CPUs
AI architectures
CDNA 3, CDNA 4 (MI350 series)
Related
NVIDIA, Tensor Processing Unit, GPU, Data centre

Advanced Micro Devices, Inc. (AMD) is an American semiconductor company founded in 1969 and headquartered in Santa Clara, California. Under chief executive Lisa Su, who took over in 2014, AMD rebuilt its product line around high-performance CPUs and GPUs, and by the mid-2020s its Instinct family of data-centre accelerators and its open ROCm software stack had made it the principal merchant-chip challenger to NVIDIA in the market for artificial intelligence hardware.[1]

AMD competes in AI primarily with large-memory accelerators designed to run very large language models on a single chip, alongside server CPUs and adaptive computing products. The company has historically held a small share of the AI accelerator market relative to NVIDIA, but its annual product cadence and adoption by major cloud providers have made it the most significant alternative supplier for AI compute.[2]

History

AMD was founded in 1969 as a second-source manufacturer of logic chips and grew into one of the two main x86 microprocessor vendors, competing with Intel through the 1980s and 1990s. Its acquisition of ATI Technologies in 2006 gave the company a graphics-processing division, and in 2017 the Ryzen and EPYC CPU families, built on a new "Zen" architecture, began a sustained recovery in both PCs and servers. In 2020 AMD announced the acquisition of Xilinx, the FPGA specialist, completed in 2022, which expanded its data-centre portfolio into adaptive computing.

AMD's AI push was built on the CDNA architecture, a compute-focused GPU design introduced for data-centre workloads. The Instinct MI300X, launched in December 2023, paired CDNA 3 with 192 GB of HBM3 memory and 5.3 TB/s of bandwidth, deliberately exceeding the memory capacity of NVIDIA's H100 so that large models could be served from a single accelerator. The MI300X became the fastest product ramp in AMD's history, with adoption announced by Microsoft Azure, Meta, Oracle Cloud, IBM Cloud and other hyperscale and enterprise providers.[3] The MI325X followed in late 2024 with 256 GB of HBM3E, and in June 2025 AMD launched the MI350 series — the MI350X and MI355X — built on CDNA 4 using TSMC's 3-nanometer-class process, with 288 GB of HBM3E, 8 TB/s of bandwidth and native support for FP4 and FP6 "microscaling" formats. AMD previewed the next-generation CDNA 5-based MI400 series, codenamed Helios, for the 2026-2027 period.[1]

Key Technologies

The technical core of AMD's AI offering is the Instinct accelerator family combined with the ROCm software platform, AMD's open-source counterpart to NVIDIA's proprietary CUDA ecosystem. ROCm provides the libraries, compilers and runtime that let mainstream machine-learning frameworks such as PyTorch, JAX and vLLM run on AMD hardware, and AMD has pushed ROCm releases — including ROCm 7 and the ROCm 10 generation announced in 2026 — on an annual cycle with an emphasis on open standards.[5]

AMD's design philosophy for AI differs from NVIDIA's in two notable ways: a memory-first strategy in which capacity and bandwidth are maximised so that models up to hundreds of billions of parameters fit on one accelerator or a small node, and the use of industry-standard networking such as Ethernet and the UALink/Infinity Fabric ecosystem rather than proprietary fabrics.[4] The CDNA 4 generation added native low-precision matrix engines for FP4, FP6 and FP8, which roughly quadruple peak throughput for inference workloads that tolerate low precision, and its MI355X liquid-cooled variant targets the same 8-GPU racks that compete with NVIDIA's Blackwell systems.[2]

Applications and Impact

AMD Instinct accelerators are used for large language model training and inference, AI image and video workloads, and high-performance computing. Cloud customers rent MI300X and MI350 instances from major providers, and AMD has positioned itself on price-performance: its largest memory chips let operators serve models with fewer accelerators, while ROCm's openness is attractive to organisations that want to avoid vendor lock-in. On the supercomputing side, AMD-powered systems such as Frontier and El Capitan were among the first exascale machines in the world.[1]

The company's impact on the AI industry has been competitive rather than dominant: analyst estimates throughout 2024-2026 placed AMD's share of the data-centre AI accelerator market in the single digits to low teens, but its credible roadmap has pressured NVIDIA on pricing, memory configurations and software openness, and it has become the default second source for organisations that cannot obtain or do not want NVIDIA supply.[2]

>See Also

🇲🇾Malaysian Context

AMD has operated in Malaysia for decades through engineering and services operations, and in 2025 it expanded its Penang presence with a new state-of-the-art global services facility and engineering laboratory at GBS by the Sea in Bayan Lepas. The 209,000-square-foot office is designed to host more than 1,200 employees and supports semiconductor product engineering and R&D, making Penang one of AMD's larger engineering sites in Asia.[6]

AMD's expansion aligns with Malaysia's broader position in the AI supply chain: the country hosts major semiconductor assembly, test and packaging operations, and its rapid build-out of data centres — encouraged by MDEC, the National AI Office and state investment agencies — is creating demand for AI accelerators of every brand. Malaysian cloud operators and GPU-as-a-service startups have begun offering Instinct-based capacity as a lower-cost alternative to NVIDIA systems, and local engineers trained on ROCm are increasingly in demand. For Malaysian organisations, the choice between NVIDIA and AMD ecosystems is often framed around software maturity versus cost and openness, with ROCm's open-source model appealing to the public sector's push for sovereign and interoperable AI infrastructure.[6]

References

  1. ↑AMD. AMD Instinct MI350 series and beyond — accelerating the future of AI and HPC. https://www.amd.com/en/blogs/2025/amd-instinct-mi350-series-and-beyond-accelerating-the-future-of-ai-and-hpc.html
  2. ↑AMD. AMD Instinct MI350X product page (CDNA 4 specifications). https://www.amd.com/en/products/accelerators/instinct/mi350/mi350x.html
  3. ↑TensorWave. AMD MI300X accelerator unpacked: specs, performance and more. https://tensorwave.com/blog/mi300x-2
  4. ↑AMD. CDNA architecture overview. https://www.amd.com/en/technologies/cdna.html
  5. ↑AMD Newsroom. (2026). AMD ROCm 10: bringing ROCm.AI AI-native developer experiences to AMD platforms. https://newsroom.amd.com/news/rocm-10-software-ai-native-developer-experiences/
  6. ↑AMD. (2025). AMD expands R&D footprint with state-of-the-art Malaysian facility. https://www.amd.com/en/blogs/2025/amd-expands-rd-footprint-with-state-of-the-art-facility.html
Back to all articles