AMD
Last reviewed
Jun 1, 2026
Sources
36 citations
Review status
Source-backed
Revision
v1 · 2,959 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 1, 2026
Sources
36 citations
Review status
Source-backed
Revision
v1 · 2,959 words
Add missing citations, update stale details, or suggest a clearer explanation.
Advanced Micro Devices, Inc. (AMD) is an American semiconductor company headquartered in Santa Clara, California. Founded in 1969, it designs central processing units (CPUs), graphics processing units (GPUs), data-center accelerators, and adaptive computing hardware such as field-programmable gate arrays (FPGAs). For most of its history AMD was known as the main competitor to Intel in the x86 processor market and to NVIDIA in consumer graphics. Since the mid-2010s it has grown into one of the largest suppliers of high-performance computing silicon, and since the early 2020s it has become a significant vendor of data center hardware for artificial intelligence.[1][2]
AMD has been led since 2014 by Lisa Su, who became chair of the board in 2022. Under her tenure the company introduced the Zen CPU architecture, regained server and desktop market share from Intel, and surpassed Intel in market capitalization for the first time in 2022.[3][4] AMD does not manufacture its own chips; like most modern chip designers it is a fabless company and outsources production to foundries, principally Taiwan Semiconductor Manufacturing Company (TSMC).[2]
In 2025 AMD reported record annual revenue of 34.6 billion dollars, up 34 percent year over year, with its Data Center segment generating 16.6 billion dollars.[5] Much of the company's recent growth comes from its EPYC server processors and its Instinct line of data-center GPUs, which compete with NVIDIA's accelerators for training and running large AI models. In October 2025 AMD and OpenAI announced a multi-year agreement to deploy 6 gigawatts of AMD GPUs, one of the largest computing commitments in the industry to that point.[6][7]
| Field | Detail |
|---|---|
| Company | Advanced Micro Devices, Inc. |
| Founded | May 1, 1969 |
| Founders | Jerry Sanders and seven colleagues from Fairchild Semiconductor |
| Headquarters | Santa Clara, California, United States |
| CEO and Chair | Lisa Su (CEO since 2014, chair since 2022) |
| Industry | Semiconductors |
| Key products | Ryzen and EPYC CPUs, Radeon GPUs, Instinct data-center accelerators, Versal and Virtex FPGAs |
| Architectures | Zen (CPU), RDNA (consumer GPU), CDNA (data-center GPU) |
| Manufacturing | Fabless; chips fabricated by TSMC |
| 2025 revenue | 34.6 billion USD |
| Employees | More than 25,000 |
| Stock symbol | NASDAQ: AMD |
AMD was incorporated on May 1, 1969 by Jerry Sanders and seven colleagues who had left Fairchild Semiconductor.[8] The company first operated out of the home of co-founder John Carey before moving to Santa Clara and then, in September 1969, to a new headquarters at 901 Thompson Place in Sunnyvale, California, where it remained for decades.[9] AMD's earliest products were logic chips and memory, and it became a second-source manufacturer of components designed by other companies.
In 1975 AMD entered the microprocessor business, initially by producing chips compatible with Intel's designs.[1] A 1982 agreement made AMD a licensed second source for Intel's x86 processors, supplying parts for IBM personal computers. The relationship later broke down and led to years of litigation over x86 licensing rights, which AMD largely won, securing its ability to make x86-compatible chips. Through the 1990s AMD competed with Intel using the Am486 and K5 processors, then the K6.
The company's competitive position improved sharply with the K7 generation. The Athlon processor launched on June 23, 1999, and in 2000 AMD shipped the first x86 processor to reach a clock speed of 1 gigahertz, narrowly beating Intel to that mark.[10] The Athlon 64 and Opteron lines that followed introduced 64-bit extensions to the x86 instruction set, an approach Intel later adopted.
In 2006 AMD acquired the Canadian graphics company ATI Technologies in a deal valued at about 5.4 billion dollars at closing, combining roughly 4.3 billion dollars in cash with 58 million AMD shares.[11] The acquisition gave AMD the Radeon graphics business and the engineering base for its later combined CPU and GPU products. Integrating ATI proved expensive, and AMD took large write-downs on the deal in subsequent years. In 2009 AMD spun off its manufacturing operations into a separate foundry, GlobalFoundries, completing its transition to a fabless model.
AMD's modern resurgence began with the Zen microarchitecture, first shipped in 2017. Zen was sold as Ryzen for desktops and laptops and as EPYC for servers and workstations.[1][12] Successive generations (Zen 2, Zen 3, Zen 4, and Zen 5) closed and in many benchmarks reversed AMD's performance gap with Intel, and aggressive pricing helped the company take market share in both consumer and server segments. By 2022 AMD exceeded Intel in market value for the first time.[4]
In February 2022 AMD closed its acquisition of Xilinx, a maker of FPGAs and adaptive computing chips, in an all-stock transaction with a total purchase consideration of about 48.8 billion dollars, widely described as the largest semiconductor deal to that date.[13] The deal added Xilinx's Versal and Virtex product lines and formed the basis of AMD's Adaptive and Embedded Computing group. AMD also acquired data-center systems company Pensando in 2022 and, in 2024, the server-maker ZT Systems to strengthen its rack-scale AI systems capability.
AMD's processor business spans consumer, commercial, and server markets. The Ryzen brand covers desktop and mobile CPUs built on the Zen architecture, and competes with Intel's Core line. Ryzen processors are widely used in gaming PCs and workstations, and AMD's 3D V-Cache technology, which stacks additional cache memory on top of the processor die, has been particularly popular for gaming.[12]
In servers, the EPYC line is the foundation of AMD's data-center growth. EPYC processors offer high core counts and memory bandwidth, and AMD has used them to take a substantial share of the server CPU market from Intel's Xeon. The fifth-generation EPYC processors, code-named Turin and based on Zen 5, launched in October 2024 with configurations up to 192 cores.[14] By the fourth quarter of 2025, industry trackers placed AMD's server CPU revenue share above 40 percent.[15]
EPYC processors also anchor several of the world's fastest supercomputers. The El Capitan system at Lawrence Livermore National Laboratory, which became the top-ranked machine on the TOP500 list in November 2024 with a measured performance of about 1.74 exaflops, is built on AMD's MI300A APU, which combines Zen 4 CPU cores and CDNA 3 GPU compute on a single package.[16] The earlier Frontier supercomputer at Oak Ridge National Laboratory, the first to exceed one exaflop, also uses AMD EPYC CPUs paired with Instinct MI250X accelerators.
AMD's next server CPU generation, code-named Venice and based on Zen 6, is built on a 2-nanometer-class TSMC process and is planned to scale to 256 cores. Venice is intended to pair with the company's MI400 GPUs in its Helios rack systems.[17]
AMD's graphics products fall into two families. The Radeon brand and the RDNA architecture serve gaming and consumer markets, while the Instinct brand and the CDNA architecture target data centers and high-performance computing. The split lets AMD optimize each line for its workload: RDNA emphasizes graphics rendering, while CDNA removes fixed-function graphics hardware and concentrates on the matrix and vector math used in scientific computing and AI.[18]
The Instinct line is central to AMD's AI strategy. The accelerators use a multi-chiplet design and large amounts of high-bandwidth memory (HBM), which is important for AI inference because it lets a single GPU hold larger models and process longer context windows without splitting work across many devices.
The Instinct MI300X, launched in December 2023, is built on the CDNA 3 architecture using a 5-nanometer process. It carries 192 gigabytes of HBM3 memory with about 5.3 terabytes per second of bandwidth, and uses 304 compute units across eight accelerator chiplets.[19] At launch its memory capacity was larger than that of NVIDIA's then-current H100, a point AMD emphasized for memory-bound inference. The companion MI300A is an APU variant that shares a unified memory pool between CPU and GPU and is used in the El Capitan supercomputer.[16]
The Instinct MI325X, announced in October 2024 with broad system availability from early 2025, kept the CDNA 3 architecture but increased memory to 256 gigabytes of HBM3E at about 6 terabytes per second.[20] AMD positioned it against NVIDIA's H200, citing larger memory capacity and bandwidth.
The MI350 series, comprising the air-cooled MI350X and the liquid-cooled MI355X, launched in mid-2025 and moved to the CDNA 4 architecture on a TSMC N3-class 3-nanometer process. Each GPU contains about 185 billion transistors and 288 gigabytes of HBM3E memory with up to 8 terabytes per second of bandwidth, and adds hardware support for the low-precision FP4 and FP6 data formats used to speed up AI inference.[21][22] A single MI355X delivers roughly 10.1 petaflops of FP8 and 20.1 petaflops of FP6 or FP4 compute, with a total board power of up to 1,400 watts in the liquid-cooled configuration.[23] AMD reported large generational gains over the MI300 generation and presented the MI350 series as competitive with NVIDIA's Blackwell B200 on inference workloads.[21]
The following table summarizes the recent Instinct data-center GPUs.
| Accelerator | Architecture | Launch | Memory | Bandwidth | Notable compute |
|---|---|---|---|---|---|
| Instinct MI300X | CDNA 3 | Dec 2023 | 192 GB HBM3 | ~5.3 TB/s | ~1.3 PFLOPS FP16 |
| Instinct MI325X | CDNA 3 | 2024–2025 | 256 GB HBM3E | ~6 TB/s | FP8/FP16 matrix |
| Instinct MI350X / MI355X | CDNA 4 | 2025 | 288 GB HBM3E | ~8 TB/s | ~20 PFLOPS FP4/FP6 |
| Instinct MI400 series | next-gen CDNA | 2026 (planned) | 432 GB HBM4 | ~19.6 TB/s | ~40 PFLOPS FP4 |
The MI400 series is AMD's next data-center generation, planned for 2026. AMD has stated the MI400 will offer 432 gigabytes of HBM4 memory, about 19.6 terabytes per second of memory bandwidth, and roughly 40 petaflops of FP4 and 20 petaflops of FP8 compute per GPU, with 300 gigabytes per second of scale-out bandwidth.[17][24] The series is designed to be deployed at rack scale through AMD's Helios platform, a liquid-cooled system that links 72 MI400-class accelerators into a single high-bandwidth domain. AMD says a Helios rack provides 31 terabytes of HBM4 memory, 1.4 petabytes per second of aggregate bandwidth, and up to 2.9 FP4 exaflops for inference and 1.4 FP8 exaflops for training.[17][25] The Helios rack is built on the Open Compute Project's Open Rack design that Meta contributed in 2025, part of AMD's effort to standardize AI rack hardware on open specifications.[26]
AMD's GPU software stack is called ROCm (Radeon Open Compute). It is an open software platform that includes drivers, compilers, libraries, and programming tools for running general-purpose and AI workloads on AMD GPUs.[27] ROCm's core programming model is HIP, a C++ runtime and kernel language designed to be close enough to NVIDIA's CUDA that developers can port existing CUDA code with relatively modest changes.[28]
Software has historically been AMD's main disadvantage against NVIDIA in AI. CUDA has a long head start, a large developer base, and deep integration with the most widely used machine-learning frameworks, which has made it difficult for buyers to switch hardware vendors. AMD's response has been to lean on openness and on tight integration with popular open-source projects rather than to build a closed competitor.[29] ROCm 7, released in 2025, added native support for the MI350 and MI325X GPUs, expanded handling of the FP4 and FP8 inference data types, broadened support to Windows and to consumer Radeon GPUs, and provided day-one compatibility with frameworks such as PyTorch and the vLLM inference engine.[30][31] AMD has compared its approach to the Linux development model, arguing that an open ecosystem can improve faster than a single proprietary stack.[29]
AMD's AI strategy combines hardware, software, and full-system products, aimed at the market for training and running large neural networks, where NVIDIA holds a dominant position. AMD's pitch to large buyers rests on three points: more on-package memory per GPU, an open software stack, and an annual product cadence intended to match NVIDIA's release pace.[18][32]
Memory capacity is a recurring theme. Because large language models and their context windows are often limited by available GPU memory, AMD has consistently shipped Instinct accelerators with more HBM than the comparable NVIDIA part of the same period, which can let a model run on fewer GPUs for inference.[19][20] AMD has also moved from selling individual accelerators toward selling complete rack-scale systems, the same shift NVIDIA made with its NVL and rack products, so that networking, cooling, and CPUs are designed together with the GPUs. The Helios platform and the ZT Systems acquisition are part of that move.[17][32]
NVIDIA still holds the large majority of the data-center AI accelerator market, helped by the maturity of CUDA and by its NVLink interconnect and networking products. AMD competes on price, memory, and openness, and has won deployments with several major cloud and AI companies, but it has acknowledged execution risk in scaling production and software support to the level its largest customers require.[32] U.S. export controls have also affected the business: AMD recorded roughly 440 million dollars in inventory-related charges in 2025 tied to restrictions on selling its MI308 accelerators to China.[5]
AMD's data-center GPUs are deployed by several large technology companies. When the MI300 series launched in December 2023, Microsoft, Meta, and Oracle were among the named early adopters.[33] Microsoft offers MI300X accelerators through its Azure cloud under the ND MI300X v5 virtual machine series, and has run production AI workloads, including Azure OpenAI Service models, on the hardware.[34] Meta uses Instinct accelerators for its own AI services and has said it works with AMD on future roadmaps, including the MI400 series.[35] Oracle Cloud Infrastructure was among the first to adopt AMD's open rack-scale design with MI355X GPUs and has announced very large planned clusters built on AMD accelerators.[35]
The most prominent partnership is with OpenAI. On October 6, 2025, AMD and OpenAI announced a multi-year agreement under which OpenAI will deploy 6 gigawatts of AMD Instinct GPUs across multiple hardware generations, starting with a 1-gigawatt deployment of MI450 GPUs in the second half of 2026.[6][36] As part of the agreement AMD issued OpenAI a warrant to purchase up to 160 million AMD shares, vesting in tranches tied to deployment milestones and other conditions, which could give OpenAI a stake of roughly 10 percent in AMD.[7][36] AMD described the arrangement as a strategic compute partnership expected to generate tens of billions of dollars in revenue over its life.[6]