This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| CDNA (microarchitecture) | |
|---|---|
| Name | CDNA |
| Designer | AMD |
| Introduced | 2020s |
| Architecture | GPU microarchitecture |
| Predecessor | RDNA |
| Successor | CDNA 2 |
| Process | TSMC |
CDNA (microarchitecture) is a GPU microarchitecture developed by Advanced Micro Devices for high-performance compute and data center acceleration. It targets scientific computing, artificial intelligence, and high-performance computing workloads, emphasizing compute density, memory bandwidth, and energy efficiency. CDNA represents AMD's strategic separation of consumer graphics and compute-focused GPUs, aligning with cloud providers, research institutions, and supercomputing centers.
CDNA was announced by Lisa Su and Dr. Mark Papermaster as part of AMD's product roadmap alongside announcements involving Radeon Technologies Group leadership changes and partnerships with NVIDIA rivals. The architecture sits within AMD's family of silicon alongside Zen microarchitecture CPUs and follows design trends seen in Intel Xe and NVIDIA Ampere lines. CDNA's launch involved collaborations with cloud providers such as Amazon Web Services, Microsoft Azure, and Google Cloud Platform, as well as integration into supercomputers like Fugaku-adjacent projects and national lab procurements including Lawrence Livermore National Laboratory and Oak Ridge National Laboratory.
CDNA's die integrates compute units, high-bandwidth memory interfaces, and interconnect fabrics inspired by works at AMD Research and influenced by standards bodies including PCI Express Special Interest Group and Open Compute Project. Its compute pipelines are organized into matrix and scalar units resembling tensor and SIMD blocks from architectures like NVIDIA Volta and Intel Nervana, while memory subsystems draw on implementations similar to HBM2E technologies used by SK Hynix and Micron Technology. The architecture includes features for virtualization compatible with software from VMware, Red Hat, and Canonical. Interconnect and coherency are designed in the spirit of protocols like CCIX and Infinity Fabric.
CDNA emphasizes double-precision and mixed-precision compute for workloads popular in institutions such as CERN, Lawrence Berkeley National Laboratory, and MIT. It provides matrix acceleration units for tensor operations comparable to Google TPU design principles and supports precision modes reflecting needs from Los Alamos National Laboratory projects. Hardware schedulers and dispatch units borrow concepts familiar to engineers from ARM Holdings and IBM Research. Power and thermal management strategies reflect collaborations with foundries like TSMC and cooling lessons from data centers run by Facebook (Meta Platforms), Alibaba Group, and Tencent.
Software ecosystems for CDNA include support from ROCm maintainers, integrations with frameworks such as TensorFlow, PyTorch, MXNet, and HPC stacks like OpenMPI and SLURM. Compiler toolchains draw on work from LLVM and GCC, while profilers and debugging trace tools reference methodologies used in Intel VTune and NVIDIA Nsight products. Partnerships with academic institutions like Stanford University, University of California, Berkeley, and ETH Zurich contributed optimizations for libraries such as cuBLAS analogues, linear algebra packages used in projects at Los Alamos National Laboratory and Princeton University.
CDNA spawned subsequent iterations and products in AMD's roadmap alongside other lines like RDNA 2 and RDNA 3, with later generations addressing feedback from deployments by Microsoft Research and Google DeepMind. Variants were tailored for rack-scale accelerators used by Hewlett Packard Enterprise and Dell Technologies and for OEMs including Lenovo and Fujitsu. Fabric and memory options evolved similarly to choices by Oracle Corporation and enterprise buyers such as Accenture.
CDNA positioned AMD in markets dominated by players including NVIDIA, Intel Corporation, and specialized vendors like Graphcore, appealing to customers in sectors represented by National Aeronautics and Space Administration, European Organization for Nuclear Research, Boeing, and General Electric. Target applications included AI training used by teams at OpenAI and DeepMind, scientific simulations by Los Alamos National Laboratory and Sandia National Laboratories, and financial modeling practiced at firms like Goldman Sachs and JPMorgan Chase.
Early reviews compared CDNA to contemporaries such as NVIDIA Ampere and Intel Ponte Vecchio in benchmarks published by outlets like AnandTech, Tom's Hardware, The Register, and research reports from Gartner and IDC. Benchmarks from HPC centers and cloud providers reported strengths in throughput for matrix workloads and energy efficiency metrics cited by Top500 and Green500 communities. Reception among academic and enterprise reviewers mirrored historical coverage patterns seen for architectures by ARM and IBM.
Category:GPU microarchitectures