Kepler (microarchitecture)

Kepler (microarchitecture)
August Köhler [1] · Public domain · source
Name	Kepler
Designer	NVIDIA
Product family	GeForce, Quadro, Tesla
Process	28 nm
Introduction	2012
Succeeded by	Maxwell

Kepler (microarchitecture) is a GPU microarchitecture developed by NVIDIA Corporation and released in 2012 for use in consumer GeForce graphics cards, professional Quadro workstations, and Tesla compute accelerators. The design targeted energy efficiency and throughput improvements over the preceding Fermi architecture and was first implemented on a 28 nm process by TSMC. Kepler powered products across desktop, mobile, and server markets while influencing CUDA compute workflows and GPGPU adoption in scientific and enterprise contexts.

Overview and development

Kepler was announced by Jen-Hsun Huang and developed by engineers at NVIDIA Research with fabrication through Taiwan Semiconductor Manufacturing Company workflows. The development cycle intersected with industry trends driven by competitors such as Advanced Micro Devices and initiatives like Intel Corporation's integrated graphics roadmaps and the broader move to 28 nm lithography. Kepler's roadmap aligned with software ecosystems including Microsoft's DirectX 11 and OpenGL extensions, as well as compute APIs like CUDA and OpenCL (standard). The platform release coincided with market events such as CES product announcements and OEM partnerships with Dell, HP Inc., and Lenovo for laptop and workstation shipments.

Architecture and design

Kepler introduced a redesigned streaming multiprocessor, the SMX, improving upon the Fermi (microarchitecture) SM. The SMX combined more CUDA cores per unit and enhanced SM scheduling to increase utilization for workloads common to 3D graphics, scientific computing, and machine learning. Kepler emphasized energy efficiency via clock-gating and a refined power-management subsystem that interfaced with ACPI and platform firmware used by Intel and AMD motherboards. The architecture implemented features relevant to Direct3D 11 feature levels and supported advanced texture compression formats used in Unity (game engine), Unreal Engine, and rendering pipelines in Autodesk applications. Kepler also integrated improvements in rasterization, tessellation performance compliant with Tessellation (graphics) specifications, and memory subsystem optimizations tied to GDDR5 interfaces standardized by memory vendors such as Micron Technology and Samsung Electronics.

Product lineup and variants

Kepler cores were deployed across NVIDIA product lines: the consumer-oriented GeForce GTX 600 series and GeForce GTX 700 series desktop cards, mobile GeForce GTX 600M series for laptop OEMs like ASUS and Acer, professional Quadro K-series models for content-creation vendors including Adobe Systems and Autodesk, and compute-oriented Tesla K20 and Tesla K40 accelerators for datacenter customers such as NASA research centers and Lawrence Livermore National Laboratory. Die variants included GK104, GK106, GK107, GK110, and GK208, each targeting segments from mainstream to high-performance computing customers served by distributors like Ingram Micro and channel partners such as Arrow Electronics. OEMs and system integrators including Supermicro and Hewlett Packard Enterprise integrated Kepler-based cards into servers and workstations certified for enterprise software from Siemens and ANSYS.

Performance and features

Kepler delivered improved single-precision throughput and enhanced double-precision options on specific SKUs like the GK110-powered Tesla K20X aimed at scientific institutions including CERN and Los Alamos National Laboratory. Architectural changes benefited real-time graphics in games published by Electronic Arts, Activision Blizzard, and Ubisoft, and accelerated compute tasks in frameworks from MathWorks and NVIDIA CUDA Toolkit. Kepler supported features such as GPU Boost for dynamic frequency scaling, hardware video encode/decode improvements used by Adobe Premiere Pro and FFmpeg integrations, and energy-aware scheduling in enterprise grids managed by software from Red Hat. Benchmarks from reviewers at Tom's Hardware, AnandTech, and PC Gamer highlighted trade-offs between throughput, power consumption, and price relative to competing Radeon GPUs from Advanced Micro Devices.

Software and driver support

Kepler was supported by NVIDIA's driver releases across Windows versions including Windows 7 and Windows 8, and by Linux distributions such as Ubuntu (operating system) and Red Hat Enterprise Linux. The CUDA ecosystem provided developer tools including cuDNN and cuBLAS optimized for Kepler cores, and middleware from companies like Palantir Technologies and Bloomberg L.P. leveraged Kepler acceleration. Game middleware companies such as Epic Games and Unity Technologies updated engines to exploit Kepler features, while professional ISVs like Autodesk and Dassault Systèmes validated drivers for certification programs. Long-term support and legacy driver channels were managed through partnerships with Microsoft's WHQL program and enterprise support agreements with IBM systems integrators.

Reception and legacy

Industry reception acknowledged Kepler for its energy efficiency and compute advances, influencing NVIDIA's subsequent Maxwell and Pascal generations and shaping strategies at competitors including AMD's Graphics Core Next roadmap. Kepler-equipped systems contributed to breakthroughs in fields represented by institutions such as Stanford University and MIT where GPUs enabled research in deep learning and accelerated simulation workflows. The architecture's balance of graphics and compute features helped expand GPU computing adoption across cloud providers like Amazon Web Services, Google Cloud Platform, and Microsoft Azure where Kepler-based instances were offered. Kepler's legacy persists in software compatibility layers and in preserved deployments within legacy HPC clusters maintained by organizations including Oak Ridge National Laboratory and academic consortia like XSEDE.

Category:Graphics processing units