NVIDIA Tesla K20-K20X GPU Accelerators Benchmarks Application Performance Technical Brief

NVIDIA Tesla K20-K20X GPU Accelerators Benchmarks Application Performance Technical Brief

NVIDIA changed the high performance computing (HPC) landscape by introducing its Fermibased GPUs that delivered an impressive leap in performance and energy-efficiency while offering a parallel programming model, CUDA, as an extension to industry standard languages like C, C++, and Fortran. Now NVIDIA has introduced the new Tesla K20-K20X GPU Accelerators that further advance the HPC industry. Built on the revolutionary Kepler architecture, these accelerators redefine the standard for energy-efficient computing and feature innovative technologies like SMX, Hyper-Q, and Dynamic Parallelism to boost application performance by up to 10x. SMX: 3x More Performance Per Watt The new SMX (Next Generation Streaming Multiprocessor) is an architectural innovation designed from the ground-up to deliver high efficiency performance. With SMX at its core, Tesla K20/K20X accelerators deliver the industry s highest single and double precision performance- 3.95 teraflops and 1.31 teraflops respectively for Tesla K20X- at an unprecedented 93% computational efficiency. Teraflop 1.5 1.0 0.5 0.0 Double Precision Performance (DGEMM) 1.22 0.43 0.17 Xeon E5-2687WTesla M2090 (Fermi) Tesla K20X Teraflop Single Precision Performance 3.0 (SGEMM) 2.0 2.9 1.0 0.89 0.0 0.35 Xeon E5-2687W Tesla M2090 Tesla K20X (Fermi) Hyper-Q: Easy Speed-up for Legacy MPI Codes Many legacy MPI codes don t generate enough work to fully occupy the GPU because they were written for CPU cores. Rather than requiring developers to refactor their codes to put more workload per MPI process, the Hyper-Q feature reduces efforts considerably because developers can now throw up to 32 MPI processes with small and medium-sized workloads at a shared GPU. Speedup vs. Dual K20 20x 15x 10x 5x CP2K-Quantum Chemistry K20 with Hyper-Q K20 without Hyper-Q 2.5x To illustrate the power of Hyper-Q, we picked a traditionally difficult code for GPUs called CP2K, a popular MPI-based quantum chemistry code, showing more than 2x performance improvement when 16 MPI ranks are run on the shared GPU. 0x 0 5 10 15 20 Number of GPUs

Dynamic Parallelism: Simplifying Parallel Programming Dynamic Parallelism allows the GPU to operate more autonomously from the CPU by generating new work for itself at runtime, from inside a kernel. The concept is simple, but the impact is powerful: it can make programming easier, particularly for algorithms traditionally considered difficult such as divide-and-conquer problems. Relative Sorting Performance Quicksort Tesla K20X Performance Advantage over Sandy Bridge CPUs 4x 3x 2x 1x 2x Without Dynamic Parallelism With Dynamic Parallelism 0x To showcase its potential, we used 0 5 10 Dynamic Parallelism on Quicksort, a well- Problem Size (Million of Elements) known algorithm for sorting methods, to reduce the lines of CUDA code in half while improving performance by 2x. K20X Benchmarks by Application Accelerating Key Scientific Applications by up to 10x Today, hundreds of applications take advantage of GPU acceleration, spanning all scientific disciplines and engineering domains, and the number of applications continues to grow. In the past year alone, the number of CUDA-accelerated applications has grown by over 60%. When Tesla K20 GPU Accelerators are added to servers with Sandy Bridge CPUs, CUDAenabled applications are typically accelerated up to 10x. The K20X benchmark below shows single node performance of leading applications in various science domains. CPU results: Dual socket E5-2687w; GPU results: Dual socket E5-2687w + 2 Tesla K20X GPUs *MATLAB results comparing one i7-2600k CPU vs. Tesla K20 GPU

Here s a closer look at three of the applications listed above. Chroma: High Energy & Nuclear Physics Chroma is used by scientists to test alternatives to the Standard Model of physics in order to develop a deeper understanding of the fundamental properties of matter and energy. The benchmark below show significant performance boost when adding one or two Tesla K20X GPU Accelerators to dual socket CPUs. 24 3 x128 Lattice Volume, BiCSTAB solver 20x 18.0x Relat iv e to 2x CP U 16x 14.6x 12x 8x 4x 0x 1.0x 7.5x 9.4x 2xCPU 2xCPU+ 2xCPU+ 2xCPU+ 2xCPU+ 1xM2090 2xM2090 1xK20X 2xK20X CPU results: Dual socket E5-2687w; GPU results: Dual socket E5-2687w + 2 Tesla K20X GPUs, 64GB per node SPECFEM3D: Earth Sciences Researchers use SPECFEM3D for a deeper understanding of the phenomena that shape seismic activity, thereby allowing engineers to assess potential hazards and to build structural countermeasures. This code was a Gordon Bell winner in 2008 prior to the GPU work, so it was a highly tuned code even before the recent work to accelerate on GPUs.

U ive elat R 16.0x x CP12.0x 8.0x 256x128 3D Spatial Discretization 5.4x 9.3x to 28.8x 4.0x 0.0x 1.0x 12.8x 2xCPU 2xCPU+ 2xCPU+ 2xCPU+ 2xCPU+ 1xM2090 2xM2090 1xK20X 2xK20X CPU results: Dual socket E5-2687w; GPU results: Dual socket E5-2687w + 2 Tesla K20X GPUs, 64GB per node AMBER: Molecular Dynamics Molecular dynamics (MD) allows the study of biological and chemical systems at the atomistic level on timescales from femtoseconds to milliseconds. Numerous software packages exist for conducting MD simulations of which one of the most widely used is AMBER. With the Tesla K20X GPU Accelerators, acceleration by up to 80% is observed on AMBER, over compared to Tesla M2090 GPUs. 10.0x SPFP-JAC_production_NVE Relative to 2x CPU 8.0x 6.0x 4.0x 2.0x 1.0x 3.4x 4.6x 7.1x 8.2x 0.0x 2xCPU 2xCPU+ 2xCPU+ 2xCPU+ 2xCPU+ 1xM2090 2xM2090 1xK20X 2xK20X CPU results: Dual socket E5-2687w; GPU results: Dual socket E5-2687w + 2 Tesla K20X GPUs, 64GB per node

Large Cluster Scaling for GPU-Accelerated Applications Large clusters around the world are increasingly deploying GPUs to accelerate their workload. In the Top500 list, from June 2011 to June 2012, number of GPU-accelerated systems grew by over 400%. Whether an application is GPU-accelerated or not, designing code to scale well across many nodes has never been an easy task. However, many GPU-accelerated applications are scaling well as developers work on extracting more parallelism by allocating more parallel work within a node and use MPI to distribute GPU workloads. WL-LSMS: Material Science WL-LSMS simulates the magnetic behavior of materials at the atomic and nanoscale in order to design lightweight, powerful magnetic components for highly efficient electric motors, generators, and magnetic storage devices. WL-LSMS was a Gordon Bell winner in 2009, so it was already a highly-tuned CPU code prior to the recent work to accelerate on GPUs. 40x 32 Atom Emsemble of Iron (Fe) Relative to a Single CPU only Node Cray XK7 -Tesla K20X 30x Cray XK7-CPU 20x 6 10x x 0x 0 20 40 60 80 100 120 140 160 180 200 # Sockets QMCPACK: Material Science QMCPACK is used by researchers to develop new insights into condensed matter physics, materials science, and chemistry by using more physically accurate methods. For large systems on the order of 200 electrons or more, simulations that traditionally would take upwards of 12 hours or more can now run on the order of 3 hours per simulation.

Compute Efficiency 1500000 Cray XK7 Tesla K20X Cray XK7 CPU 3x3x1 Graphite 1000000 4x 500000 0 0 500 1000 1500 2000 2500 # of Compute Nodes NAMD: Molecular Dynamics NAMD (Not (just) Another Molecular Dynamics program) is a free-of-charge molecular dynamics simulation package written using the Charm++ parallel programming model, noted for its parallel efficiency and often used to simulate large systems (millions of atoms). NAMD also scales well across many nodes accelerated with GPUs. ns/day 100x STMV 2.0 1.5 Cray XK7 K20X Cray XK7 CPU 1.0 4x 0.5 0.0 128 256 512 768 # of Compute Nodes For more information on Tesla K20/K20X GPU Accelerators, visit www.nvidia.com/tesla.

Notice ALL INFORMATION PROVIDED IN THIS TECHNICAL BRIEF, INCLUDING COMMENTARY, OPINION, NIVIDA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, MATERIALS ) ARE BEING PROVIDED AS IS. NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMEN, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or system without express written approval of NVIDIA Corporation. Copyright 2012 NVIDIA Corporation. All rights reserved. NVIDIA, the NVIDIA logo, Tesla, Kepler, and CUDA are trademarks and / or registered trademarks of NVIDIA Corporation. All company and product names are trademarks or registered trademarks of the respective owners with which they are associated. Features, pricing, availability, and specifications are all subject to change without notice. NOV 2012 www.nvidia.com