Architecture Comparisons Between Nvidia and ATI Gpus: Computation Parallelism and Data Communications

Architecture Comparisons Between Nvidia and ATI Gpus: Computation Parallelism and Data Communications

Architecture Comparisons between Nvidia and ATI GPUs: Computation Parallelism and Data Communications Ying Zhang1, Lu Peng1, Bin Li2 Jih-Kwon Peir3, Jianmin Chen3 1 Department of Electrical and Computer Engineering 3 Department of Computer & Information Science and Engineering 2 Department of Experimental Statistics University of Florida Louisiana State University, Baton Rouge, LA, USA Gainesville, Florida, USA {yzhan29, lpeng, bli}@lsu.edu {peir, jichen}@cise.ufl.edu Abstract — In recent years, modern graphics processing units be filled with valid instructions, the VLIW architecture will have been widely adopted in high performance computing areas outperform the traditional design. Unfortunately, this is not to solve large scale computation problems. The leading GPU the case in practice because the compiler may fail to find suf- manufacturers Nvidia and ATI have introduced series of prod- ficient independent instructions to generate compact VLIW ucts to the market. While sharing many similar design concepts, instructions. On average, if m out of n slots are filled during GPUs from these two manufacturers differ in several aspects on an execution, we say the achieved packing ratio is m/n. The processor cores and the memory subsystem. In this paper, we actual performance of a program running on a VLIW proces- conduct a comprehensive study to characterize the architectural sor largely depends on the packing ratio. On the other hand, differences between Nvidia’s Fermi and ATI’s Cypress and the Nvidia GPU uses multi-threading execution to execute demonstrate their impact on performance. Our results indicate code in a Single-Instruction-Multiple-Thread (SIMT) fashion that these two products have diverse advantages that are re- flected in their performance for different sets of applications. In and explores thread-level parallelism to achieve high perfor- addition, we also compare the energy efficiencies of these two mance. platforms since power/energy consumption is a major concern The second difference between two GPUs exists in the in the high performance computing. memory subsystem. Both GPUs involve a hierarchical organ- ization consisting of the L1 cache, L2 cache, and the global Keywords - GPU; clustering; performance; energy efficiency memory. On the GTX 580 GPU, the L1 cache is configurable to different sizes and can be disabled by setting a compiler I. INTRODUCTION flag. The L1 cache on the HD 5870 is less flexible and can With the emergence of extreme scale computing, modern only be used to cache image objects and constants. The L2 graphics processing units (GPUs) have been widely used to caches on both GPUs are shared among all hardware multi- build powerful supercomputers and data centers. With large processor units. All global memory accesses go through the number of processing cores and high-performance memory L2 in GTX 580, while only image objects and constants use subsystem, modern GPU is a perfect candidate to facilitate the L2 in HD 5870. Given these differences, we will investi- high performance computing (HPC). As the leading manufac- gate and compare the memory system of the target GPUs. turers in the GPU industry, Nvidia and ATI have introduced Thirdly, power consumption and energy efficiency is be- series of products that are currently used in several preemi- coming a major concern in high performance computing nent supercomputers. For example, in the Top500 list released areas. Due to the large amount of transistors integrated on in Jun. 2011, the world’s second fastest supercomputer chip, a modern GPU is likely to consume more power than a Tianhe-1A installed in China employs 7168 Nvidia Tesla typical CPU. The resultant high power consumption tends to M2050 general purpose GPUs [11]. LOEWE-CSC, which is generate substantial heat and increase the cost on the system located in Germany and ranked at 22nd in the Top500 list [11], cooling, thus mitigating the benefits gained from the perfor- includes 768 ATI Radeon HD 5870 GPUs for parallel compu- mance boost. Both Nvidia and ATI are well aware of this tations. issue and have introduced effective techniques to trim the Although typical Nvidia and ATI GPUs are close to each power budget of their products. For instance, the PowerPlay other on several design specifications; they deviate in many technology [1] is implemented on ATI Radeon graphics architecture aspects from processor cores to the memory hie- cards, which significantly drops the GPU idle power. Similar- rarchy. In this paper, we measure and compare the perfor- ly, Nvidia use the PowerMizer technique [8] to reduce the mance and power consumption of two recently released power consumption of its mobile GPUs. In this paper, we GPUs: Nvidia GeForce GTX 580 (Fermi) [7] and ATI Rade- measure and compare energy efficiencies of these two GPUs on HD 5870 (Cypress) [4]. By running a set of representative for further assessment. general-purpose GPU (GPGPU) programs, we demonstrate One critical task in comparing diverse architectures is to the key design difference between the two platforms and illu- select a common set of benchmarks. Currently, the program- strate their impact on the performance. ming languages used to develop applications for Nvidia and The first architectural deviation between the target GPUs ATI GPUs are different. The Compute Unified Device Archi- is that the ATI product adopts very long instruction word tecture (CUDA) language is majorly used by Nvidia GPU (VLIW) processors to carry out computations in a vector-like developers, whereas the ATI community has introduced the fashion. Typically, in an n-way VLIW processor, up to n data- Accelerated Parallel Processing technology to encourage en- independent instructions can be assigned to the slots and be gineers to focus on the OpenCL standard. To select a com- executed simultaneously. Obviously, if the n slots can always mon set of workloads, we employ a statistical clustering tech- (a) GTX 580 (b) Radeon HD 5870 (c) VLIW processor in HD 5870 Figure 1. Architecture of target GPUs nique to select a group of representative GPU applications TABLE I. SYSTEM INFORMATION from the Nvidia and ATI developer’s SDK, and then conduct GPU information our studies with the chosen programs. According to the expe- Parameter GTX 580 Radeon HD 5870 riment results, we can make several interesting observations: Technology 40nm 40nm #Transistors 3.0 billion 2.15 billion For programs that involve significant data dependen- Processor clock 1544 MHz 850 MHz cy and are difficult to generate compact VLIW bun- #Execution units 512 1600 dles, the GTX 580 (Fermi) is more preferable from GDDR5 clock rate 2004 MHZ 1200 MHz the standpoint of high performance. In contrast, the GDDR5 bandwidth 192.4 GB/s 153.6 GB/s ATI Radeon HD 5870 (Cypress) is a better option to Host system information CPU Intel Xeon E5530 AMD Opteron 6172 run programs where sufficient instructions can be Main memory type PC3-8500 PC3-8500 found to compact the VLIW slots. Memory size 6GB 6GB The GTX 580 GPU outperforms its competitor on double precision computations. The Fermi architec- these two GPUs along with a description of the host system is ture is delicately optimized to deliver high perfor- listed in Table I [4][7]. mance in double precision, making it more suitable in A. Fermi Architecture solving problems with high precision requirement. Fermi is the latest generation of CUDA-capable GPU ar- Memory transfer speed between the CPU and GPU is chitecture introduced by Nvidia [13]. Derived from prior fam- another important performance metrics which impact ilies such as G80 and GT200, the Fermi architecture has been the kernel initiation and completion. Our results show improved to satisfy the requirements of large scale computing that Nvidia generally has higher transfer speed. Be- problems. The GeForce GTX 580 used in this study is a Fer- sides the lower frequency of the device memory on mi-generation GPU [7]. Figure 1(a) illustrates its architectural the ATI HD 5870 GPU [4][7], another reason is that organization [13]. As can be seen, the major component of the memory copy in CUDA has smaller launch over- this device is an array of streaming multiprocessors (SMs), head compared to the ATI OpenCL counterpart. each of which contains 32 CUDA cores. There are 16 SMs on According to our experiments, the ATI Radeon HD the chip with a total of 512 cores integrated in the GPU. With- 5870 consumes less power in comparison with the in a CUDA core, there exist a fully pipelined integer ALU GTX 580. If a problem can be solved on these two and a floating point unit (FPU). In addition to these regular GPUs in similar time, the ATI GPU will be more processor cores, each SM is also equipped with four special energy efficient. function units (SFU) which are capable of executing tran- The remainder of this paper is organized as follows. In scendental operations such as sine, cosine, and square root. section II, we introduce the architecture of these two graphics The design of the fast on-chip memory is an important processing units. We describe our experiment methodology feature on the Fermi GPU. In specific, this memory region is including the statistical clustering technique in section III. now configurable to be either 16KB/48KB L1 cache/shared After that, we analyze and compare the different characteris- memory or vice versa. Such a flexible design provides per- tics of the target GPUs and their impact on performance and formance improvement opportunities to programs with differ- energy efficiency by testing the selected benchmarks in sec- ent resource requirement. Another subtle design is that the L1 tion IV. We review the related work in section V and finally cache can be disabled by setting the corresponding compiler draw our conclusion in section VI. flag. By doing that, all global memory requests will be by- II.

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    11 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us