Prev | CPSC 418 Home Page | Next |
Benchmarking Programs for question #2.
SPEC website for question #3.
------------------------------------------------------------------------------ HOMEWORK ASSIGNMENT #4 Out: Tues Mar 18/97 Due: Fri Mar 28/97 CS 418 Spring Session 1997 Marks: 100 ------------------------------------------------------------------------------ Question 1: The price half of price - performance (30 marks, 12-6-12) ---------------------------------------- a) die cost (this question is adapted from HP 2.13 and 2.14 on p.85): Let us consider the effects of defect densities on different sized wafers. You may assume the models I presented in class (which are also available in HP 2.3). One version of the Sun Ultrasparc processor has the rather enormous die size of 315 mm^2. At the other end of the spectrum would be the MIPS R3000, which was only 72 mm^2. In the early 1990s, semiconductor manufacturers were claiming to reach the stage where defect densities were approaching zero -- either an entire wafer failed or there were no defects. i) Evaluate the die cost (ignoring testing and packaging) for the Ultrasparc and R3000, assuming defect densities of 3, 2, 1, and 0 defects/cm^2. ii) In quantitative terms, how would essentially zero defect densities affect the design decisions of chip designers? ---------------------------------------- b) Based on the cost --> price models presented in class (and available in HP 2.3), and knowing that Intel sells the Pentium-150 for 278 USD in bulk (as of 96/10/28), give a range in which you think Intel's cost per chip lies. In practice, their cost/chip is estimated to be slightly lower. I would guess that the added difference is due to profit taking and high R&D costs (but that's just my guess :). ---------------------------------------- c) Your company, BlacknBlue Inc., is currently selling a system package consisting of a compiler and dual-pipeline processor. You estimate that this pair of components manages a CPI of 0.7 -- 60% of instructions manage to dual issue. Your company has two options for its next R&D effort: a more advanced processor or a more advanced compiler. The advanced compiler, running on the current processor, will allow 90% of instructions to dual issue. It will take 6 months to produce the production version. The advanced processor will feature a quad-issue pipeline. Using the current compiler, 60% of instructions will quad-issue, 20% will dual-issue, and 20% will single-issue. It will take 15 months for the first customer delivery of the chip. i) Calculate the CPI for the two options. ii) The industry standard system doubles in performance every two years. If the current system is industry standard, should BlacknBlue pursue either research effort, and if so, which one? Why? Remember: performance change is (old perf - new perf) / (new perf) ------------------------------------------------------------------------------ Question 2: Benchmarking (45 marks, 20-25) Ziff-Davis Benchmark Operation (a division of the Ziff-Davis Publishing Company, publisher of PC Magazine) has created a series of PC based benchmarks for comparing machines -- available to the public and used for machine reviews in their magazine. One of the benchmarks, PCBench 9.0, is available at the ZDBOp web page and is small enough to download in a reasonable time. This question is not meant to be an endorsement of either the company or the product -- I simply wanted a commercial level benchmark that was freely available. I have temporarily made the self-extracting PCBench 9.0 zip files available on the course web page. Take these files home (they are over 2 MB, so you will probably need to download them). Following the instructions in the Word document, extract and execute the "DOSMark suite" to benchmark your home machine. Running PCBench requires you to agree to a licensing agreement. If you do not wish to agree to this license, contact me and I will arrange a time for you to run the benchmark copy registered to me on a department machine. Answering these questions does not require you to read the entire PCBench 9.0 manual, although a quick skim of the first parts may help. If you are running Windows95 at home, reboot your machine into DOS before running the benchmark. I haven't had a chance to test PCBench's behaviour under WindowsNT extensively, so if you do have problems, come talk to me. If you do not have a home computer capable of running DOS, contact me and I will arrange a time for you to run the benchmark on a department machine. a) Fill out the disclosure form supplied with PCBench 9.0 after completing the DOSMark suite. Export the results (in the file menu) and submit a printed version. You may clip everything on the results page between "ADDITIONAL INFORMATION" and "WEIGHTED SUITE". I will post an example of the disclosure parts that I want to the class newsgroup. b) Does this benchmark appropriately determine the performance of your home machine with respect to your computing needs? In a short page, describe what the normal use of your home computer is, and the specific reasons why your feel this benchmark is appropriate or inappropriate. Don't forget to consider the benchmarking issues described in class. --------------------- Note: Some more recent machines appear to have problems displaying certain text fields in PCBench. The contents of large text fields (such as the license, system description and list of comparison machines) come out garbled. ZDBOp has provided one TSR they claim will fix the problem for Cirrus Logic GD543X-based graphics card (it didn't work for my Pentium, so I guess I don't have that type of graphics card). If you have this problem, there are work-arounds for the two important text fields: 1) Read the license agreement in text form in the file README.TXT in the same directory as PCBENCH.EXE. I'm SERIOUS; READ the license agreement before running the software. If you do not wish to agree to the license, we can arrange for you to run the pcbench on a department machine. 2) Fill out the "disclosure questionnaire". Even if you cannot read the results in the "disclosure form" menu option, it will still export properly. ------------------------------------------------------------------------------ Question 3: Choosing an appropriate machine (25 marks) I have a new verification problem we need to solve in the lab: Binary Decision Diagrams (BDDs) are VERY large trees used to study boolean functions. In memory, BDDs are represented by data structures that consist of a large number of objects (nodes), each with pointers to several other objects. A huge table of these objects is kept as a large array, but processing involves searching for and modifying individual objects in the BDD database in arbitrary order. The algorithm for this problem is inherently sequential -- porting to a parallel architecture would not be worth the bother. You may assume that these BDD jobs are run in batch mode, that all other users will be kicked off the machine in question, and that we will only be working on one BDD at a time. Additionally, I am not a compiler expert and I frequently make major modifications to the inner loops of the program, so I don't want to have to use complex compiler options to get my executable. Using the SPEC benchmarks available at the SPEC website, find the five machines you think would have the best performance on this problem. You must choose the 5 machines from 5 different manufacturers (otherwise all the machines would be fast Alphas :). Justify your choice of benchmark(s) and machines (for the latter, I would suggest cutting selected stats out of the text version of the SPEC report -- don't submit 5 pages of SPEC reports). Suggest what changes to the tested configurations might increase the performance of machine(s) on this problem. ------------------------------------------------------------------------------ Question 4: Evaluation For each of the questions 1-3, how long did it take to answer? Do you have any other comments about the assignment? ------------------------------------------------------------------------------
Prev | CPSC 418 Home Page | Next |