home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.sys.super
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!cs.utexas.edu!tamsun.tamu.edu!snorkelwacker.mit.edu!bloom-picayune.mit.edu!athena.mit.edu!solman
- From: solman@athena.mit.edu (Jason W Solinsky)
- Subject: Precision-performance tradeoffs
- Message-ID: <1992Nov23.002211.7809@athena.mit.edu>
- Sender: news@athena.mit.edu (News system)
- Nntp-Posting-Host: m37-318-6.mit.edu
- Organization: Massachusetts Institute of Technology
- Date: Mon, 23 Nov 1992 00:22:11 GMT
- Lines: 76
-
- My thanks to everybody who answered my last question.
-
- I have an other question reguarding how much the
- support of lower precision fp numbers is worth to
- people buying supercomputers, where a supercomputer
- is defined as any machine which attains 1 GFLOP
- sustained while running "typical" code.
-
- Suppose that somebody were designing a machine that
- only directly supported a 2^N bit FP format and
- 2^N bit integers. Suppose that the amount of I/O
- bandwidth (and blinking lights) have already been
- fixed, but the designer believes that he/she
- may be loosing a significant portion of the
- market which uses 2^(N-1) and 2^(N-2) formats.
-
- Without changing the cost at all, The designer
- realizes that he/she can create machines with
- the following five specifications.
-
- Option 1: A 2^N bit FP GFLOPS
- A 2^N bit Integer GOPS
-
- Option 2: B 2^N bit FP GFLOPS
- B 2^N bit Integer GOPS
- (2*B) 2^(N-1) bit Integer GOPS
-
- Option 3: C 2^N bit FP GFLOPS
- C 2^N bit Integer GOPS
- (2*C) 2^(N-1) bit FP GFLOPS
- (2*C) 2^(N-1) bit Integer GOPS
-
- Option 4: D 2^N bit FP GFLOPS
- D 2^N bit Integer GOPS
- (2*D) 2^(N-1) bit Integer GOPS
- (4*D) 2^(N-2) bit Integer GOPS
-
- Option 5: E 2^N bit FP GFLOPS
- E 2^N bit Integer GOPS
- (2*E) 2^(N-1) bit FP GFLOPS
- (2*E) 2^(N-1) bit Integer GOPS
- (4*E) 2^(N-2) bit FP GFLOPS
- (4*E) 2^(N-2) bit Integer GOPS
-
- Suppose that 2B>2C>A>B>C
- 2D>B>D
- 2E>C>E
- D>E
-
- How high must B be for customers to prefer it
- over A? How high must D be for customers to
- prefer it over B? ect.
-
- In other words, when it is possible to increase
- tremendously the speed of lower precision
- operations at the cost of higher precision
- operations, AT WHAT POINT DOES THE GAIN IN
- PERFORMANCE OF THE LOWER PRECISION OPERATIONS
- JUSTIFY THE DECREASE IN HIGHER PRECISION
- PERFORMANCE.
-
- A corrolary question is:
-
- If machine A gives X GFLOPS of 64 bit power,
- machine B gives 2*Y GFLOPS of 32 bit power, and
- machine C gives 4*Z GFLOPS of 16 bit power, and
-
- X = Y = Z, Which would you prefer?
-
- If X = 1.1*Y = 1.2*Z, Would that still be the case?
-
- Thanks, Jason W. Solinsky.
-
-
-
-
-