The cost of IEEE 754 exceptional operands and instructions

Last update: Thu Dec 5 17:59:33 2002

Since IEEE 754 arithmetic is often implemented by a combination of hardware and software, operands that are exceptional values (subnormal, Infinity, or NaN), or that result in exceptional values, can be expensive at run time, compared to normal operands, because they must be handled in software.

To investigate this further, the benchmark program timops.c and the shell script timops.sh to run it for each supported precision, together with the associated files Makefile, ieeeftn.h, second.c, and store.c, were used to measure the performance hit from exceptional values in a wide range of architectures.

The benchmark program contains a loop whose trip count for normal operands is adjusted to be at least one second, and then the same trip count is used to run that loop again with up to six types of operands whose

1. product is normal;
2. product underflows to a subnormal;
3. product underflows to zero;
4. product overflows to Infinity;
5. operands are both Infinity;
6. operands are both (quiet) NaN.

The particular operand values depend on the floating-point precision and on a IEEE 754 floating-point system, but are otherwise independent of the host CPU architecture.

On most current RISC systems, exceptional values are handled by software, but the trap to that software is transparent to the user, apart from taking longer than a hardware implementation would require.

One notable exception to user transparency is the Compaq (formerly DEC) Alpha architecture. Its designers chose to implement a heavily pipelined CPU that (except for the most recent Alpha 21264 and 21364 CPUs) cannot handle exceptional values. The default for C, C++, and Fortran compilers under both Compaq/DEC OSF/1 and GNU/Linux operating systems is to flush underflows abruptly to zero, and to immediately terminate execution on encountering an operand that is subnormal, Infinity, or NaN, or for which the instruction would generate a NaN or Infinity.

In order to produce IEEE 754 nonstop behavior on Compaq/DEC Alpha systems, special compilation options are required:

• -ieee for native C and C++ compilers,
• -fpe3 or -fpe4 for native Fortran 77, 90, and 95 compilers, and
• -mieee for GNU C, C++, and Fortran compilers.

These options cause the compilers to generate different floating-point instructions that cause traps to software for exceptional operands or results, and in addition, cause the insertion of trap barrier instructions after floating-point operations. The purpose of the latter is to flush the instruction pipeline, allowing precise determination of the interrupt location, so that the software handler can find the instruction and its operands, and complete the job.

Because instruction pipelining is extremely critical for modern high-performance CPUs, it should be expected that the performance hit from IEEE 754 nonstop behavior on Alpha processors may be severe, and that expectation is clearly demonstrated in the tables below.

Notes on the tables

The complete output data from which the tables below are derived are recorded in timops.raw, which should be consulted for details of operating systems, and absolute times. The timops.awk program filters that file to produce the table entries. Numerical entries in the last 5 columns are the slowdown (when > 1) compared to the loop with normal values.

There are several observations to make about the data in the tables below:

• The results are not expected to be particularly sensitive to the compiler used, or the optimization level chosen: the loop statement contains function calls which cannot be expanded inline, and which add sufficient overhead that loop unrolling is likely to be ineffective.
• The version of gcc on most systems is 2.95.3, or rarely 2.95.2 or earlier. gcc 3.x has yet to prove itself stable enough for use as a default production compiler across the wide range of systems that I work on, and support.
• Because the benchmarks use the same operands on each loop iteration, the results are essentially worst-case, and thus, overly pessimistic. In most practical computations, exceptional values are normally rare, although underflows are common in certain fields of computation (e.g., calculations in quantum chemistry usually involve exp(-x), with x often large enough that exp(-x) underflows).
• Entries with fp_size == 12 correspond to Intel IA-32 and IA-64 80-bit data, the preferred format on those architectures, and there is no penalty for underflow.
• Exceptional values in 32-bit and 64-bit formats on AMD Athlon and Intel IA-32 are 2 to 6 times more expensive than for normal values. This may be due partly to the fact that conversions to and from the native 80-bit format can be done only by load and store operations, which must access memory; there are no on-chip register-register instructions for format conversion.
• Entries with fp_size == 16 correspond to software implementations of 128-bit IEEE 754 arithmetic on all the systems measured. Only IBM S/390 G5 and HAL/Fujitsu SPARC64-GP implement such arithmetic in hardware, but no results for this benchmark are yet available on those systems.
• Newer processor generations in a particular architecture typically reduce the penalty for exceptional values: see the MIPS and Alpha results. This does not always hold true: UltraSPARC has a significant penalty for subnormals, but the older SuperSPARC does not.
• Implementations by different vendors of the same instruction set may exhibit marked differences: the AMD Athlon handles NaN and Infinity without penalty, while the Intel Pentium II and III suffer a 2.0 to 3.2 penalty for those operands.
• There are no penalties for exceptional values on Alpha 21246, IBM PowerPC, and SuperSPARC.
• AMD Athlon and Intel Pentium show no penalty in 80-bit native mode, but 32-bit and 64-bit underflows are 3 to 5 times slower on Athlon, and all exceptional values are 2.6 to 6 times slower on Pentium.
• NaN and Infinity carry similar penalties on most systems, except on MIPS and PA-RISC, where NaNs are 10 to 100 times slower than Infinity.
• Software implementations of 128-bit arithmetic are usually faster for NaNs and Infinity than for normal or subnormal values. This is to be expected, since the exceptional cases have to be explicitly tested for first.
• SGI IRIX on the MIPS architecture does not implement subnormal numbers, so underflows are always abrupt, unless a special run-time library routine is called to enable support for subnormals. Unfortunately, there does not appear to be any way with SGI Fortran, C, or C++ compilers to enable this support at compile time. SGI's "man sigfpe" gave me the necessary hints to fix this, and hoc versions 7.0.4.beta and later implement it. Here is the necessary code fragment, followed by call to it (from a C main() program):

  #if defined(__sgi)
  #include 
  
  static void
  flush_to_zero(int on_off)		/* see "man sigfpe" on SGI IRIX 6.x for documentation */
  {
  	union fpc_csr n;
  
  	n.fc_word = get_fpc_csr();
  	n.fc_struct.flush = (on_off ? 1 : 0);
  	set_fpc_csr(n.fc_word);
  }
  #endif
  ...
  #if defined(__sgi)
  	flush_to_zero(0);		/* to get support for subnormals! */
  #endif
  

• 128-bit arithmetic on IBM AIX PowerPC and SGI IRIX MIPS systems is implemented by paired doubles which exhibit many aberrations with exceptional values, both in the basic arithmetic, and in the runtime elementary function libraries.
• There are two sets of results for Alpha processors. The first, identified by values labeled -n/a- in the last three columns, is for the default terminate-on-exceptional-value operation; tests that involve NaN and Infinity then have to be suppressed. The second is for IEEE 754 nonstop operation, and except for the more recent 21264 processor, which handles exceptional values entirely in hardware, the results show a dramatic slowdown (8 to 66) compared to normal values.
• Examination of the absolute times in the two sets of Alpha result show no significant differences for the normal loop: there are simply too few floating-point operations in each loop iteration for the pipeline flushing to be evident. Real code with denser loops would show that difference.

Relative timing: sorted by CPU type

-----------------------------------------------------------------------------------------------
CPU                                 MHz Cmpiler fp_size   ufl->   ufl->   ofl->     NaN     Inf
                                                        subnorm    zero     Inf
-----------------------------------------------------------------------------------------------
AMD Athlon                         1400     gcc       4   4.974   3.376   1.000   1.000   0.991
AMD Athlon                         1400     gcc       8   4.802   3.198   1.009   1.000   1.009
AMD Athlon                         1400     gcc      12   1.007   1.000   1.000   1.007   1.013
DEC Alpha 21064 EV4                 100     gcc       4   1.001   1.007   -n/a-   -n/a-   -n/a-
DEC Alpha 21064 EV4                 100     gcc       4   8.252   8.139   8.261   7.532   7.517
DEC Alpha 21064 EV4                 100     gcc       8   1.006   0.999   -n/a-   -n/a-   -n/a-
DEC Alpha 21064 EV4                 100     gcc       8   8.835   8.697   9.572   7.886   7.815
DEC Alpha 21164 EV5                 466     c89       4   1.000   1.000   -n/a-   -n/a-   -n/a-
DEC Alpha 21164 EV5                 466     c89       4  43.636  34.727  21.879  21.121  21.439
DEC Alpha 21164 EV5                 466     c89       8   1.000   1.000   -n/a-   -n/a-   -n/a-
DEC Alpha 21164 EV5                 466     c89       8  66.043  49.217  27.913  27.130  27.333
DEC Alpha 21264                     667     c89       4   1.000   0.989   -n/a-   -n/a-   -n/a-
DEC Alpha 21264                     667     c89       4  53.359  42.239   0.989   1.000   0.989
DEC Alpha 21264                     667     c89       8   1.000   1.010   -n/a-   -n/a-   -n/a-
DEC Alpha 21264                     667     c89       8  78.552  57.885   1.000   1.021   1.000
DEC Alpha 21264                     667     c89      16   0.986   1.014   -n/a-   -n/a-   -n/a-
DEC Alpha 21264                     667     c89      16   1.057   1.000   0.986   0.957   0.986
HP PA-RISC 1.1 7100LC                80      cc       4  12.058  12.178   1.000  92.251   1.000
HP PA-RISC 1.1 7100LC                80      cc       8  16.955  16.841   1.000  11.278   1.000
IBM PowerPC                         133      cc       4   0.981   1.000   0.981   0.991   0.981
IBM PowerPC                         133      cc       8   1.007   1.007   1.014   1.000   1.007
IBM PowerPC                         133      cc       8   1.014   1.014   1.014   1.007   1.000
IBM PowerPC                         166      cc       4   0.991   1.000   0.991   0.991   0.991
IBM PowerPC                         166      cc       8   1.014   1.020   1.020   1.014   1.000
IBM PowerPC                         166      cc      16   1.009   1.009   0.991   0.991   0.991
IBM PowerPC                         233     gcc       4   1.006   1.013   1.026   1.000   1.000
IBM PowerPC                         233     gcc       8   1.006   1.013   1.019   1.000   1.000
IBM PowerPC                         533      cc       4   1.009   1.009   1.018   1.000   1.009
IBM PowerPC                         533      cc       8   0.991   0.991   0.991   0.983   0.991
IBM PowerPC                         533      cc       8   0.991   1.000   1.000   0.991   1.000
Intel IA-64 (emulated on IA-32)     600     gcc       4   1.012   1.018   1.009   0.941   0.953
Intel IA-64 (emulated on IA-32)     600     gcc       8   1.015   1.009   0.994   0.915   0.921
Intel IA-64 (emulated on IA-32)     600     gcc       8   1.015   1.011   0.998   0.917   0.923
Intel Pentium II                    450      cc       4   5.967   3.383   3.367   3.333   3.083
Intel Pentium II                    450      cc       8   3.655   2.236   2.227   2.291   2.145
Intel Pentium II                    450      cc      12   0.984   1.000   1.000   2.129   2.000
Intel Pentium II (Klamath)          300      cc       4   5.982   3.390   3.373   3.302   3.035
Intel Pentium II (Klamath)          300      cc       8   5.824   3.249   3.213   3.301   3.036
Intel Pentium II (Klamath)          300      cc      12   1.000   0.999   2.468   2.659   2.467
Intel Pentium III                  1266     gcc       4   6.014   3.408   3.394   3.317   3.056
Intel Pentium III                  1266     gcc       8   5.852   3.268   3.232   3.317   3.056
Intel Pentium III                  1266     gcc      12   1.010   1.010   1.000   2.588   2.402
Intel Pentium III (Katmai)          600     gcc       4   6.266   3.538   3.545   3.490   3.224
Intel Pentium III (Katmai)          600     gcc       8   6.176   3.437   3.423   3.514   3.246
Intel Pentium III (Katmai)          600     gcc      12   1.036   1.018   2.518   2.491   2.321
MIPS R10000                         180     c89       4   0.991   1.000   1.000  27.596   0.991
MIPS R10000                         180     c89       8   0.991   0.991   1.000  27.254   1.000
MIPS R10000                         180     c89      16   1.134   1.134   1.127   0.606   0.606
MIPS R10000                         195     c89       4   1.010   1.010   1.000  26.346   1.000
MIPS R10000                         195     c89       8   1.010   1.010   1.000  26.798   1.000
MIPS R10000                         195     c89      16   1.113   1.113   1.120   0.624   0.632
MIPS R4400                          150     c89       4  25.635  25.912   1.081  22.858   0.993
MIPS R4400                          150     c89       8  26.074  26.007   1.074  22.107   1.013
MIPS R4400                          150     c89      16  31.128  10.701   1.137   8.493   0.531
MIPS R4400                          175     c89       4  27.354  27.562   1.054  24.492   0.977
MIPS R4400                          175     c89       8  27.902  27.826   1.045  23.977   0.962
MIPS R4400                          175     c89      16  33.945  11.522   1.132   9.495   0.533
MIPS R5000                          180     c89       4   1.062   1.076   1.055  26.090   1.055
MIPS R5000                          180     c89       4   1.076   1.076   1.069  31.472   1.083
MIPS R5000                          180     c89       8   1.047   1.068   1.054  24.223   1.061
MIPS R5000                          180     c89       8   1.054   1.068   1.047  24.439   1.054
MIPS R5000                          180     c89      16   1.206   1.198   1.222   0.532   0.540
MIPS R5000                          180     c89      16   1.222   1.198   1.230   0.532   0.540
Sun UltraSPARC                      400     c89       4  16.675   1.031   0.995   1.015   1.015
Sun UltraSPARC                      400     c89       8  14.527   1.015   1.053   1.008   1.015
Sun UltraSPARC                      400     c89      16   1.015   1.026   1.031   0.701   0.716
Sun UltraSPARC II                   167     c89       4  16.761   1.017   1.009   1.009   1.017
Sun UltraSPARC II                   167     c89       8  12.586   1.006   0.994   1.000   1.006
Sun UltraSPARC II                   167     c89      16   1.002   1.000   1.002   0.705   0.701
Sun UltraSPARC II                   270     c89       4  18.618   0.993   0.993   1.000   0.986
Sun UltraSPARC II                   270     c89       8  14.640   0.995   1.000   0.995   1.000
Sun UltraSPARC II                   270     c89      16   1.000   1.000   1.000   0.697   0.701
Sun UltraSPARC II                   300     c89       4  16.961   1.008   1.008   1.008   1.008
Sun UltraSPARC II                   300     c89       8  13.068   1.000   1.000   1.011   1.000
Sun UltraSPARC II                   300     c89      16   1.000   1.004   0.989   0.706   0.709
Sun UltraSPARC II                   400     c89       4  16.777   1.005   1.000   1.000   1.000
Sun UltraSPARC II                   400     c89       8  12.818   1.008   1.000   1.000   1.000
Sun UltraSPARC II                   400     c89      16   1.010   1.010   1.010   0.694   0.694
Sun UltraSPARC II                   440     c89       4  16.824   0.995   0.995   1.000   1.000
Sun UltraSPARC II                   440     c89       8  13.198   1.008   1.016   1.016   1.000
Sun UltraSPARC II                   440     c89      16   1.021   1.000   1.021   0.688   0.704
Sun UltraSPARC IIe                  500     c89       4  16.981   1.000   1.013   1.000   1.006
Sun UltraSPARC IIe                  500     c89       8  13.179   1.000   1.000   1.009   1.000
Sun UltraSPARC IIe                  500     c89      16   0.994   1.000   0.994   0.697   0.690
Sun UltraSPARC III                  750     c89       4  13.417   0.942   0.897   0.942   0.910
Sun UltraSPARC III                  750     c89       8  11.223   1.000   0.995   1.000   1.000
Sun UltraSPARC III                  750     c89      16   0.992   1.000   0.992   0.659   0.675
TI SuperSPARC Viking                 40     gcc       4   1.000   1.009   0.991   0.991   0.991
TI SuperSPARC Viking                 40     gcc       8   0.996   1.000   0.984   0.988   0.984
TI SuperSPARC Viking                 40     gcc       8   1.016   1.012   1.000   1.000   1.000
TI SuperSPARC Viking/MXCC            50     gcc       4   1.005   1.000   0.995   0.995   0.989
TI SuperSPARC Viking/MXCC            50     gcc       8   1.005   1.000   0.990   0.990   0.995
TI SuperSPARC Viking/MXCC            50     gcc       8   1.010   1.010   1.000   1.000   0.995
-----------------------------------------------------------------------------------------------

Relative timing: sorted by precision and CPU type

-----------------------------------------------------------------------------------------------
CPU                                 MHz Cmpiler fp_size   ufl->   ufl->   ofl->     NaN     Inf
                                                        subnorm    zero     Inf
-----------------------------------------------------------------------------------------------
AMD Athlon                         1400     gcc       4   4.974   3.376   1.000   1.000   0.991
DEC Alpha 21064 EV4                 100     gcc       4   1.001   1.007   -n/a-   -n/a-   -n/a-
DEC Alpha 21064 EV4                 100     gcc       4   8.252   8.139   8.261   7.532   7.517
DEC Alpha 21164 EV5                 466     c89       4   1.000   1.000   -n/a-   -n/a-   -n/a-
DEC Alpha 21164 EV5                 466     c89       4  43.636  34.727  21.879  21.121  21.439
DEC Alpha 21264                     667     c89       4   1.000   0.989   -n/a-   -n/a-   -n/a-
DEC Alpha 21264                     667     c89       4  53.359  42.239   0.989   1.000   0.989
HP PA-RISC 1.1 7100LC                80      cc       4  12.058  12.178   1.000  92.251   1.000
IBM PowerPC                         133      cc       4   0.981   1.000   0.981   0.991   0.981
IBM PowerPC                         166      cc       4   0.991   1.000   0.991   0.991   0.991
IBM PowerPC                         233     gcc       4   1.006   1.013   1.026   1.000   1.000
IBM PowerPC                         533      cc       4   1.009   1.009   1.018   1.000   1.009
Intel IA-64 (emulated on IA-32)     600     gcc       4   1.012   1.018   1.009   0.941   0.953
Intel Pentium II                    450      cc       4   5.967   3.383   3.367   3.333   3.083
Intel Pentium II (Klamath)          300      cc       4   5.982   3.390   3.373   3.302   3.035
Intel Pentium III                  1266     gcc       4   6.014   3.408   3.394   3.317   3.056
Intel Pentium III (Katmai)          600     gcc       4   6.266   3.538   3.545   3.490   3.224
MIPS R10000                         180     c89       4   0.991   1.000   1.000  27.596   0.991
MIPS R10000                         195     c89       4   1.010   1.010   1.000  26.346   1.000
MIPS R4400                          150     c89       4  25.635  25.912   1.081  22.858   0.993
MIPS R4400                          175     c89       4  27.354  27.562   1.054  24.492   0.977
MIPS R5000                          180     c89       4   1.062   1.076   1.055  26.090   1.055
MIPS R5000                          180     c89       4   1.076   1.076   1.069  31.472   1.083
Sun UltraSPARC                      400     c89       4  16.675   1.031   0.995   1.015   1.015
Sun UltraSPARC II                   167     c89       4  16.761   1.017   1.009   1.009   1.017
Sun UltraSPARC II                   270     c89       4  18.618   0.993   0.993   1.000   0.986
Sun UltraSPARC II                   300     c89       4  16.961   1.008   1.008   1.008   1.008
Sun UltraSPARC II                   400     c89       4  16.777   1.005   1.000   1.000   1.000
Sun UltraSPARC II                   440     c89       4  16.824   0.995   0.995   1.000   1.000
Sun UltraSPARC IIe                  500     c89       4  16.981   1.000   1.013   1.000   1.006
Sun UltraSPARC III                  750     c89       4  13.417   0.942   0.897   0.942   0.910
TI SuperSPARC Viking                 40     gcc       4   1.000   1.009   0.991   0.991   0.991
TI SuperSPARC Viking/MXCC            50     gcc       4   1.005   1.000   0.995   0.995   0.989
AMD Athlon                         1400     gcc       8   4.802   3.198   1.009   1.000   1.009
DEC Alpha 21064 EV4                 100     gcc       8   1.006   0.999   -n/a-   -n/a-   -n/a-
DEC Alpha 21064 EV4                 100     gcc       8   8.835   8.697   9.572   7.886   7.815
DEC Alpha 21164 EV5                 466     c89       8   1.000   1.000   -n/a-   -n/a-   -n/a-
DEC Alpha 21164 EV5                 466     c89       8  66.043  49.217  27.913  27.130  27.333
DEC Alpha 21264                     667     c89       8   1.000   1.010   -n/a-   -n/a-   -n/a-
DEC Alpha 21264                     667     c89       8  78.552  57.885   1.000   1.021   1.000
HP PA-RISC 1.1 7100LC                80      cc       8  16.955  16.841   1.000  11.278   1.000
IBM PowerPC                         133      cc       8   1.007   1.007   1.014   1.000   1.007
IBM PowerPC                         133      cc       8   1.014   1.014   1.014   1.007   1.000
IBM PowerPC                         166      cc       8   1.014   1.020   1.020   1.014   1.000
IBM PowerPC                         233     gcc       8   1.006   1.013   1.019   1.000   1.000
IBM PowerPC                         533      cc       8   0.991   0.991   0.991   0.983   0.991
IBM PowerPC                         533      cc       8   0.991   1.000   1.000   0.991   1.000
Intel IA-64 (emulated on IA-32)     600     gcc       8   1.015   1.009   0.994   0.915   0.921
Intel IA-64 (emulated on IA-32)     600     gcc       8   1.015   1.011   0.998   0.917   0.923
Intel Pentium II                    450      cc       8   3.655   2.236   2.227   2.291   2.145
Intel Pentium II (Klamath)          300      cc       8   5.824   3.249   3.213   3.301   3.036
Intel Pentium III                  1266     gcc       8   5.852   3.268   3.232   3.317   3.056
Intel Pentium III (Katmai)          600     gcc       8   6.176   3.437   3.423   3.514   3.246
MIPS R10000                         180     c89       8   0.991   0.991   1.000  27.254   1.000
MIPS R10000                         195     c89       8   1.010   1.010   1.000  26.798   1.000
MIPS R4400                          150     c89       8  26.074  26.007   1.074  22.107   1.013
MIPS R4400                          175     c89       8  27.902  27.826   1.045  23.977   0.962
MIPS R5000                          180     c89       8   1.047   1.068   1.054  24.223   1.061
MIPS R5000                          180     c89       8   1.054   1.068   1.047  24.439   1.054
Sun UltraSPARC                      400     c89       8  14.527   1.015   1.053   1.008   1.015
Sun UltraSPARC II                   167     c89       8  12.586   1.006   0.994   1.000   1.006
Sun UltraSPARC II                   270     c89       8  14.640   0.995   1.000   0.995   1.000
Sun UltraSPARC II                   300     c89       8  13.068   1.000   1.000   1.011   1.000
Sun UltraSPARC II                   400     c89       8  12.818   1.008   1.000   1.000   1.000
Sun UltraSPARC II                   440     c89       8  13.198   1.008   1.016   1.016   1.000
Sun UltraSPARC IIe                  500     c89       8  13.179   1.000   1.000   1.009   1.000
Sun UltraSPARC III                  750     c89       8  11.223   1.000   0.995   1.000   1.000
TI SuperSPARC Viking                 40     gcc       8   0.996   1.000   0.984   0.988   0.984
TI SuperSPARC Viking                 40     gcc       8   1.016   1.012   1.000   1.000   1.000
TI SuperSPARC Viking/MXCC            50     gcc       8   1.005   1.000   0.990   0.990   0.995
TI SuperSPARC Viking/MXCC            50     gcc       8   1.010   1.010   1.000   1.000   0.995
AMD Athlon                         1400     gcc      12   1.007   1.000   1.000   1.007   1.013
Intel Pentium II                    450      cc      12   0.984   1.000   1.000   2.129   2.000
Intel Pentium II (Klamath)          300      cc      12   1.000   0.999   2.468   2.659   2.467
Intel Pentium III                  1266     gcc      12   1.010   1.010   1.000   2.588   2.402
Intel Pentium III (Katmai)          600     gcc      12   1.036   1.018   2.518   2.491   2.321
DEC Alpha 21264                     667     c89      16   0.986   1.014   -n/a-   -n/a-   -n/a-
DEC Alpha 21264                     667     c89      16   1.057   1.000   0.986   0.957   0.986
IBM PowerPC                         166      cc      16   1.009   1.009   0.991   0.991   0.991
MIPS R10000                         180     c89      16   1.134   1.134   1.127   0.606   0.606
MIPS R10000                         195     c89      16   1.113   1.113   1.120   0.624   0.632
MIPS R4400                          150     c89      16  31.128  10.701   1.137   8.493   0.531
MIPS R4400                          175     c89      16  33.945  11.522   1.132   9.495   0.533
MIPS R5000                          180     c89      16   1.206   1.198   1.222   0.532   0.540
MIPS R5000                          180     c89      16   1.222   1.198   1.230   0.532   0.540
Sun UltraSPARC                      400     c89      16   1.015   1.026   1.031   0.701   0.716
Sun UltraSPARC II                   167     c89      16   1.002   1.000   1.002   0.705   0.701
Sun UltraSPARC II                   270     c89      16   1.000   1.000   1.000   0.697   0.701
Sun UltraSPARC II                   300     c89      16   1.000   1.004   0.989   0.706   0.709
Sun UltraSPARC II                   400     c89      16   1.010   1.010   1.010   0.694   0.694
Sun UltraSPARC II                   440     c89      16   1.021   1.000   1.021   0.688   0.704
Sun UltraSPARC IIe                  500     c89      16   0.994   1.000   0.994   0.697   0.690
Sun UltraSPARC III                  750     c89      16   0.992   1.000   0.992   0.659   0.675
-----------------------------------------------------------------------------------------------