By Mark Duell
I was watching the keynote last night, and I was amazed by the SPEC scores
presented. So amazed in fact, that I had to go to spec.org to see if they were
real. While I won't call Apple's numbers fraudulent outright, they had a number
of curious details.
1) Selective use of hyperthreading. On the dual processor tests, where HT
would help the most due to the program using multiple threads, HT was disabled;
but on the single processor tests where spec only uses one thread and HT would
be detrimental due to resource contention, Apple conveniently enabled HT.
2) Compiler choice. By choosing gcc for both platforms, Apple puts the Dell
at a disadvantage. Better compilers exist for x86 hardware, namely ICC, and
gcc has different optimizations for each platform. Similar circumstances on the
benchmarks that are written in Fortran, Apple using NAGWare Fortran on the PC
instead of the superior Intel Fortran or Compaq Visual Fortran. Note Jobs's
error of omission, since he only talked about using gcc for the benchmarks.
3) OS choice. Instead of using a publicly released OS, Apple chose to use a
internal beta version of 10.2.7 on their hardware; this is fine, as I'm all in
favor of using the best available (to debate the availability of an internal
beta is a discussion for another time). But on the Dell they decided to use
RedHat Linux. This seems odd, since all of the officially published spec
benchmarks for x86 use Windows 2000/XP/2003 (all of which have been publicly
released), and Windows is used far more often in homes and businesses.
4) The little things. According to page 27 of VeriTest's report, on they Dell
they use -O3 -march=pentium4 -mfpmath=sse. According to the Pentium 4
optimization manual, scalar SSE/SSE2 math is slower than the plain old 387 math
on the Pentium 4. Specifically, the 387 math can execute one FP addition per
cycle, whereas SSE2 executes one scalar FP addition every two cycles.
Incompetence or malice, neither looks good for VeriTest.
Debunking the price myth:
Take a look at the low end
|
Apple PowerMac G5 |
Dell Precision 650 |
| Dual 2.0Ghz G5 (fp: 15.7, int: 17.2, average:
16.45) |
Dual 2.4Ghz Xeon (fp: 14.7, int: 18.7,
average: 16.7) |
| 512MB PC3200 RAM |
512MB PC2100 RAM |
| 160GB SerialATA HDD |
Dual 80GB ATA/133 HDD |
| DVD/CDRW |
DVD/CDRW |
| ATI Radeon 9600 Pro |
nVidia Quadro NVS 280 |
| Built in audio and FireWire |
Sound Blaster Audigy II with 1394 |
| 17" Cinema Display |
17" UltraSharp Display |
| Accessory Kit (Mouse and Keyboard) |
Mouse and Keyboard |
| Mac OS X 10.2 |
Windows XP Professional with SP1 |
| AppleCare 3 year warranty with bring-in
service |
Dell 3 year warranty with onsite labor |
| Price: $3,718 not including sales tax and
shipping |
Free Palm Zire |
| Shipping: August |
Price: $3,242, not
including sale tax and shipping |
| |
Shipping: Today |
What about the high end?
|
Apple PowerMac G5 |
Dell Precision 650 |
| Dual 2.0Ghz G5 (fp: 15.7, int: 17.2, average:
16.45) |
Dual 2.4Ghz Xeon (fp: 14.7, int: 18.7,
average: 16.7) |
| 2.0GB PC3200 RAM |
2.0GB PC2100 RAM |
| 250GB SerialATA HDD |
Dual 120GB ATA/133 HDD |
| SuperDrive DVD-RW |
4x DVD+RW |
| ATI Radeon 9800 Pro 128MB |
nVidia QuadroFX 500 128MB |
| Built in audio and FireWire |
Sound Blaster Audigy II with 1394 |
| Dual 20" Cinema Displays |
Dual 20" UltraSharp Displays |
| Accessory Kit (Mouse and Keyboard) |
Mouse and Keyboard |
| Mac OS X 10.2 |
Windows XP Professional with SP1 |
| AppleCare 3 year warranty with bring-in
service |
Dell 3 year warranty with onsite labor |
| Price: $7,121 not including sales tax and
shipping |
Free Palm Zire |
| Shipping: August |
Price: $5,492, not
including sale tax and shipping |
| |
Shipping: Today |
The quoted fp/int numbers are from Apple (for the PowerMac), and SPEC (for
the Precision). The Dell lacks FireWire 800 and the memory in the Dell is about
33% slower. The graphics cards are different due to options on the BTO sites,
but have similar street prices.
The real SPEC numbers:
| |
Dual 2.0Ghz G5 |
Dual 3.06Ghz Xeon |
Dual 1.8Ghz Opteron* |
| SPECfp_rate_base2000 |
15.7 |
15.7** |
24.7 |
| SPECin_rate_base2000 |
17.2 |
21.7 |
25.0 |
* For comparison only, server oriented machine
** Highest score on a workstation oriented machine. A server oriented machine
scored 16.7.
Although I advocate using the best software available for each architecture,
when introducing the SPEC benchmarks Jobs says "We're going to use gcc 3.3
because we want to use an open source compiler that's the same compiler on all
machines, and the compiler we all use, by the way." If he wants "open source"
and "the same" why not use Linux on both?
Have a comment? Post it here. |
