Slashdot Mirror
x86 vs PPC Linux benchmarks
Jay Carlson writes "We've all heard about how Apple's hardware is really fast compared to PCs. ("Supercomputer!" "Twice as fast as a Pentium!" "Most powerful laptop on the planet!") So, if you *aren't* going to use Photoshop and Final Cut Pro, how fast is it? I care more about integer apps like compilers, so I did some careful benchmarking of a few x86 and PPC Linux boxes. Submissions welcome."
From http://fampm201.tu-graz.ac.at/karl/timings40.html.
Numbers are relative to a G3-300MHz (higher are faster).
There are more numbers on the homepage.
Athlon 1.2 GHz, 512MB, Windows 2000 [73]: 4.78993
Athlon 1.2 GHz, 512MB, Linux [72]: 4.4734
Gateway Select 1000, AMD Athlon 1000 MHz (1GHz), 512KB L2, 192 MB, Linux [65]: 3.77305
Kryotech 1GHz AMD Athlon, 512k cache, 512MB, Linux [66]: 3.69674
Gateway Select 1000, AMD Athlon 1000 MHz (1GHz), 512KB L2, 192 MB, Linux [60]: 3.57748
Dell Dimension XPS B1000r, 512MB Ram, Win98 SE [64]: 3.38084
Dell 4100, 933MHz, 128MB, Linux [63]: 3.19988
COMPAQ AlphaStation XP1000, 2 GB RAM, 4MB L2, Digital Unix 4.0F [50]: 3.16987
AMD Athlon, 800 MHz, 512 KB L2, 256 MB, Linux [61]: 3.00154
Dual Xenon 866, 512 MB, Windows 2K [71]: 2.9618
PenguinComputing, dual 800Mhz PIII, 128Mb, Linux [67]: 2.76745
Athlon 700, asus k7m, 512 mb, win98 [43]: 2.70199
Dell 800 Mhz, Pentium III, 512 MB RDRAM, Win 98 SE [57]: 2.69821
Athlon 700 MHz, 128 MB, Linux [51]: 2.67027
Athlon 650 MHz, 256 MB, 100 MHz, Win NT 4.0 [42]: 2.60662
Compaq-Digital Alpha 8200, 625MHz, 1GB RAM, DEC-UNIX [11]: 2.48495
SONY VAIO F409 notebook, PIII 650 MHz, 128MB RAM, Red Hat LINUX 6.1 [55]: 2.21137
Athlon 650 Mhz, 32 Mb, Windows 98 [32]: 2.14558
Athlon 550MHz, 128MB, Red Hat Linux 6.1 [45]: 2.13797
Dell XPS T600r, P3 Copper wl 256kb, 128MB, Linux-2.0.36 [39]: 2.09939
Dell Precison 410, 2 PIII 550MHz, 256Mb RAM, Windows NT 4 [28]: 1.8561
Dell Precision 210, 550 MHz, 128 Mb, NT 4.0 [34]: 1.83606
Dell XPS T550, PIII-550, 128MB, RedHat-5.2, Linux [24]: 1.777
Gateway GP7-500, 500 MHz, 192 MB, WinNT 4 [6]: 1.70318
PowerMac 8500, 500 MHz MACh Carrier G3, 1 MB L2, 256 MB, MacOS 8.6 [35]: 1.68249
Instead of pitting his machines against one another, he should've assembled them into a BEOWULF CLUSTER so they could WORK TOGETHER.
The sole reason Photoshop is faster on Macs than PCs (recently), and the reason Steve always brings it out to show off, is altivec. We need benchmarks of vectorized vs non-vectorized versions of the same task, running on the same G4 or against an x86 box. Unfortunately I can count the programs that use altivec in the real world on one hand; hopefully this will change with OS X arriving (altivec in system libraries).
According to an article in the linux journal
the Altivec unit handles only single precision
floats
-> only useful for special tasks or for calculations were you can take care for the
reduced precision,
-> not for general scientific numerical calculations
No, the Apple Store price I quoted in that table is for a single-processor G4/533. Go price it yourself.
(I'd give a URL but the Apple Store is too sessional.)
All this means is that compiling on a RISC architecture is bound to be a great deal slower.
I'm aware of that, and I talk about it on that page. See the "Choice of workloads" section.
The install-egcs test does measure native compilation performance. This is relevant for people who use the box for development.
The install-glibc and cross-gcc tests are both compiling to a single RISC architecture, little-endian MIPS. The amount of effort required to optimize for PPC or x86 doesn't factor into this.
If you don't care about development compile times, just look at the cross-only numbers.
Though it might more accurately be described as a GCC/Linux PPC vs. x86 benchmark, it's still interesting to those of us who are committed to GCC and Linux. Those two items, and their spawn, suck up almost all the (used) cycles on all of my machines.
Geeky modern art T-shirts
Hi. I'm a GCC maintainer. I don't work on this part of the compiler, but I can speak to this point:
Nearly all of the optimizations in GCC are not machine-specific. Those kinds of optimizations, ones which are specific to the processor, are called peephole optimizations, and while every little bit helps, they don't make that much of a difference. The big ones are done at an intermediate level, before the compiler "knows" what processor it's using and starts to chunk out the opcodes.
More specifically, unlike the Linux kernel, glibc, and other major projects, GCC is not designed for and targeted primarily at Intel chips. The x86 is just one more back-end like any other; sometimes it falls behind and sometimes it pulls ahead, development-wise.
Some have, some have but won't be in the upcoming 3.0 release in a few weeks, and some are yet to come.
The biggest problem is that many of the really cool optimizations -- the ones that make a big difference and aren't CPU-specific -- have been patented by IBM and other major players.
You cannot apply a technological solution to a sociological problem. (Edwards' Law)
I can give you over a thousand examples, with the help of our f[r]iendly patent office (my tax dollars at work). Just go to http://www.uspto.gov/, look under the green Patent Grants area, follow the Advanced Search link, and search on "compiler and optimization". Doing this today, I got 1,261 patents, but some of them don't apply here.
Er, that is to say, I got 1,261 search results each representing a patent. I don't have 1,261 patents myself. :-)
You cannot apply a technological solution to a sociological problem. (Edwards' Law)
Lies,
Damn Lies, and
Benchmarks
I read through these comments and get the impression that slowly a conclusion is being reached: PPC-type hardware is good for some things, x86 hardware is good for others. Nothing really new there, is there? For running Linux, it seems from this little (and far from in-depth) benchmarking session that PCs are a bit better, especially given costs. You can probably get a 1.2GHz Athlon box for the cost of a 533MHz G4, and it'll be better for Linux, so if you run Linux, why not?
MacOS X, stuff like Maya, Final Cut Pro, etc. etc. quite obviously runs better on PPCs, barring some strange circumstances. I imagine that with enough "brute force" (RAM, dual processors, etc.) one could get a PC to run this stuff faster than a Mac.. but what's the point? You might as well just keep it simple and buy a Mac that'll run it pretty well outta the box.
I agree though, that cost is an important consideration. With 760MP around the corner, if it ever does surface in quantities making it available, dual Athlons might give dual G4s a bit of a whippin', especially considering AMDs prices as of late. In general I find you can buy a PC with a much faster proc, more RAM, etc. for the same cost as a Mac from the Apple store.
Still, even a 1.2GHz Athlon would probably choke on OS X, and the G4 will at most hiccup...
"Caffeine is not an option. Caffeine is a way of life."
I wind up doing my own internal PPC vs X86 benchmarks almost every year.
I'll set up whatever current game I am working on to run with the graphics stubbed out so it is strictly a CPU load. We just did this recently while putting the DOOM demo together for MacWorld Tokyo.
I'll port long-run time off line utilities.
I'll sometimes write some synthetic benchmarks.
Now understand that I LIKE Apple hardware from a systems standpoint (every time I have to open up a stupid PC case, I think about the Apple G3/G4 cases) , and I generally support Apple, but every test I have ever done has had x86 hardware outperforming PPC hardware.
Not necessarily by huge margins, but pretty conclusively.
Yes, I have used the Mr. C compiler and tried all the optimization options.
Altivec is nice and easy to program for, but in most cases it is going to be held up because the memory subsystems on PPC systems aren't as good as on the PC.
Some operations in Premier or Photoshop are definitely a lot faster on macs, and I would be very curious to see the respective implementations on PPC and X86. They damn sure won't just be the same C code compiled on both platforms, and it may just be a case of lots of hand optimized code competing against poorer implementations. I would like to see a Michael Abrash or Terje Mathison take the x86 SSE implementation head to head with the AltiVec implementation. That would make a great magazine article.
I'll be right there trumpeting it when I get a Mac that runs my tests faster than any x86 hardware, but it hasn't happened yet. This is about measurements, not tribal identity, but some people always wind up being deeply offended by it...
John Carmack
I'm sure this person THINKS he's testing the same thing on each machine. Unfortunately, he's not.
Firstly, the action of compiling on different architectures is very different, even without considering optimisation strategies. To compile code into the CISC code of the x86 architecture is very different from that of a RISC chip such as the PowerPC. For a start, instruction ordering etc. for a RISC chip, even for not really optimised code can take far more processing time. Then, if you add optimisation, which in a RISC architecture is a FAR more complex task.
All this means is that compiling on a RISC architecture is bound to be a great deal slower.
Basically, this "benchmark" is measuring not only the intrinsic speed differences of the architectures and chips but also degree of optimisation the native compilers used can cope with and the extra processing power needed to generate the code during the comile stages.
Basically, using compilation as a benchmark is not at all useful, other than to test the difference in speed of two similar peices of equipment using identical software (ie. compilers & OS) or the difference between two versions of the same OS or two versions of the same compiler.. Basically, you can only change one variable to deliver a meaningful benchmark if using the method chosen in this "study."
The only way to get a half-way meaningful benchmark for the two systems used here would be to write a program which did lots on disk I/O and integer manipulation, worrying about whether it's being biased for or against certain types of architecture or use (eg. loops sitting in cache etc.). This would give you an idea of the real-world speed differences between the two systems. However, you won't be just measuring the intrinsic speed of the machine but also the different ways the kernels have to do things on the two architectures, the degree of optimisation the compilers building the kernel and the program could generate and the speed of the hard disk built into the machine.
As you can see, it's a tricky thing comparing two types of machine.
Agrajag: "Oh no, not again!"
Everybody knows Linux is so fast it can execute an endless loop in 5 seconds flat.
"A door is what a dog is perpetually on the wrong side of" - Ogden Nash
I don't thnk that you will ever get an agument from anyone over how very, very, very badly Motorola has messed up with being unable to deliver faster versions of the PowerPC. However, you should look at a couple of issues with these benchmarks: 1) 450/733 = .61 and 533/733 = .72. Thus any scores over .61 and .72 respectively, indicate that the PowerPC is doing more per clock cycle than then PIII. If Motorola can ever get their act together (and that is not a certain), normal code on the PowerPC will run every bit as fast and faster than the x86 processor. Combined with the fact that the PowerPC has a nice quiet and fairly energy efficient air-cooled chip you might have some nice machines. Unfortunately, all benchmarks can have some rather un-intentional bias. My 1.2 Athlon would do 105 Scimarks in Windoze 98, 113 Sciemanrks in Wine under Redhat, and 119 under Windoze 2000 for Tim Wilkin's Science Mark benchmark.
Same machine, same memory, same disks, etc. the only difference was the os. Even given the same os, the tweaks that it goes through are also a function of the author's machine. Please pass the salt, I need a grain.
Sex is heriditary, if your parents didn't have it chances are good you won't either.
http://www.jdkoftinoff.com/eqtest.tar.gz
(gpl'd with source)
450 mhz g4:
1.7 gigaflops with altivec
410 megaflops without altivec
500 mhz pentium:
220 megaflops
--jeff
ipv6 is my vpn