[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Alpha recommendations

On Thu, 15 Oct 1998, Christopher C Chimelis wrote:
> On 15 Oct 1998, John Goerzen wrote:
> > I am seriously considering a new Alpha in the 600 MHz range as my next 
> > computer.  The benchmarks I've seen indicate that such a machine will
> > easily trounce a Dual PII/450 that I had been looking at, and the
> > Alpha is about the same price, so -- the decision is not terribly
> > difficult.  However, I've got some questions for y'all:
> Well, benchmarks can often be deceiving.  While I love my Alphas, I do
> only compare them to Intels when applicable.  For instance, application
> support on the Alpha tends to be "not as rich" as the Intels, mostly
> because it's a newer platform supported by the Linux kernel.  In terms of
> FP performance, Alphas usually kill Pentiums easily, but it's not as good
> under Linux as under DEC Unix due to unoptimised math libraries and
> routines.  This is being worked on heavily in glibc 2.1, however, and
> can only improve at this point.  Integer performance, however, tends to be
> on par with Intels or sometimes below depending on which Alpha processor
> is being used to compare.

Ah yes, benchmarks.  People most often quote SPECint95/SPECfp95.  That's 
only one view of Alpha performance, and a very narrow one at that... 
I always suggest folks try their own benchmarks.  I downloaded the 
BYTEmark source code to test out various compiler options.  The BYTEmark 
integer index is normalized to a P90.  Here's my results for several 

    CPU/OS                   Integer      FP
    -----------------------  -------  ------
    Pentium 233MHz, WinNT4:     2.63    2.22
    Sun UltraSPARC, 350MHz,
    Solaris 2.6:                3.54    3.11
    Alpha 164SX, 533MHz,
    Debian/Linux:               8.17    2.92

(Tests were performed with GCC 2.7.2 on Pentium and UltraSPARC, and 
egcs-1.1 on Alpha.)  Clearly my 164SX does well on integer.  The fp 
results are disappointing, but not suprising since they depend on sqrt(), 
sin() and cos() which are poorly optimized in glibc.

My point is that there is no such thing as a "typical" benchmark.  People 
quote SPEC saying that Alpha is comparable to Intel on integer, and 
superior on fp, but that's just one data point, and may be a very poor 
representation of the real-life appliactions you'll be using.

More surprising is how all these benchmarks seem to focus on CPU 
performance.  System performance can be measured by CPU, memory, I/O, 
network, video and other factors... it would seem that the Alpha is 
superior too in memory and I/O areas but I don't have numbers to back it 
up, other than the standard "hdparm -tT" utility (which gives me buffered 
disk writes near 100MB/s on my 164SX... I don't have an Intel machine 
that gives me over 50MB/s on the same test).

Anyway, forgive my ramblings, back to work now :)


Reply to: