Boxx shipinabottle benchmark speed

Well, I had problems with the Boxx "laptop" (if you've got a pretty
hefty and flame resistant lap), so I traded them for a tower with much
better specs and slightly less portability.

Anyway, so far it's a screamer. 3 Gb RAM, Athlon FX57, 10,000 rpm HDD,
single nVidia 3450. Gets through 50 builds of the ship bench in 17.4
seconds. Moving on to the Spec next.

matt

 

Consider my FX53 beat.

What did you get on STAR and Patbench?

 

I've had problems downloading from Spec.org, and I've been working at
customer site. I'll get to the others this weekend and post results.

 

Great numbers, Matt! Any chance of posting a complete
hardware list for those of us about to go shopping?

Art

 

-Boxx 3204 (off the shelf - didn't pay the listed price)
-AMD Athlon 64 FX57 (single processor, single core)
-3Gb DDR400 RAM
-nVidia FX 3450 PCIe 256 Mb (single card, dual capable)
-74 Gb 10,000 rpm SATA drive
-XP Pro (32 bit)
-Asus mobo, nForce chipset, SLi capable
-500W power supply

http://www.boxxtech.com/products/cf_step2.asp?ModelInstanceID=601
&cfg8728=7351&cfg8729=5635&cfg8759=7350&cfg8734=3784&cfg8735=3784
&cfg8740=6556&cfg8744=3094&cfg8745=2446&cfg8758=2446&cfg8748=2446
&cfg8749=2446&cfg8750=2446&cfg8751=4439&cfg8752=2446&cfg8753=2446
&cfg8754=2399&cfg8755=3730&cfg8756=2446&cfg8757=2446&cmdUpdateTotal2.x=
40&cmdUpdateTotal2.y=10

 

Interesting Star 2.1 results

sw05 sw06
35.8 38.2 time
3.70 3.35 rebuild

So the "time" gets longer, but the "rebuild time" is less in 2006.
What's the difference between time and rebuild time?

Patbench

4 = 2.39 s
8 = 6.18 s
10 = 15.4 s
11 = 105.0 s

Specapc (best scores from spec.org site in parentheses)

sw2005.Result.score = 2.88 (2.2)
sw2005.Result.graphics = 1.6 (2.34)
sw2005.Result.cpu = 5.78 (2.03)
sw2005.Result.io = 3.38 (2.18)

The graphics score is kind of middle of the road, but the cpu is way
better than the top scores on the Spec.org site. I didn't think the
number was valid, but I ran it twice and got similar results. I/O is
also about double the average score from the spec.org site.

 

STAR is timed by a real time timer. As such you have to average a few
scores to get a good idea because if other things are running it will
slow down. The first time is on the recursive routine that models the
cubes. It is run outside of the graphics system using some tricks SW
has in their API toolbox. The second time, for rebuild, is just that,
a timer on either side of a call to rebuild and it is again outside the
graphics system.

The first routine simply draws a box as a set of coordinates passed to
a subroutine and then extrudes it. The algorithm came from a book
called Algorithms in C by Sedgewick. When you watch it work in the
graphics system (turn off dbadd and display calls) you will see that as
the feature tree gets longer it also takes longer for each sketch and
box. The only explanation for this is that the nature of the algorithm
coupled with the way SW establishes parent child relations increases
the time for each feature. If you analyze the feature tree you will see
that the last feature is actually the child of all the features in the
feature tree even though parent child will not necessarily tell you
this. The second routine simply calls a forced rebuild. This suggests
that there is some difference between the checking that goes on when a
feature is built and when it is first created.

PATBENCH is older. I discovered by accident that it can run a system
out of memory. I think originally I wanted to see what all this
limitation on the size of a pattern was all about. It turned out to
also show that patterns are real memory hogs during creation. Curiously
different releases of SW behave quite differently to Patbench. I think
2006 actually showed a small performance improvement on this benchmark.
The big jump between 10 and 11 iterations may be going from L2 cache
to main memory. I'm not 100% sure about that though.

On SPECapc it doesn't surprise me that CPU and I/O are higherl. The
FX57 is just plain fast. It looks like they put a Western Digital
Raptor in. That coupled with the large L2 cache and memory bandwidth of
the AMD chips should help I/O.

One more small favor. Would you run Ship in a Bottle per the
instructions in http://www.engtran.com/SWBM001.html? That will give a
better feel for cpu vs graphics performance.

 

Boxx doesn't have "preconfigured" systems just waiting to be shipped.
You spec the box from the list of components on their site, they put it
together, test and ship it. What I meant by "off the shelf" was that I
haven't done any customization or added components. I haven't done any
bios or registry tweeks. I got a system without an OS and installed
that myself so there was no unnecessary junk and the hard drive was
configured the way I wanted it.

If you configure the system the way I did on their site, you come up
with a number roughly 10% higher than what I paid. Boxx doesn't have
discounts per se (except for volume, and even then it's not much), but
maybe they had a price increase since I ordered.

If you're looking at a Boxx system (www.boxxtech.com), I'd encourage you
to also look at Xi (www.xicomputer.com), because they seem to be
equivalent machines at a slightly lower price. Boxx comes at a premium
for some reason (they cater to the CGI/movie industry).

The reason I picked what I did, which is actually their lowest model
line, is that it was the only machine that you could spec an Athlon FX
chip in (as opposed to an Opteron). A Boxx rep told me, and it seems to
be true, that the Opterons are optimized for multiprocessor use, so
getting a single Opteron didn't seem like a good idea. SolidWorks
doesn't benefit much from multiprocessors, so I decided to go single
processor, single core, but get the best one available. The Athlon FX
series was touted as being the absolute nuts for single threaded games,
and since kelner (TOP) had reported great results with one of the FX
chips, I figured it wasn't a bad bet.


Good luck,

Matt

 

HI LO
24.4 18.2 edges shown
23.4 17.5 no edges
26.4 18.5 wireframe (not shaded)
27.3 18.9 hlr (not shaded)
25.3 19.2 hlg (not shaded)

Using low quality transparency, zooming out until invisible, changing
from 16 to 32 bit colors and turning off the gradient background don't
seem to improve things at all.

Moving the ship off the screen improved an 18.5 run to 17.75.

Reducing the size of the SW window from approx 1024x728 to 800x600
(system set up at 1600x1200 resolution) improved an 18.5 run to 17.5.

Reducing the graphics window to nothing (only can see the feature
manager) reduces the 18.5 time to 15.8.

On the Spec.org site, the computer that seemed to really crank is one
with dual video cards. I've got a very graphics intense project looming
on the horizon. Maybe I'll see if I can pop in a second 3450 to save
some time on that.