Computer Hardware Performance and Benchmarks

I have been thinking about computer system performance a lot of late...

Our current engineering systems are Dell Precision 380's, Intel P4 3.2 Ghz single core workstations.  Each workstation has 2 gigs of RAM, Nvidia FX 1400 graphics cards and Western Digital Raptor SATA hardrives, 80 gig, 10K rpm.  My system is also a Dell Precision 380 but with a Pentium D940 3.2 Ghz dual core processor, 4 gigs of RAM, a Nvidia FX 3450 graphics card with the same WD Raptor harddrive.

On most of the designs we work on, computer performance is not something that is a major concern for us.  However, with one of our die asm designs, we are punching a pretty big linear array of cavities and I am noticing some long re-build times for my models.  The main array contains a large number of punches.  Plus there are tractor hole punches and other punches of various shapes.  Then you have all the extra holes for punch keeper attach, dowels and attaching the various die, stripper, punch holder and punch back-up plate details to the die asm.  You end up with somewhere in the range of 1200 holes in the punch holder.  No surprise on the longer re-build times as big linear arrays are very processor intensive and push a computer system to it limits.  To keep the re-build time reasonable during the design phase I have needed to keep the main array of punch cavities suppressed in my model.

This new design is giving me a great opportunity to start the development of a library of our own models that are representative of the work we do at Auer as a reference set for comparing our existing computer hardware in engineering against the newer computer hardware that has been introduced in the last year.

The dataset I plan to develop will make it easier for us to justify to management the need to upgrade our computer hardware with the new, speedier hardware platforms that have been introduced or are scheduled for launch in the later part of 2007.

I am taking each of these models and keeping track of the time that is required to do a control-Q re-build of each model.  Verification on Rebuild is turned off during this re-build process.  I am tracking the re-build time reported by Feature Statistics in SolidWorks and recording the time for the various computer platforms that I have access to.

So far I have recorded re-build times on the four different systems that I regularly do design work on.  Two of the systems are what we currently have at Auer and the other two are my systems at home.

For the punch holder model I have pictured above I have recorded the following re-build times:

• AMD Opteron 144 - 290 seconds
• Intel Pentium 4 640 - 264 seconds
• Intel Pentium D940 - 224 seconds
• Intel Core 2 Duo E6600 - 129 seconds

The new Core 2 Duo is very impressive, it is in my home Dell Precision 390 workstation.  It re-builds the punch holder model in about 58% of the time required on my dual-core Dell Precision 380 at work.

I am very interested in seeing the re-build times on the higher end Core 2 Duo's and AMD CPU's.  I am making the spreadsheet and punch holder model available to anyone who may want to test their systems. It can be downloaded from the following link:

Download Punch Holder Model and Processor Benchmarks (6.5Megs)

If you do use the model for some benchmarking with your systems please drop me a note with your results and I will add your data to the spreadsheet.

There are some exciting new platforms for both Intel and AMD that are scheduled to be launched towards the end of 2007.  I am hopeful that these new platforms will show even more performance gains then what I am seeing with the first generation Core 2 Duo's on the typical models we create.  If the performance improvements are significant it will make a very compelling case for us to add hardware upgrades to our engineering budget for 2008.

Comments

[matt]

Anna,

Interesting part. I wish I had more time and patience to work with it.

I CtrlQed it on 2 machines:

- Hypersonic laptop, AMD 4800+ (dual core), 2 GB RAM, nVidia GoFX1400 = 205 seconds.

- Boxx desktop, AMD FX57 (single core), 3 GB RAM, nVidia FX3450 = 245 seconds

It is interesting to me that the dual core made pretty good use of the second core, with the cpu usage averaging between 67-72%. This is probably why the laptop got better results than the desktop, which is supposed to be superior all around. Both of these systems are 1-2 years old.

[Anna Wood]

Matt,

Thanks for the results on your systems; I have added them to the spreadsheet.

If you have a moment can you send me a screen shot of your Features Statistics? I will re-run on my Dell's and save a screen shot of the Feature Statistics as well.

I am assuming that the dual cores are helping with the linear patterns. If the patterns take advantage of multi-threading I would think we would notice that the biggest percentage of improvement is in the patterns on the dual cores over the single cores.

I need to keep an eye out for some models that take time to resolve but are not so pattern intensive. That would give a more balanced perspective on system performance. For us though patterns are the big hitters when it comes to system slowdowns.

Thanks,

Anna

[Steve O]

First time I've read your bog and came across this Punch Holder benchmark. I thought I'd give it a whirl.

HP nw9440 - Intel Core 2 Duo T7400 2.16 GHz - NVIDIA Quadro FX1500M

2007 SP3.1 - 225 seconds for CTRL Q

Quick question - why wasn't the LPattern5 created with Geometry Pattern turned on. I know it affects the way the previous pattern is handled, but the results are much better. In 2007, my rebuild time dropped to 95 seconds. Maybe you just wanted to leave it as is for testing - which is cool. I was just curious since you had all the other patterns with this setting on.

Steve O

[Anna Wood]

Steve,

Thanks for your results, I have added them to the spreadsheet. Compared to the other Core 2 Duo laptops that are listed your machine seems to be a bit off the pace. How much RAM is in you system?

Good observation on the Geometry Pattern on LPattern5. My real model that is in my tool design has the LPattern4 suppressed and Geometry Pattern turned ON for LPattern5. This does bring the rebuild time down considerably.

For testing purposes I left the model in the worst case scenario for the Crtl Q rebuild. Keeps the results consistent for the spreadsheet and is a great tool to show how modeling technique can greatly affect the rebuild speeds of our models.

Thanks,

Anna

[Steve O]

Anna - 2gb of RAM. I probably got the RAM that was sitting in the back of the class when learning about speed.

I had a ton of crap running in the background when I first ran your test. I tried it again today with everything off, and Windows Visual Effects turned off, etc. That didn't make a difference, though. I was still around 225 seconds.

Question - When do you stop time on the benchmark? When you can physically click on something in SolidWorks, or when the status bar is clear of the Rebuild info?

Thanks,

Steve O

[Anna Wood]

Steve,

We are looking for the time reported in SolidWorks under Tools > Feature Statistics. I suspect you will find that your time reported there is in line with the other Core 2 Duo laptops.

Thanks,

Anna

[Steve O]

Boy do I feel dumb. I did a CNTR+Q and sat and watched the clock waiting and waiting . . .

In that case - 212 seconds.

Steve O

[Steve O]

I forgot to put it this in my previous comments. 2008 Beta 3 gave me 201 seconds.

Steve O

[Kirk L. Jess]

Anna,

I clocked out at 182.63 seconds
Solidworks 7.0 x64 Edition Sp4

Boxx system
Windows XP 64
AMD Athlon 64 FX-60 Dual Core 2.73 Ghz
4 GB Ram
Dual Western Digital Raptor Drives Raid 0
Nvidia Quadro FX 3500
I have the screen shot saved if you need it.

Kirk

[Dennis Messer]

I have some interesting results... for this test Core Quad is slower the Core Duo.

I have two nearly identical Dell Precision 390s that I bought at the same time, the only diffrence is one is a Core Quad and one is a Core Duo

Both systems have these stats:
RAM 2gig
Windows XP -32bit
Nvidia Quadro FX 550
Same Dell Motherboard
2.4ghz CPUs in both. One is a Core Duo and one is a Quad.

Core Duo 2.4ghz E6600 129 seconds
Core Quad 2.4ghz Q6600 136 seconds

I ran the test several times and got the same results.

I thought to myself, hmmm... Why is the Quad slower? What if I set the processor affinty on the solidworks process to only use two of the four cores, I did just that and re-ran the test and it came in at 129 just like the Duo machine.

Looks like maybe Solidworks is optimized around two cpus. I am going to test that theroy with the solidworks benchmarks from SPEC.ORG and once I figure it out I'll post the results.

--Dennis

[matt]

Anna,

I got a new computer, and decided to add a very interesting data point to your spreadsheet.

The machine is a HP tc4400 12" convertible tablet PC. It has a mobile Core 2 Duo T7200 processor, with 3 GB RAM, and mobo integrated Intel video.

The benchmark speed for your punch holder part was 144 seconds. This is 25% faster than either of my other machines. I bought this thing for $1100 as an occasional use and travel machine, but even with the crap graphics, I may rethink that.