"Cheap" new CFD software
The January issue of Racecar Engineering has an article about an interesting new CFD application called Caedium. It utilizes a GUI and a Navier-Stokes solver with various turbulence models. Apparently it can also run on a Windows platform, with a 64-bit version to be released soon. License costs are $300/quarter or $1000/year, which might fit within the budget of at least some small formula car development projects.
CFD
Just be aware this application is rather "challenging" until you get used to it. It is NOT intuitive or easy to use.
Be prepared to invest a lot of time to understand/use it.
Tom
Tom Owen
Owner - Browns Lane and Racelaminates.com
OpenFoam is free. That said, I think it is pretty difficult to use.
Cory
The solver it's using is basically their adaptation of openFOAM if I recall correctly. Which means you need to know what you're doing to use it! It could be good for the racing community because for those who know CFD properly there is now a cheaper option, but then again as a CFD guy I'm already able to use openFOAM so my software costs were already zero!Originally Posted by Rick Ross
My guess is that it basically does some of the tedious work commercial codes do for you. For example, in a commercial code, if I want something to be a wall I just go define > BC, click the zone and highlight wall, or its already intelligently assumed which surfaces are walls. It then assumes standard, common wall properties. In openFOAM I have to spell all that out separately for each zone in several text files using C code, and set a bunch of other things as well in other files that relate to that wall. This Caedium probably fills all those files out for you, but you still need the CFD expertise to know what to tell it to do. And the big question is meshing. Yes, it has a great solver, but 90% of the accuracy comes from the mesh, so who's the "brains" behind the mesh? I've never seen a good automatic mesher except for ones that I or my colleagues have written that are too project-specific to ever be sold or generalized.
Also, without lots of memory and a 64 bit version you could never do proper half car (symmetry plane) runs. Oh and forget windows, it will be slooooowwww!
Just my unsolicited 2 cents though. Approach with caution :-)
-Robert
I wonder what the memory and processor requirements would be to do a half-car run on a Windows platform (assuming it is even possible)? According to the magazine, the 64 bit version of the software is not yet available.
Well, with 32bit you have a max of 4GB, which with commercial codes means about 3.5M cells max. With FOAM its less, and foam's mesher is the real memory hog. IMO that's enough for upper rear wing development only, meaning rear wing in isolation with no car modeled at all, meaning you'll miss the wing-diffuser interaction.
Full car, if you're a brilliant mesher with the best commercial tools, is about 10M cells *minimum*, or 16GB RAM. With FOAM's mesher you better have 20+ GB, and ideally you'd use more cells. You could run on any processor, its just how long you'll wait . . .If you want to kick out more than one run a day you need a small cluster.
-Robert
Process memory is restricted to 3GB in windows on a 32bit OS, with tweaks. Of note for Rick is that you can run 32bit apps on a 64bit OS and get the memory benefits (8TB of addressable space) but still slowish as the OS will kick the CPU back down to a hobbled 32bits for that process. This adds a lot of context switching at the hardware level, which tends to degrade overall performance, but it's doable.Well, with 32bit you have a max of 4GB, which with commercial codes means about 3.5M cells max. With FOAM its less, and foam's mesher is the real memory hog. IMO that's enough for upper rear wing development only, meaning rear wing in isolation with no car modeled at all, meaning you'll miss the wing-diffuser interaction.
Full car, if you're a brilliant mesher with the best commercial tools, is about 10M cells *minimum*, or 16GB RAM. With FOAM's mesher you better have 20+ GB, and ideally you'd use more cells. You could run on any processor, its just how long you'll wait . . .
The caveat of course is that just because the OS will allow extended memory access doesn't preclude the application from having been architected by a moron who took a literal / linear approach to memory management. There are many applications like this which are essentially blind to the extra 64bit memory space because they programmed it to only expect what they would get from a 32bit OS. I do not know which side of the fence FOAM sits on in this regard.
Clusters have their own issues.
Cheers,
Rennie
Hmmm.......so if I ever want to try out this software I may need a "slight" upgrade from my Dell laptop with 4GB of RAM. At least I have the 64 bit OS.
Interesting stuff, how bad would you guess the performance loss is running a 32 bit app on a 64-bit OS vs. the 64 bit app?
Well, you can try it out on your laptop, but if you like it and want to use it for meaningful motorsport work you'll need a significant hardware upgrade!Hmmm.......so if I ever want to try out this software I may need a "slight" upgrade from my Dell laptop with 4GB of RAM. At least I have the 64 bit OS.
-Robert
You could try! Be prepared to limit your model field, and wait a looooong time for the calculation to converge, but yeah it would probably chug through it eventually. Your laptop will be useless for other things during that period, of course.Hmmm.......so if I ever want to try out this software I may need a "slight" upgrade from my Dell laptop with 4GB of RAM. At least I have the 64 bit OS.
Luckily enough, good multi-core server hardware is becoming cheaper and cheaper by the minute. You can probably find a decent quad/quad machine on eBay with memory and hard drives for about the price of a good laptop these days. Intel is making 16-core chips these days on new machines and keeps promising an 80-core mesh chip later in the year. Moore's Law and all, but all of this talk today about "unobtainium" hardware requirements to process CFD on an adequate scale will become chump change in a year or two.
Depends on how many other background processes are running, how active they are in terms of interrupting the CPU, and a few other esoteric factors like how the software itself is architected, how many CPU cores are available, etc., but you can figure on a 30-60% performance hit. The basic problem though, is context switching. In other words, the CPU has a big hopper of instruction sets that it processes one by one, and it expects all of those instruction sets to be in a certain format (32bit or 64bit). The formats can't be mixed within the hopper. So when a 32bit application comes along with a request, any new 64bit instructions are halted, the CPU empties out what is currently in the hopper, and the 32bit instructions are now fed into the hopper for processing. And vice versa to go back to 64bit.
It's a bit like waiting at one of those stupid road construction checkpoints where they are only allowing one lane of traffic, so each end of the construction zone backs up while they let a whole train of cars through going one way, then the other.
We have a TILE-Gx machine here at the office with 2TB of memory on board. 100 cores using mesh architecture for CPU traffic instead of a bus. I'm constantly tempted to see how well it would run Fluent...
Cheers,
Rennie
Caedium Speaks!
Full disclosure: I'm the developer of Caedium from Symscape, so I'm severely biased when it comes to my software
When the Racecar Engineering article was written we didn't have the 64-bit version up and running, but we do now and parallel too - so it can take advantage of those multi-processors/cores. For more try:
http://www.symscape.com/news/faster-...-goes-parallel
The generic open wheel simulation shown in the article was about 1.2 Million cells (half model), which ran on a 32-bit Windows machine with 4 GB RAM and took a few hours. The same simulation using a 4 processor (8 threads) gaming laptop now takes about 30 minutes.
In terms of the mesh count it all depends on where you cluster your cells. Caedium's mesh (Accuracy) tools provide fine tuning of mesh size and growth rate to keep your cell count down, but don't take my word for it...
The easiest way to see whether Caedium is intuitive or easy-to-use or is slow on Windows (which it isn't) is to pop over to the Symscape website and take a free trial for a spin:
http://www.symscape.com/product/professional
Some fluid mechanics background is definitely a help, even better is some prior CFD experience, but you don't need to know anything about OpenFOAM.
Hi Rich,
Thanks for your comments. Could you tell me a little more about the "gaming" laptop on which you ran the 30 minute half-car simulation? Were you using the 64-bit Windows 7 OS? How much RAM? Quad-core processors?
I am not a gamer and am not familiar with the typical capabilities of gaming PC's.
Gaming Laptop Specs
[FONT=Verdana]It's an ASUS G51J and yes it runs 64-bit Windows 7, with:
[/FONT]
- [FONT=Verdana]Intel Core i7 (1.6GHz Quad-Core)[/FONT]
- [FONT=Verdana]6 GB RAM[/FONT]
- [FONT=Verdana]Nvidia GEFORCE GTS 360M 1 GB[/FONT]
- [FONT=Verdana]500 GB HDD
[/FONT]
I've found that good gaming machines are usually quite acceptable for desk-bound CFD. Added bonus is the relatively high-spec consumer graphics cards too.
Thanks, Rich. My current machne has a 2.2 GHz Intel Core 2 Duo processor, and 4 MB of RAM. I am running 64-bit Windows 7.
How sensitive is the application run time to RAM availability? In other words, if you received a solution in 30 minutes with 6 MB, what change (if any) would you expect with 8 MB installed?
CFD speedup
As long as the simulation fits within your RAM, you won't see any difference if you increased your RAM. However, if the simulation is too big for your RAM and is therefore paging then you'd see a significant speedup with more RAM - maybe by an order of magnitude.
For a given simulation, say of greater than 1 million cells, the best way to see a speedup is to add more processors/cores. You can expect such a simulation would be just less than twice as fast if you doubled the number of processors/cores from 2 to 4.
Thanks. I saw that some of the quad core machines come with 6 MB of RAM, but can be upgraded to 8 MB. Just wondering if there was any advantage to the RAM increase for a given quad core processor.
Just a note here to mention that the paging topology restrictions change dramatically when using SSD's. In fact we often see applications which experience better performance while paging from an SSD than straight from RAM, though granted it's highly dependent on memory speed and the bus speed comparison to the HD controller. But when you've got an SSD which can retrieve paged memory faster than the RAM chip, it does create an interesting conundrum.
Might be something for you to test.
Rick, it's a bit of a tuning question. There's no hard and fast rule when it comes to performance tuning a computer, it depends on how loaded each aspect is (memory vs CPU utilization). It's a bit like a race car really - optimum lap times occur when all is in harmony, rather than simply piling on more of a good thing. Even for something as benign as Microsoft Exchange (their email server platform), they will provide a recommendation of between 1 and 4GB of memory per CPU core, depending on the role of the server. Maximum throughput typically is achieved at around 80% resource utilization across the board - more than that and you are flirting with overheating issues. In general, the software manufacturer should be able to give you their performance baseline configurations and findings.
Cheers,
Rennie
Thanks, Rennie. I have never used any applications that even approached the limits of my PC, so this is a steep learning curve for me. I am considering the purchase of a new laptop this year, and am trying to figure out if there might be something out there that would let me at least experiment with an application like Caedium. I have some basic fluid dynamics knowledge, but have never had the opportunity to experiment with CFD (when I started college they were still using punch cards!).
Looks like you can get a laptop with a quad-core i7 processor and 6 MB of RAM for around $1500, depending on options. There also appear to be a few six core laptops on the market.
I would caution against trying this sort of work on a laptop of any sort. Manufactures do not usually design them for the thermal output you get when running the proc 100% flat out as this sort of math would. You may be able to get one that can put up with it but that is going to come at 3k price tag for a boutique gaming laptop.
As an aside has anyone with experience in this seen any products written to take advantage of the current video cards capability to do really just this sort of math work? Though maybe memory is a limitation for the video path since you only have small amounts of on board memory in your average card?
I agree that a desktop would be better for this application. However, I travel a lot and a laptop just makes more sense for me. I am not looking for a dedicated "CFD" computer, just something that would perhaps allow me to experiment with Caedium (or similar) and see how it goes. I could always upgrade machines in the future in the unlikely event that I ever started using these types of applications on a regular basis.
If you are wanting to experiment then just install Caedium on whatever you have available with the understanding that it will be compromised when compared to a dedicated machine. If you are looking at new machines anyway then go for 6-8GB RAM and a 64 bit OS. Leave the 6 core processors and monster video cards for the gamers.
Besides, a CAD laptop is a bit of an oxymoron. I have one but it's weight, heat, and batterylife means it sits on a desk plugged in 99% of the time. Not very portable, but it will keep your coffee warm if you set it next to the CPU/GPU exhaust vent.
Are there any plans to add GPU processing support? Mind you, just knowing that the software was developed to take advantage of multi core processors is a huge plus.
-Kyle
GPU Acceleration
There is GPU support already in the form of accelerated 3D graphics, but I presume you mean using the GPU to accelerate the CFD simulations. This is something I'm investigating for the next Caedium release, so watch this space.
Like you said earlier, I think it's best to test the water with whatever computer you have available, then worry about hardware specs later. If the software doesn't meet your needs then hardware specs are mute.
I'll vouch for that. Up until last year I was traveling to shipyards all over the world and had to be able to take a laptop with me that would easily run Solidworks, Autocad and other programs. Some of the drawings and 3D models would grind a normal laptop to a stand still. I have a Dell M6400 Covet with a 2.54 GHz Quad Core processor, 8 GB ram, Nvidia Quadro FX 3700 1 GB video card and Win 7 64 bit. This thing will suck a battery down in under 20 minutes doing moderately hard work. Running a hard core processing program would kill it even quicker. You also need to keep a chiropractor on speed dial because this thing weights over twice many "normal" laptops. That being said, it is very impressive to have that much processing power packed into a portable package. Price is also killer at around $5K for a nicely equipped one. To my knowledge only Dell and HP even make a "portable workstation" class laptop.
Chris Ross
09 NovaKBS F600 #36 Powered by '09 600 Suzuki GSX-R
"If all else fails, immortality can always be assured by spectacular error." John Kenneth Galbraith
I have no doubt that you guys are 100% correct regarding weight, power consumption, and cooling. And I know that this would not be considered a "workstation" class laptop, but the following HP Pavillion can be had with quad-core i7 and 8 MB RAM for under $2K, depending upon configuration.
http://www.shopping.hp.com/webapp/sh.../dv7tse_series
I think Sony also sells a Vaio with similar capabilities.
Aside from running parts in isolation you can't do proper CFD on a laptop. If you wanted to accurately do a Formula car you'd need at least 16GB RAM. Now if you just want answers and graphs, you can run a laptop, but good answers will cost you.
-Robert
Thanks, Robert. If I understand you correctly, total cell count is essentially limited by RAM availability. In a previous post you also mentioned 10 million cells for a formula car, requiring 16 MB RAM. Is this for a full or half car simulation? Have you found that half-car simulations of open-wheel formula cars provide useful data?
Yes, total cell count is a RAM issue, not CPU limited. The 10M cell figure is for a half car. Realistically unless Caedium supports hexahedral meshing you'll need even more (you need many more tetrahedral cells to resolve the wake and vortices than you would with hexahedrals). That said I am talking half car models, which will give you perfectly good data at half the cell count (i.e., you would need 20M for a full car model). Full car is rarely necessary.
-Robert
Hold out for a Sandy Bridge based PC if you are going with an Intel chip. There should be several released in March. The model number of the processor will be i7-2740QM instead of i7-740QM. Four digit model number starting with a 2 is the key (ix-2xxx).
Sandy Bridge is the code name for the "Second Generation Intel Core Processor". The improved architecture offers a rather large improvement over the first generation chips. So, next month you should be able to get better performance for the same money, or the above laptop marked down a fair bit.
Cheers,
-Kyle
Thanks, Kyle. You are the second person to recommend that I wait a few months for the next generation processors.
It seems that the primary limitation with a laptop for this type of computation-intensive application is RAM. Even an 8 GB laptop doesn't have adequate memory for the number of cells required for complex shapes such as that required for an accurate half-car simulation. This is not so much an issue for desktops, of course. I did a quick check of the HP website and found that I could get a six core desktop with 24 GB of RAM for around $2000. A used machine on Ebay would probably be much cheaper.
Wonder if I could log into and use a home computer remotely while I was on the road? I've done this to other machines using the Windows Remote Desktop application, but I've never checked to see if I could log into my own wireless network from the internet. If so I could perhaps travel with a cheap, lightweight laptop and use it to access a desktop at home.
That is certainly possible. I do it all the time. The easiest method is to use a web proxy like gotomypc.com.
Using windows remote desktop will take a bit more understanding of configuring port forwarding on your home router, and also using dyndns so you don't have to worry about your isp changing you ip address while your on the road.
The only issue would be if these programs are rendering 3d displays while your doing inputs for the program. Most remote desktop programs cannot transmit directx or open GL redered things to a remote PC.
Right, I was asking about GUGPU computing capability using something like OpenCL, or a specific Nvidia CUDA/ATI Stream solution. Also, I am curious if Caedium has the capability to distribute work over a beowulf cluster.
Thanks,
-Kyle
Windows HPC Server 2008 Cluster
I've never tried it, as I don't have access to a Beowulf Cluster. However, Caedium works well on a Windows HPC Server 2008 cluster. For details try:
http://www.symscape.com/node/930
re: Intel Sandy Bridge chipset SATA flaw
This is 1 1/2 months old news, but I just ran across this 1/31/2011 CNET article, and I feel it bears repeating. It's impact will likely delay Sandy Bridge system availability into this April.
Rick
So, let's assume for a moment that you get the huge processor and huge amounts of memory, and you buy the software and load it all up......
How do you get the car into the software? Do you have to have it scanned somehow? Is there software that can look at photos and make a model? How does that part work?
I have a bunch of CAD drawings of the car, but it is all frame. The bodywork is not in any of the drawings.
Body into CFD
You will need to do one of 2 things
1.) Laser scan the body and generate an STL file. Then convert the stl file into a CFD compatible file with a program like ANSA
2.) Generate proper 3D cad surfaces or solids of bodywork (from drawings, scanning, black magic-whatever). Then use these surfaces in the CFD package.
Whatever way you go you are looking at a lot of work.
There is no such thing as "low cost" CFD right now.
There is cheap CFD that doesn't work-there is expensive CFD that doesn't work.
F1 teams use it but still require any CFD results to be verified by real tunnel testing.
There has been real spotty results of "CFD only" designed racecars. Not a good track record right now.
But things will change.
All Tools have Strengths and Weaknesses
If only it were so easy, then we could discount CFD and move on...
It's not a one way street. CFD is also used to validate wind tunnel results. They form a much more potent combination together than either does alone. Also CFD is great for closed loop optimization. Different tools have different strengths.
I presume you are referring to Virgin F1 and their CFD-only approach?
I think you have to decide whether CFD-only is a symptom or a cause of their poor performance. I believe it's a symptom (of course I'm likely biased too). Their problem seems to be under funding. In F1 money = results, with few exceptions.
Things changed a long time ago. F1 has made extensive use of CFD since 1996, increasing every year. The aerospace industry has been using CFD in various forms since the 1960s.
CFD's strengths are in comparison studies, not absolute numbers. Wind tunnels can be very accurate, but typically for the wrong Reynold's number due to scaling issues.
My company does a lot of CFD in the motorsports industry. A lot. We are quite aware of all you mention. Aerospace use is way different animal. In some ways easier--I have done a bunch of aero CFD (No ground, no rotating wheels,).
yea-Virgin, Aston Martin and others. Issue is putting all of your eggs in the CFD basket. I am not anti CFD-far from it. My issue is that there is a growing (and wrong) assumption that CFD will give the "typical" racer a quick and easy and cheap way of getting aero data. It just ain't so. (Yet)-give it 10 years to mature but even then it still will not be a "press a button" and get a car design.
I just got done doing a very extensive CFD model versus full scale moving ground plane tunnel test on the same ground effects/tunnel car. Probably first time done on this continent-and still more than many F1 teams do. Expensive-yes-but the results (not for general disclosure-sorry) are very interesting and show a lot of things you do not get with just CFD. A lot of things that most CFD engineers and software are not aware of (as they usually have no or little actual wind tunnel experience with all the types of cars that are out there.) navier-stokes has some fundamental issues with cars running on the ground. And we have the proof now.
CFD + Wind Tunnel = Complimentary Relationship
Seeing as though it's down with CFD I guess I'll have to argue its case, but I'm still of the opinion that:
CFD + Wind Tunnel = Complimentary Relationship
I'd stick with different rather than easier. There's plenty of ground effects in aerospace, e.g., landing/takeoff, VTOL. Throw in compressible flow with shock waves, stores release from internal bays and you can get difficult aplenty.
The fundamental problem is that the alternatives - on track testing or wind tunnels are extremely expensive, so they don't happen for 'typical' racers. In this light CFD is attractive. You don't need perfect matches with experiments to act on trends.
Again exact matches aren't CFD's strength, as you seem to be aware. As long as the trends are in the right direction you can nudge your design in the right direction. Then for those lucky few who have access to track time or wind tunnels you can always cross-check and refine.
I presume you mean Reynolds-Averages Navier-Stokes (RANS) when you say 'navier-stokes.' Given immense computing resources (huge meshes) for Large Eddy Simulations (LES) I'm sure that accurate CFD aero results for a car are possible.
How do you validate your wind tunnel results? E.g., wall corrections for blockage, force balance mounts (stings), boundary layer trips for scale models. I ask because it seems that everyone is comfortable assuming the wind tunnel is the exact result, yet mistakes happen, assumptions aren't cited, etc.
There are currently 1 users browsing this thread. (0 members and 1 guests)
Posting Permissions
Reply With Quote