|
[Sponsors] |
June 8, 2017, 10:36 |
On the CFD market and trends
|
#1 |
Senior Member
|
Inspired by a series of blog posts by Keith Hanna (https://blogs.mentor.com/khanna/blog...-mind-the-gap/), I'd like to try to put together some toughts about the current status and trends in the CFD market and, if possible, on applied CFD in general.
You will find more precise numbers and details in the blog cited above and its references but, roughly speaking, we are today in a market with 300k-500k paying users with revenues around 1.2 Billion $ and, apparently, this number increases at 10% yearly, but is somehow slowing down. The main theme of the post series is the "democratization" of CFD, making CFD more and more available to uneducated and casual users, which is the biggest trend today in the market. That is, we can probably sell products to 10x more users, why on earth are we still not doing it? Some of the factors which have been globally pursued to help in this process are (this is a personally reinterpreted list): - Ease of use trough CAD-ification, APP-ification, embedding in PLM tools - Leverage of GPU power (so that you don't need anymore a cluster) - Cloud (so that you can still use a cluster easily even if you don't have one) - (Supposed) mesh-free techologies and the like (Lattice-Boltzmann, Immersed Boundaries, etc.) - OpenSource tools Now, before delving into some of these points, I'd first like to highlight the flaws I see in this picture. First of all, none of the above directions consider the possibility that, maybe, the underlying tools might not be mature enough for this. My current experience is that there are still a lot of cases where the solver (both commercial and opensource), even if carefully tuned, still miserably fails and I need to develop a solution strategy which, nonetheless, doesn't get me to full convergence (2nd order being completely out of reach). Just try the 5 shock tube tests on the Toro book and see what happens... now imagine how that would work in a real case. APP-ification might be of help on this (but I'd like to hear from those using these vertical apps if they really find themselves always away from problems), but to which extent? I mean, I can always come up with boundary conditions making an appified case difficult or impossible to converge with the underlying methods. And for me, democratic also means safe fail, or fail by degradation. Case blow up with correct case setup is not part of the picture. The second aspect that is obviously missing from the picture is the price. I mean, order 10k $ is high for casual users or small companies providing CFD as a service. And it still is one of the main choices for OpenSource codes over commercial ones (I have first hand experience that most people using it have no idea how a certain feature is implemented or where it comes from, so the story of looking into the code is pure bullshit, even in academia). Let's face it, almost nobody would use a pc or smartphone if the price tag was, respectively, 10k $ and 2k $. And we can afford them at much lower prices not because of magic, but because of low production costs in developing countries (just imagine producing iPhones in Switzerland). Also, most people don't pay, say, for LinkedIn, which instead has revenues from a bunch of premium users using niche features (still, would be interesting to know who and why is sustaining the present startup economy with unprecedented cash flows, is this just the effect of low interest rates? Maybe this is going to be the next bubble no one sees on the radar). Now, a CFD code is nothing similar to the two products above but, somehow, shares a lot with their union. And if there was, say, an usable product, at the video-game level of usability, at 15 $ per month, things would be different. My idea is that the price can't go down enough because the CFD industry is flooded by costs, and it is so because, as any other business related software, users can't read a damn manual (or do not want to), and when they do, it might be poorly written, so you need to give a lot of support. Honestly, I don't think this is going to be any better if CFD is going to be used by even less educated people than today. This somehow proves that CFD is still far from being democratic and, probably, not even scalable as a business. Note that the video-game example is quite interesting. That industry regularly produces technical software (it has a complexity comparable with CFD ones) easily reaching 10Mln users with similar prices. And those users don't need to be marketed (not one by one at least) or supported. That might really be the key to democratization, to the extent that the video-game industry might actually reach it before the CFD one. However, one should not forget the main topic: a tool supposed to solve a mathematical problem representative of a physical problem, which brings me to a conclusion. Making CFD democratic trough ease of use might never meet its goal just as much facebook would never had its fortune if users were required to pay and because there is a minimum amount of knowledge which, in order to use it, simply can't be dropped. To keep the size of this post manageable, I'll leave why and how, in my opinion, the price and not the other factors alone can make CFD democratic. In the meanwhile, what do you think on this? Is there something which would enlarge the CFD reach? Are the current efforts really making CFD more democratic? Is the prize an obstacle at all? |
|
June 8, 2017, 11:10 |
|
#2 |
Senior Member
|
It is probably relevant to add that, in contrast to the points above, the academic research is clearly going in quite the opposite direction now, with uncertainty quantification and high order methods being the most relevant research topics in the field.
For what concerns GPU computing, which also got its share of visibility in research, my guess is that it was mostly industry piloted and I still have to see a clear, fair comparison between a classical mpi code and an identical hybrid one with GPU running on the exact same problem at the exact same energy cost. What would be the time gain in that case? Who knows... |
|
June 8, 2017, 11:42 |
|
#3 |
Senior Member
Filippo Maria Denaro
Join Date: Jul 2010
Posts: 6,839
Rep Power: 73 |
this democratization seems like a videogame advs
|
|
June 8, 2017, 12:11 |
|
#4 |
Senior Member
|
||
June 8, 2017, 12:15 |
|
#5 | |
Senior Member
Filippo Maria Denaro
Join Date: Jul 2010
Posts: 6,839
Rep Power: 73 |
Quote:
iVenturi ... |
||
June 9, 2017, 01:06 |
|
#6 |
Senior Member
Arjun
Join Date: Mar 2009
Location: Nurenberg, Germany
Posts: 1,283
Rep Power: 34 |
The perception that until you are not paying 10k or 20k in licenses you are not doing real CFD does run deep. Making licenses cheap does not work for this reason as a company because then no one buys that product.
Open source code is different but hard to sustain. Open FOAM has been successful but lots of things come in place for it, other softwares may not be so lucky. One example of this thing for openfoam, say is that from the time it becomes open source , it has multiple forums created for it on cfd-online. I remember no other software for which cfd-online was so quick to make forums. They usually wait until there are enough users. |
|
June 11, 2017, 13:04 |
|
#7 | ||
Senior Member
Lucky
Join Date: Apr 2011
Location: Orlando, FL USA
Posts: 5,740
Rep Power: 66 |
We are entering an age where everyone has some level of access to solving ode's and pde's on a computer rather than analytically by brain & hand. So your ability to get a solution is now no longer dependent on your intellectual prowess in manipulating the physics. Even if you don't completely understand the physics, you can experience the result.
Quote:
I see the three main barriers to entry are licensing cost (say $10k) hardware cost ($10k for mid/high-end workstation) and know-how. Many vendors will waive the software cost for you if you are a small company, and they offer help with getting you cloud computing. But as soon as you are able to do CFD well enough, you become a big company and now the costs become real again. These are just the barriers, then there are the nuances that make CFD in real PITA like always crashing, needing good mesh, etc. Quote:
GPU power. This was the supposed hail mary around 5 years ago thanks to CUDA and many asked me for their opinions on it and even many more tried to convince me it was how all CFD (and matlab) would be done next year. However, a quick search led me to discover that GPU's simply don't offer any benefit in terms of dp units for their cost (at best they scale linearly with CPU count). GPU computing could probably be viable in our lifetime but certainly not soon. It will take a major architecture overhaul and commitment from GPU and CPU manufacturers with high dp core unit counts (in consumer grade cards) to make this a reality. The way it is done now in my opinion is complete bullshit. The feature is only viable for the super high-end market right now which is anti-democratization. Cloud computing is an interesting concept. To me it is a vector but certainly not a solution. At the end of the day, there is something called IP which large companies are extremely interested in protecting and they will not easily trust the security of this sensitive information to an open cloud. Cybersecurity is also a growing concern. However, for a CFD hobbyist (if those exist) who is not trying to make money cloud computing is certainly an enabler for serious calculations. Mesh-free technologies has been poorly mis-advertised. These really are alternative strategies for solving PDE's apart from FVM and FEM. They have their advantages, but also have their disadvantages. In terms of democratization, I think they bring nothing to the table and likely cause more problems than they solve, i.e. they typically have even poorer documentation than popular tools. OpenSource tools mitigate one of the barriers-to-entry but it's still never free. The source code might be free to use, but service will always cost you something (as it should). I like the analogy to the video game industry. For democratization to work, it must be robust like a video game as you said. The user doesn't need to deal with all the nuances, as long as they can play with it; they're even willing to pay a reasonably small price. Video game makers also don't post their source codes online. |
|||
June 12, 2017, 11:25 |
|
#8 |
Senior Member
Hector Redal
Join Date: Aug 2010
Location: Madrid, Spain
Posts: 243
Rep Power: 17 |
Hi,
First of all, I would like to say that I am only a beginner in the CFD field (started in 2007), so if many of what stated in here by me sounds wierd or non sense, I would like to request apologies for it, :-). In my opinion, the CFD software is still far away to the state where it has reached some kind of "democratization" state. One important aspect in the CFD field is the analysis of the results. You can have the best CFD software ever produced at a cheap prize, but it won't help you unless you know how to analyze the results obtained with it or unless you define the correct Boundary and Temporal conditions, so as to make it work. You cannot forget the engineer part in this process. I think that there is still a big gap filled by engineers in this process, that cannot be substituted by software alone. I don't think that the current state of the CFD Market provides users with the background obtained after having studied four/five years (or even more) in a University in order to obtain a degree in Mechanical / Mathematical / Aerospace Engineering. Maybe, one day, when Artificial Intelligence (AI) has developed enough, this kind of software may allow to solve any engineering problem without human intervention, or little intervention, to say the least. I don't think that this is the moment in time when we have reached it. AI is trying to solve similar problems in other fields, or they are facing same problems like the pointed in here. Maybe when some algorithms / schemes in other discipliness are found, similar solutions can be adopted to fix some problems in the CFD (such as skewness, etc.). I recall on this point, Experts Systems, that try to substitute human intervention in some fields. Is it not the same as what we are discussing here? I see also two distincts set of costs in this regards: -Software costs -Hardware costs Software costs can be divided by the amount of software licenses you sell. The higher number of licenses sold the better. If you sell a lot of licenses your flat costs can be divided between those licenses. And this more or less matches the Video Game industry, that sells software as well. But, I think there is an important point in the equation (Software + Hardware), that is related to hardware costs, Scale economy costs can be used, but i don't think that they can be reached up to a point where you can find a super-computer at a very low price. Top high end computers needed for CFD will be expensive. Just only commenting aloud my dwelling thouhts on this discussion. I will be glad to receive any comment or open criticism to my statements. Best regards, |
|
June 12, 2017, 11:43 |
|
#9 | |
Senior Member
Lucky
Join Date: Apr 2011
Location: Orlando, FL USA
Posts: 5,740
Rep Power: 66 |
Democratization doesn't mean that everyone will be doing DNS and running on supercomputers. Democratization means more that everybody (and their grandmothers) will have some reasonable access to CFD. Today, the typical person has 3 PC's (1 at work, 1 at home, and 1 smaller portable device such as a laptop,tablet, or phone). And all of these are capable of running simple CFD problems. Licensing costs can be made way cheaper to eliminate it as a barrier to entry. You can easily imagine that it is sold by compute hours to completely level the playing field. Compute hours on supercomputers are sold for as little as cents per hour. For licensing to be competitive, it would need to be sold for a fraction of this cost.
Quote:
This forum is also a good example. Just take a look at how many people are doing multi-phase reacting flows with moving bodies as their first CFD case! It is insane how CFD newbies are tackling problems more complicated than the most complicated problem I've ever solved. |
||
June 12, 2017, 11:50 |
|
#10 | |
Senior Member
Filippo Maria Denaro
Join Date: Jul 2010
Posts: 6,839
Rep Power: 73 |
Quote:
This forum is also a good example. Just take a look at how many people are doing multi-phase reacting flows with moving bodies as their first CFD case! It is insane how CFD newbies are tackling problems more complicated than the most complicated problem I've ever solved.[/QUOTE] So, the question is: democratization of CFD means only democratization of the codes? |
||
June 12, 2017, 12:09 |
|
#11 | |
Senior Member
Arjun
Join Date: Mar 2009
Location: Nurenberg, Germany
Posts: 1,283
Rep Power: 34 |
Quote:
:-D :-D Indeed. |
||
June 13, 2017, 03:38 |
|
#12 | ||
Senior Member
Hector Redal
Join Date: Aug 2010
Location: Madrid, Spain
Posts: 243
Rep Power: 17 |
Quote:
But this only rules out one part of the equation (hardware). The problems related to software are still there. Quote:
In my humble opinion, only if you have the background needed, you will be able to understand the results and analyze them. If not, then I don't see reason in trying to carry out a job that you don't know if you are doing right or wrong. |
|||
June 13, 2017, 09:37 |
|
#13 | |
Senior Member
Lucky
Join Date: Apr 2011
Location: Orlando, FL USA
Posts: 5,740
Rep Power: 66 |
My opinion is that hardware and software licensing could be democratized. For hardware, everyone and their grandmother (sorry to overuse this Russian expression) has reasonable computing power today. For HPC, cloud or cloud-like computing will most likely finish the work. By cloud-like, I mean someone will make an app that will run my CFD on some supercomputer somewhere in the world, but since I don't have to interact with this distant computer, it is effectively like a cloud to me. Software licenses is (imaginably) easy to turn off. The economics of scale are in favor of both these issues, when they have a large number of users, they can make a lot of money (just like video game industry). Again, we are talking cents per compute hour, which sounds cheap until you rack up a million compute hours. =)
Quote:
To me, the obstacle is getting people comfortable enough with any software, with "how to do it," so that they can do it. E.g. how can we eliminate posts in those forum that asks: Are my results converged? How do I fix it? This sounds deceptively like the human not having sufficient background, which is probably also true, but the point of democratization is that we are talking specifically about the uneducated ones. So I must say I agree with the original point, is that it basically boils down to the robustness of the tools and how easy they are to use. I think we have answers for the other aspects, but getting the tools to be user friendly instead of user hostile is still a challenge. |
||
June 13, 2017, 10:59 |
|
#14 | |
Senior Member
|
Quote:
I personally like the definition of Keith Meintjes of CIMData (at ASSESS, 2016) who defined “democratization” with respect to computer aided engineering as: “Making simulation accessible and usable by a much wider audience in a way that supports how they get their work done.” I think that this definition is as good a starting point as any for the topic of CFD democratization especially in the sense of making the software available – and being used – by a much wider community (say x10 or even x100) than the current userbase of today. Thus, in practice, the term is mostly intended from a software vendor point of view. Which, basically, is: we want to sell more, how can we achieve this? The reason I highlighted the price and robustness points is simply that those are just kept out of the equation in the whole discussion. I think that a key aspect of the discussion is the part: in a way that supports how they get their work done. What this exactly means for software vendors is probably reflected in the products they try to sell today. However, I think, this inevitably involves taking CFD knowledge out of the picture. At least according to the vendors. Which brings to my mind the following analogy: it is like a F1 championship where you want to take pilots out of the picture and rely, say, on AIs or robots to drive the cars (...damn it, I just realized it could actually happen soon, considering the recent self-driving direction taken by the automotive industry). The idea for this discussion was twofold: 1) Is democratization actually good in general? 2) Is the current approach to it really going to give any results? On point 1 I don't feel like it is bad, in general, as idea. Actually, it might really be a gamechanger in the industry. If you look at how the products are packaged today (i.e., their GUI and workflows), with respect to, say, 15 years ago, they have simply moved the complexity from the knowledge part to the workflow part. I recently had a live online demo with one such products from a major vendor, and I had the following toughts (mostly in sequence): - congrats, you really managed to skyrocket the complexity of achieving even the most simple task - I'm not Einstein, but damn, I've been in CFD for 12 years, how comes I can't even find a clue where to set a boundary condition - I'm not launching a mesher whose only control is a slider ranging between coarse and fine - incredible, I was order of magnitude more productive in OpenFOAM - ok, this is where I leave you and this is where you realize how Kodak or Nokia managed to fail so embarassingly despite their monopoly. This is not democratization, it is just shifting focus to the largest possible target at a given moment, no matter who they are. And this feels just as bad as designing cars around those who want to use the cell phone while driving instead of actual drivers (...damn again, that might be the whole thing around the self-driving thing of these days). I feel that democratization (an ideal one), if it is ever going to happen, will be good in this respect. As soon as experts (yeah, at least I feel like that) already prefer the raw, no GUI, OpenFOAM to your product, how can even less educated users pick it up? Ironically, the CAD part is the one most suitable to be scripted or automated, but vendors decided to invest where full automation could be near to impossible and trash their GUI with CAD-like bells and whistle (otherwise, how can you still sell at 10k$ or more?). Democratization is definitely requiring another switch, even, eventually, removing CAD from the picture (in the form we are used to it today, at least). Maybe AR will manage to do this (eventually, Magic Leap, or someone else, is going to deliver something great). Going back to point 2, some of you pointed out that hardware might require a democratization as well. I don't feel this has ever been a solid issue, in the sense that, until today, CFD is still used by engineers which, honestly, should have several things going on on some heavy machine, not just CFD. Where feasible (and most of the times it isn't), however, Cloud is probably going to help. What I feel as most dangerous about the OpenSource factor, however, is that it isn't really making anything more democratic. Actually, it managed to do something really opposite just because of the missing price tag. Today, a lot of people just use them without actually knowing even the rough details, not to mention those who actually just package them and sell. I'm not saying this is bad, but it is probably taking even more knowledge out of the picture. An expert can use it and, in case, fix it. A non expert who can just use it, will likely use it like a black box, just without the "safety" that a commercial product provides, or at least should. This brings me to one of the other factors, which besides Lattice-Boltzmann and immersed boundaries, seems to affect most of the new players of the market in general: where, in the hell, are your verification cases and/or theory manual? Can you actually prove that what you're selling is what you're marketing and that it has engineering value (if I can't give an error bound to my computation, honestly, I don't know what to do with it)? This might also make sense from the marketing point of view (nobody reads anything anymore, why caring about a manual?). But the damn thing still has to accomplish one very specific task. Why the hell should i prefer it over real time video games already priced very low if verification is not anymore part of the picture? |
||
June 13, 2017, 11:48 |
|
#15 |
Senior Member
Filippo Maria Denaro
Join Date: Jul 2010
Posts: 6,839
Rep Power: 73 |
I don't know what can be a future for CFD. But that seems a market strategy to increase selling codes, no democratization. CFD like videogames: plug & play?... Boh...
Would people like the idea of democratization in medicine? I mean some software that give you a diagnosis and a care whatever it is your knowledge? Maybe we are thinking about a future where hardware and softwares are our substitute in working and thinking? Honestly, I am puzzled too while reading many questions in this forum where people with no background in CFD (and FD...) try to solve complex problems. Are we able to democratize the FD before adding the "C"? |
|
Thread Tools | Search this Thread |
Display Modes | |
|
|
Similar Threads | ||||
Thread | Thread Starter | Forum | Replies | Last Post |
Future of CFD article-Have you read this? | Nail | Main CFD Forum | 1 | May 11, 2008 12:30 |
Trends of CFD | A.M.Yang | Main CFD Forum | 12 | November 11, 2001 12:19 |
CFD - Trends and Perspectives | Jonas Larsson | Main CFD Forum | 16 | August 7, 1998 16:27 |