Vladislav M. • This website says: ["Too many tools, steps and PLD chips to choose from" says FPGA and CPLD users]

While I do find cloud computing interesting, and it's not like anyone asked me for my opinion but the above statement is one of the dummiest things I have ever heard. Should be taken off the website and replaced with something that would make it more appealing.

Manoj V. • Why would anyone want to run freeware simulation tools on 1/2 of a server ? Running Altera and Xilinx tools on their servers probably violates the licenses for said tools.

While this business model is interesting, it will take a bigger player like Mentor, Cadence, or Synopsys to pull it off but why would they sell their wares on a pay-per-use model? They make more money making you pay for yearly licenses.

Vladislav M. • I think they may have licenses of their own or most likely, cracked tools :o)

Regardless, I believe that most of this is a waste of effort. I could be wrong, but as an experienced FPGA design engineer, I use scripting stuff for the design automation. I also setup things when I migrate projects from one product to another, and still, I polish every single synthesis option when I start the new project.

I see no reason to use such tools in my design process, although I see this as a noble idea. I also dont see these tools appealing to experienced engineers who have certain work habits and ways to do things. These may appeal to some brain-less managers, who usually follow the salesperson's talk and ignore what their engineers are telling them. Sorry if I was too direct but this is what I think.

Will M. • It's got some interesting concepts -- but as Manoj states, it will take a big iron player to make it really interesting for most projects --

The Web based simulator is too limited for large jobs as is -- it really needs the capabilities of Modelsim/or similar to deal with a large or even moderate project --

Vladislav M. • Not to mention security and integrity of what you send ovre the web lol

Jeremy R. • I think many of the technical/business objections people have with cloud based EDA will be overcome with time, and convenience the way Online Banking, Online Travel, Google Apps and that sort of thing has gained acceptance. There are some aspects of cloud-based tools that really make sense. And of course all the big EDA users have their own internal clouds / server farms, but I don't think these giant cos. are the initial market for EDA cloud tools. I think it's the small resource constrained companies.

At PDTi we offer a register management tool as a web-app in both onsite and online versions. Some companies are using the online version and like the way we maintain the tool for them, saving a lot of setup, upgrade, configuration and maintenance work. Again, this is especially useful for smaller, resource constrained cos.

Saar D. • There's always a tradeoff between usability/functionality and confidentiality (and authenticity!) or information, and as systems are built, it is often the case that one of them has to give way to the other. But overall I agree with Jeremy on this -- companies will evaluate the security vs utility (correctly or not) and some will conclude that computing on the cloud is worth their while. Some never will, of course.

At boldport (http://www.boldport.com) we decided to start with a build generator "as a service", but let the users compile the designs on their own machines. Farming out jobs to the cloud is an obvious next step, but that doesn't necessarily have to go through boldport's servers. Again, it's a tradeoff.

Vladislav M. • Can you imagine how many managers or even engineers will fall into sending you the code just so that your application? I am not talking even about defence sector where things are ITAR(US) or SLSC(Canada). Apart from that, the improvement such tools pride is so small that sending you an IP that may be of a huge value is not something they would fall into. Imagine this: if you have his cloud computer cool stuff, you are step away form opening your own design services and selling IP cores :o)))))

Evgeni S. • The biggest advantage of such a service is scalability. Especially in a small company that doesn't have server farm but occasionally needs to run multiple concurrent builds that exceed capacity. Security is an issue. Ideally, FPGA design tools should provide end-to-end encryption, and they will if there is enough "market pressure". Right now, designers can do synthesis locally and send a netlist to the cloud. Although a lot of design information can be gleaned from the netlist as well.

Phil S. • IP security can be addressed via the deployment of the encryption technology that the FPGA vendors use on their own IP. Certain segments will not be satisfied with this, such as Defense, however others may. The benefit that I see is the ability to compile multiple jobs in parallel, on high end machines, to hit timing closure faster for challenging designs. My question is how much would FPGA designers be willing to pay for this service?

Anil K. • The business model and monetization model for cloud based EDA is still a challenge. The non-eda SAAS model vis a vie Salesforce etc has taken hold because of the economies of scale: something that will not happen in the EDA space. As Jeremy rightly points out, a viable ,per use model will no doubt be eventually flushed out. Especially for services such as emulators that are both expensive and are are on target to obsolescence: these can be replaced with cloud based FPGA based prototyping in a very cost effective manner.

Security aspects can be addressed as Phil points out. More mission critical information at rest and in motion is on the cloud then ever before.

On the technical front: I have worked on build infrastructure for offloading compute-cycle intensive Quartus runs and there is value to be had: just a random seeding strategy with a good shell script buys you more , owing to the heuristic (VPR algorithm like) nature of fitting/P

- -. • It's no new idea to make FPGA as a service but yet not implemented, may be this startup http://www.plunify.com will do. The pricing can be compared to amazon web services for instance they provide HPC servers with optional GPU on the board. How mach? It depends on the whole project budget, designer should be free :)

HarnHua N. • Just for the record, we (Plunify) aren't running and do not plan to run/use cracked tools/IP : )

We focus instead on providing on-demand scalability to engineers - what Evgeni, Anil and Phil mentioned - as the immediate value for a cloud-based platform like ours.

As many have mentioned, security can be addressed by both technical and business initiatives, and it will be hard to satisfy certain segments like defense at first. For example, on top of encryption in vendors' tools, we also provide an additional layer of user authentication and encryption via symmetric and asymmetric key technologies. If people are willing to bank online, designing online won't be far away.

That said, a web-based design flow would be hard-pressed to give experienced engineers the control they need - not impossible, but it takes time. That's why we're also building an API so users who are more command-line-inclined can use us from their existing scripts.

How much is it worth to get results for 10 compilations in, say an hour's time compared to getting the same results sequentially in 100 hours' time? The utility question seems to garner different opinions depending on the end-user's resources, comfort level with the cloud, etc.

Great to read everyone's comments here. We're a young company and are talking to as many FPGA users - from beginners to experts, GUI-based or commandline users, as possible to find what value and concerns a cloud-based service might be associated with.

Looking forward to hearing more of your thoughts! We'll be at DAC in San Diego this week so if you'd like to chat, just let me know.

Richard P. • Is the premise behind the idea of an FPGA based cloud intended for large scale design synthesis, or verification through emulation/simulation? The cloud of course offers such a high level of abstraction that either would be possible.

HarnHua N. • The lazy answer is both ; ) because the potential to add value to the FPGA workflow using cloud computing is enormous. Especially during crunch periods where a surge of compute resources are needed, a cloud provides that kind of on-demand scalability.

Peter V. • HarnHua,
I do have a problem with long compilation time on some projects - so I like your solution. But I have been testing with several compilation setups, and have found that the amount of parallization the compiler uses is not in that degree that it counts in compilation time. If you want a fast compilation time, you need a high processor speed and a lot of cache - the difference between the number of cores is neglectable. For things like SmartXplorer, multiprocessing does help. For a normal compilation, my tests show the difference is neglectable.
Can you please show numbers on the speed gain you get per single compilation, compiling with 1 computer, 2 computers, 3, 4, 5?
How many computers does it take to go from 10 hour to 1 hour single compilation time?
Please real test results.

Will M. • I think in some ways the tools have to catch up to the concept -- Many of the tools are not yet fully optimized for Multi-core/Multi-processor methods

Phil S. • The Quartus II software can take advantage of multiple processors/cores on a single workstation to reduce compile time. 4 processors will reduce the compile time by an average of 23%. There is a lot of data sharing from memory, hence the restriction to a single workstation.
If you partition the design to take advantage of incremental compilation, you can then get much bigger gains and compile the individual partitions on separate workstations, using a final top level compile to link the partitions. This will get you closer to the 10 to 1 compile time reduction, although I doubt that you will hit the 10:1 ratio. The challenge is partitioning the design such that you do not lose performance, as this approach limits cross-boundary optimization from both synthesis and fitting. Another benenfit of this approach can also limit the impact of small design changes.........we're now getting into another discussion beyond the "Cloud".

Will M. • The big deal breaker for me at the moment is the GTKWAVE viewer which will bog down with 4GB+ Waveform files frequent in DO-254 checks of even simple state-machines to all combinations / all permutations for all inputs -- Modelsim, and Logic Analyzer's on real H/W seem to be able to withstand that "Acid Test" -- plus the coverage reporting is nice on Modelsim

Anil K. • @Peter..indeed my findings with the processing tool. Parallelizing heuristic/probabilistic algorithms is not too feasible. As you mention, in Tools like Quartus the full parallelism is never achieved with 1.23-1.5processors used on average on the projects that I have worked on.
History repeats itself: the bottom-up approach (partition-stitch) that Phil alludes to, was the team synthesis approach for asic projects in the early days, for larger chips. In the development phase, with the code base in flux, this is not as optimal, but is a potentially feasible approach on a code freeze/release branch. This fork-partitions-compile-join at top-delta compile approach can be scripted and parallelized using gridengine or some other job submission engine or via a custom fork-join algo riding on top of an MPI layer: a potential cloud based offering-likely not a TAM to justify the effort by those other than Xilinx/Altera?.