I don't know how I managed to miss this thread all this months, heheh... anyways, int19h checking in. Atm my interests are mainly with top-level architectural design, eg mapping algorithms to array structures, augmenting existing processor core designs with custom instructions... simple noob stuff like that. I'd rather stay away from the low-level stuff, will leave that to you guys

[edit: clarified something]

This post has been edited by int19h: Apr 17 2006, 03:40 PM

...then you may get a kick out of this!

Detection right, not recognition? Detection is pretty simple... well I guess you can always make it as complicated as you like, but I've seen a pretty simple one in the past that worked well considering the level of effort. But algorithmic simplicity is good, you can have fun putting together an application specific system on an FPGA, or download the Cell BE simulation suite and implement the algorithm on that platform.

I did rudimentary face detection on an FPGA once, but it wasn't for a graded project. But face detection on a commodity RISC... I dunno... the compilers do a lot of the work for you, I doubt an implementation there would warrant high marks, there would have to be some novel algorithmic work if you ask me.

I would suggest you do this project only if you have a supervisor who can walk you through the pattern recognition algorithms. If you have to muck around in Matlab implementing and testing various algorithms you may not have enough time to really enjoy the real-time implementation side of things. The same is true for any project really, if you have your heart set on real-time implementation work.

Unless of course you're keen on digging into the algorithms themselves, in which case more power to you.

Good luck!

I find it interesting that you left out IBM, especially considering that Sony came close to cutting Nvidia out of the PS3 gig completely because there was the initial estimate that the Cell would have enough juice to do the graphics on it's own [sorry, I can't seem to find the citation now]... anyway it turned out not to be the case (for now), but heck even the thought must've been enough to make nVidia and ATi worry. And they do worry. So much so that they recognise (to their credit) that they can't just sit pretty; they're not just worried about each other anymore (in the same way that Intel and AMD got into that MHz war); they realise they both have bigger problems to deal with. The nature of computer systems is evolving, as it always will, with workload characteristics. At one time the CPU simply did processing, ALL the processing... then the GPU "revolution" happened. Now nVidia wants to do audio processing on GPUs (and there's this whole community doing all sorts of things with them), meanwhile if IBM has their way they'll put GPU functionality back into their CPU. Who knows what's next.

I think your presumption that we can't "catch up" to the Intels of the world is misplaced, because while we probably can't build a better Pentium 4 than Intel, would we really want to? As workload characteristics change architectures must evolve.

SoC is big. Media processing is big. Sensor networks is going to be big (good luck putting a Pentium into one of those, lol). When the whole triple-play bla bla takes off and people can have another 10k channels with nothing good to watch there will be another round of router upgrades to be done, things like pipelined memory etc will come into play then, as will specially designed line cards etc etc etc... and fantastic progress in FPGAs massively reduces the barrier to entry for startups with special skills and ideas.

It's a mighty big pie. We don't have to have all of it, but we need to have the foresight and balls to claim some of it when the time is right.

Sorry, my bad, I should have noticed the context

I see your point, but correct me if I'm wrong, besides IP design the rest of the job becomes much simpler if PLDs (especially FPGAs) are used right? I think that's good enough for Malaysian companies to contribute to strategic niche industries, and anyway FPGA growth atm is outpacing ASIC growth, according to Gartner:

(A bit outdated, I know, but Gartner charges USD10k for the latest report, so... erm... no thanks )

That's assuming you limit your definition of "CPU" to von Neumann architecture cache based processors that will sit in a PC to do general-purpose computing tasks... which certainly isn't the only kind of "computing" there is. FPGAs have beaten Pentiums for pattern analysis. Keith Underwood of the SDSC had shown 2 years ago that FPGAs can beat Pentiums for floating-point arithmetic (and that was before the Virtex 5, with it's beefed up multipliers). For general purpose and/or control-intensive irregular code FPGAs will lose out, but that's just one part of the whole picture.

My intention with my last few posts to this thread was to point out that there's a lot more opportunity out there, and limiting oneself to "competing with Intel" is likely to be both unproductive and unrewarding. But that's just my opinion

At the end of the day there's no accounting for personal desire, and there's no discounting personal will power, so if someone really wants to put "MalaysiaBoleh Inside"... I probably won't buy stock right away, but they will have my moral support, FWIW

[edit: added the word "irregular"]

This post has been edited by int19h: Jul 14 2006, 09:42 AM

... there's one more area where FPGA are used out of necessity: low volume systems. I've used production quality 4-socket ISDN line cards (meant for ISDN routers) that were powered by Xilinx FPGAs.

And no, I don't work for an FPGA vendor

good call...

Btw guys, first of all lemme just say I've been enjoying this discussion, it's nice to get different information, views and arguments! But now I should get back to work lol... if any new posts I will follow up tmr or day after k

First you've got to reach the mass market. I'm not proposing that FPGAs replace ASICs. As others have pointed out, you start small and grow. Utilising FPGAs allows you to get products into the market early, and that allows you to:
a) Tweak your product as time goes by - EXTREMELY important in networking applications where vendors often screw up their implementation of protocols, so your product has to compensate for the screw-ups of others if customers are to think highly of you.
b) Wait for the market to grow.
If b) doesn't happen, it's not so tragic because you haven't over-committed yourself in terms of NRE.

The counter-argument would be to use general purpose programmable processors in the above 2 situations. The problem with that is
a) getting a favourable power/performance ratio within a specified form factor (big heatsinks may be a no-no)
b) you lose out on the possibility to eventually seamlessly take it to structured ASIC.
An obvious compromise is to utilise softcores for the "contentious" parts of the application, and move some suitable and universally standardized parts to custom logic.

And anyway, if you serve a niche low-volume market, there are perks too... less support issues, and you get to charge a premium for your product. We paid USD5k for our E1 line card. Starbridge Systems charges USD100k to oil companies for a special-purpose 7-FPGA seismic processing accelerator. I'm not sure how much TimeLogic charges for their bioinformatics accelerators. None of these companies is public-listed, so I'm not sure how well they are doing, but these niche markets wouldn't even exist at all if it weren't for breakthroughs in FPGA performance. And for the record, in bioinformatics there was a company called Paracel that tried to do acceleration on ASICs... FWIW they lost out in the market; I won't say exactly why because of course I don't have all the facts and figures (again they were a private company), but I'm sure you can detect my bias

I'm not saying FPGAs pwn ASICs, I'm just saying that their recent performance gains opens up a lot of possibilities that previously simply weren't available. FPGAs may not end up in consumer's homes, but the market is bigger than that.

(Now we've gone a bit OT for this thread, but anyway...)
E1 PRI (a standard for carrying multiple ISDN channels) is still used by enterprise and carriers as the backbone for digital voice due to it's ability to provide better QoS guarantees. Clearly VoIP seeks to, and will eventually, change that, but at least up until recently ISDN still got (gets?) some love. I'm out of that industry now though, the last system I commisioned in Malaysia was 2 years ago, so consider this info 2 years old.

I should probably clarify at this point: when I spoke of "4-socket ISDN" in an earlier post, I actually meant 4-socket E1... I said ISDN instead of E1 because I find more people know about ISDN than E1.

KBU! I was there for 3.5 semesters, foundation then e&c for 3/4 of 1st year... wayyy back when we had to pay Rm60 for 1 month of internet access. Ahhh, the memories... :-)

I left KBU and went to Monash after that... from my experience, KBU's curriculum was much more practical and hands-on, Monash a bit too heavy on the theory side, but now Monash have changed their course structure dramatically to be more design-oriented. KBU had more project work. All the best with it!

Btw, I just have to ask... are you still expected to etch PCBs with ferric chloride? That was the kind of old-skool stupid but fun project that I enjoyed doing so much at KBU! I think that was year 1 sem 1... or maybe foundation...

That's good to know what skills are especially in demand? I did some quick-n-dirty ™ jobstreet searches to see what's out there:

pcb - 59 jobs
firmware - 27 jobs
fpga - 13 jobs
verilog - 9 jobs
vhdl - 9 jobs

Does that look about right?