FPGA based FFT - Could it work?

[airsquirrels]

Originally Posted by GP2

The f1.2xlarge instance type on AWS cloud is one Xilinx Virtex UltraScale+ VU9P, which has 1182K LUTs, which obviously is not quite as high-end as 2500K (which model has that many?). Current spot prices are about 50 cents an hour in Oregon or N. Virginia, which is $12/day (it was over $1 an hour a couple of months ago). But that's with no upfront cost, no additional charges for electricity, and four Xeon E5-2686 cores thrown in which could do LL testing or whatever else in parallel.

On the other hand, 4500W would be $10.80 a day for electricity assuming 10 cents / kWh, plus the upfront purchase costs for that array of GPUs plus supporting hardware.

Well you found one of my errors - I was glancing through the spec sheet and jotted down the marketing “logic cells” value rather than the LUTs for the VU9P. That certainly reduces the advantage a bit more. Interesting concept on the amazon instances though. If I created a bitstream for the VU9P it would certainly be easily deployable on those instances.

[drkirkby]

A field programmable gate array (FPGA), which are made by companies such as Xilinx, may offer a way to test exponents quickly. They would probably outperform a GPU. I wonder if anyone has ever tried this? There are some quite low-cost (under $50) FPGA development system around, so it would probably be fairly inexpensive to try. However, the low cost ones would be too slow and have too little memory to test large exponents.


I am an electronics engineer, but my skills do not lie in digital design - I'm an RF/microwave engineer. As such, this is probably not something I would want to do, but I can see it could form a project for those motivated to do.


What I see as an obstacle is that the low-cost units don't have the building blocks of higher cost units, so maybe time spent programming a low-cost device will be largely wasted if a higher cost device allows for a completely different implementation.



Anyway, I'm sure such a project could make a BSc/MSc project for a student.



Dave

[Batalov]

I am sure some Uni will welcome very much a generous donation that would allow funding such a BSc/MSc project.

Of course, you can try Verilog design in a garage. The software will cost more than the garage, though. The chips cost nothing - that's true.

[drkirkby]

Originally Posted by Batalov

I am sure some Uni will welcome very much a generous donation that would allow funding such a BSc/MSc project.

There must be plenty of academics on this site that could approach Xilinx about that. They would probably have a reasonable chance of success. When I worked in academia I got a decent(ish) DSP board from Texas Instruments.

Originally Posted by Batalov

Of course, you can try Verilog design in a garage. The software will cost more than the garage, though. The chips cost nothing - that's true.

I don't know what the software is like with the development kits, but it must be usable. There are open-source tools too.

I have very little knowledge of these things, but I think the use of an FPGA is at least worthy of consideration.

Dave

[LaurV]

I have tried in the past to make a 96-bits modular multiplier/squarer using a Spartan kit that my colleagues "discarded" after use, but I stopped because my VHDL knowledge sucks, and the learning curve needs more time than I can afford (or a better mind, haha). I chose 96 because of it divides nicely, and there are 16-, 32-, and especially 24-bits hardware multipliers in almost all of those FPGAs. The idea was to use some of those in a "karatsuba" fashion to multiply, do modular reduction (here I was on mined terrain, it needs a lot of experiments to get it right and fast), then expand the mulmod later into a powmod, and at the end, tackle the TF task. People around me who knows more about the subject tells me that such "enterprise" is not complicate, but the calculus we made would raise to about $8000 or more, to have a board that would do the equivalent work of a water-cooled 2080Ti (i.e. about 4.3THzDays/Day). So, we gave up. Unless you have the money to go for the ASICs, it doesn't really worth, and if you DO have the money, better make mining rigs...

[VBCurtis]

Originally Posted by drkirkby

I have very little knowledge of these things, but I think the use of an FPGA is at least worthy of consideration.
Dave

Straight from the Dilbert book of management, this. Doubly so when it's used a a reply to someone who *does* have knowledge of these things indicating that it has (a) been thought of before and (b) isn't cost-effective.

[drkirkby]

Originally Posted by VBCurtis

Straight from the Dilbert book of management, this. Doubly so when it's used a a reply to someone who *does* have knowledge of these things indicating that it has (a) been thought of before and (b) isn't cost-effective.

I think you must be in a time-warp, as I wrote the comment

"I have very little knowledge of these things, but I think the use of an FPGA is at least worthy of consideration."

about it worth trying about 1.5 hours before LaurV said he had tried a Spartan kit and concluded it is was not cost-effective.

[xilman]

Originally Posted by Batalov

I am sure some Uni will welcome very much a generous donation that would allow funding such a BSc/MSc project.

Of course, you can try Verilog design in a garage. The software will cost more than the garage, though. The chips cost nothing - that's true.

A few years ago I went on a one-day course in FPGA design. The software cost absolutely nothing and the hardware a few quid.

Designed a 9600baud UART (well a UAT as there wasn't time for the R portion) and tweaked a RISC-V design which ran a full Linux on said chip.

[drkirkby]

Originally Posted by xilman

A few years ago I went on a one-day course in FPGA design. The software cost absolutely nothing and the hardware a few quid.

Designed a 9600baud UART (well a UAT as there wasn't time for the R portion) and tweaked a RISC-V design which ran a full Linux on said chip.

I recall years ago someone telling me that writing code for FPGAs was a career in itself, or words to that effect. I forget the exact wording, but the implication was this was a specialised skill. So I was somewhat surprised some years later when I was at a medical physics lecture, where someone had used an FPGA is a small part in a much bigger project. He was a medical physicist.

I'm guessing that FPGAs are something where you can dip your foot into the water using cheap hardware and free software, but also a professional would be paid well, use expensive hardware & software.

[retina]

Digging ditches is also a career in itself. But amateurs can also partake in the fun if they wish.

If you want a fantastic ditch dug quickly, hire a professional. If you want to spend no money and lots of time farting about with a shovel to create a ditch that does its job only under a narrow range of special conditions then you can do it yourself.

[kriesel]

Originally Posted by drkirkby

I'm guessing that FPGAs are something where you can dip your foot into the water using cheap hardware and free software.

Go for it.
It took only seconds online to find:

https://www.intel.com/content/www/us.../overview.html

https://www.ebay.com/sch/i.html?_fro...board&_sacat=0