Author

Topic: BFL Single - Monte Carlo (Read 2534 times)

brand new
Activity: 0
Merit: 250
April 09, 2012, 06:02:50 AM
#15
Doing floating-point calculations on FPGAs is quite inefficient use of the resources. Therefore most of the numerical modeling software (including Monte Carlo methods) is not easily portable into FPGAs. In order to get sensible FPGA resource utilization you have to rewrite your software to use fixed-point or other specific hardware-efficient representation. This is nontrivial undertaking.
I'd love to know whether the end-result of such an undertaking would have monstrously superior (cost / time) performance *if* some rocket scientist actually managed to code up their valuations using FPGA-friendly ops.

This is only *my* view (and hence irrelevant since I'm not directing strategy for a significant player in the markets), but I reckon most of the asset valuation models commonly used in fixed income by institutionals could do with a damn hard look at due to the issues surrounding sovereign risk and the validity of ratings-agency data. Many algos aren't really geared up to model assets that have a negative real yield (which includes a large proportion of liquidity and short-term sov debt, traditionally used in equations as 'risk-free' but clearly not now). If you're operating a plain long-only bond fund then the risk universe has changed dramatically. Mandates binding fund ownership of debt considered 'low-risk' decades ago aren't going to have the same predictable returns variance given the crap going around these days - loads of opportunity for the hedge funds, but the level of political risk is a dangerous unknown. I'd be careful with pure algo trading unless you're big enough to be the manipulators in the first place Wink

Perhaps the new market paradigm (ZIRP, QE to infinity, etc.) could be modelled better in ways that *are* amenable to FPGA optimisation... just thinking out loud... for all I know, PIMCO are already doing this with their 'new normal'... Cheesy
newbie
Activity: 39
Merit: 0
April 09, 2012, 08:17:32 AM
#14
Probably overkill for a individual trader... monte carlo sims are usually used to revalue your derivative positions against a large number of statistically generated possible market variables, giving a set of positions that can be statistically analysed to work out the worst-case scenario for your positions at the 95% confidence level (or whatever your risk appetite is).

Well it's been a long time since I did credit risk Wink That was with a cluster of Pentium Pros - all you'd get with FPGAs (assuming you could program the bitstream to implement the valuation algo you particularly prefer) is the ability to run a large number of sim-sets (and hence statistically have a better view of your position risk), hopefully in real-time. After all, all you're doing is running your valuation algorithm for your derivatives against a large number of different possible market scenarios.

And that's the kicker. What's the valuation algo? Well, that depends both on the derivative you're trading, and any proprietary val that you've cooked up in Excel. Translating this Excel rocket-science valuation algo onto an FPGA bitstream would show *awesome* skillz if you understood both the trading, the product, the valuation algo AND the EE stuff needed to program FPGAs... I doubt that you'd be here asking questions like that if you *did* know it all Cheesy

I guess BFL could provide a basic implementation of Black-Scholes - but markets have moved on since that was an effective options val purely because BS assumes the concept of the 'risk-free rate' (as per many other older asset valuation models), and recent developments in the sovereign debt markets have made that concept somewhat nonsensical. Risk has popped up and 'assumptions' are the mother of all clusterfucks. There are other issues with it too but this isn't a hedge-fund forum.


You're right though - being able to program your custom valuation algo into an FPGA bitstream and use your ex-BTC-mining-farm to provide real-time high-resolution risk numbers would make derivs trading somewhat less casino-like, regardless of the name of the technology used Cheesy

(of course, I could be talking like a dinosaur since it's been 10 years since I did this stuff)

Very very informative, thanks! I didn't really think about it, but yeah FPGAs would suck for FLOPs. Probably a job better suited for CUDA/OpenCL on a GPU. Unfortunately their doesn't seen to be any stable BTC options markets (that i can buy into, anyways), so risk assessment isn't going to be terribly useful for me. I was trying to find a good algo for trending/forecasting mostly, but I'm pretty sure I'd still have to identify the right variables.

Anywho, i might restart this thread in another forum to keep the talk going on trading algos. Thanks again for all the input guys!
hero member
Activity: 504
Merit: 504
Decent Programmer to boot!
April 08, 2012, 03:27:25 PM
#13
Interested in what you do with this BUT I doubt you will get far.

The Monte Carlo and other bitstreams is just a pathetic excuse for them to promote this as "not only a BTC miner" ( which it blatantly is ) but that it can also do other things.


Them 'drivers'

http://www.butterflylabs.com/drivers/

Indeed. It seems they provide this to attract large customers for their BFL units but I think this will not appeal to many becuase the costs of "custom" solutions is often significant.

The broken links for the other drivers have been there since before they had products. The drivers seem to have been meant for previous products that don't exist.
hero member
Activity: 518
Merit: 500
April 08, 2012, 02:06:10 PM
#12
Interested in what you do with this BUT I doubt you will get far.

The Monte Carlo and other bitstreams is just a pathetic excuse for them to promote this as "not only a BTC miner" ( which it blatantly is ) but that it can also do other things.


Them 'drivers'

http://www.butterflylabs.com/drivers/

Indeed. It seems they provide this to attract large customers for their BFL units but I think this will not appeal to many becuase the costs of "custom" solutions is often significant.
hero member
Activity: 504
Merit: 504
Decent Programmer to boot!
April 08, 2012, 01:15:30 PM
#11
Interested in what you do with this BUT I doubt you will get far.

The Monte Carlo and other bitstreams is just a pathetic excuse for them to promote this as "not only a BTC miner" ( which it blatantly is ) but that it can also do other things.


Them 'drivers'

http://www.butterflylabs.com/drivers/
legendary
Activity: 2128
Merit: 1073
April 08, 2012, 01:08:17 PM
#10
Doing floating-point calculations on FPGAs is quite inefficient use of the resources. Therefore most of the numerical modeling software (including Monte Carlo methods) is not easily portable into FPGAs. In order to get sensible FPGA resource utilization you have to rewrite your software to use fixed-point or other specific hardware-efficient representation. This is nontrivial undertaking.
newbie
Activity: 39
Merit: 0
April 05, 2012, 08:34:31 AM
#9
Interested in what you do with this BUT I doubt you will get far.

The Monte Carlo and other bitstreams is just a pathetic excuse for them to promote this as "not only a BTC miner" ( which it blatantly is ) but that it can also do other things.

and... you are correct.

Just got a response from Sonny:

Quote
Hi Colin,

MonteCarlo calculation firmware are a custom function which is built on contract specific to the end use needed.

Kind regards,
Sonny K
BF Labs Inc.

the product specs (http://www.butterflylabs.com/product-details/) has it listed alongside the hashing... i considered emailing them back and verifying i dont have to pay for bitcoin support... but i dont want it to affect my shipping date Wink
hero member
Activity: 518
Merit: 500
April 05, 2012, 06:31:56 AM
#8
Interested in what you do with this BUT I doubt you will get far.

The Monte Carlo and other bitstreams is just a pathetic excuse for them to promote this as "not only a BTC miner" ( which it blatantly is ) but that it can also do other things.
newbie
Activity: 39
Merit: 0
April 04, 2012, 11:19:50 AM
#7
The BFL singles use an unknown chip with encrypted single purpose bitstream.  BFL indicated probing via JTAG wouldn't work which would indicate some other form of card level obfuscation/encryption/security.  Exactly how hard it is to bypass is unknown.  Obviously once someone gets a BFL Single there is significant time value.  Tinkering with it means no revenue so it isn't surprising that we haven't seen any users reporting details. 

So right now you can mine with it and nothing else unless either BFL provides programming information and/or someone does some reverse engineering/hacking.



Checked out the cgminer code... definitely an application-specific protocol. I'm really hoping they get back to me with some info. It looks like they're interested in providing more general purpose APIs, but I doubt they'll ever release an affordable SDK... looks like they're centered mostly around consulting.

I'm just hoping the monte carlo is abstract enough that I could hook it up to something worthwhile, I don't think I'll hold my breath as far as custom application goes...
hero member
Activity: 532
Merit: 500
April 04, 2012, 10:58:47 AM
#6
I just ordered my first BFL single Smiley I'm excited by the other real world applications this might have, while I am a programmer and could learn to do my own FPGA algos, the BFL Single comes with some sort of monte carlo application.

I'm hoping I could use this in my custom trading software. Anybody else think this could work? It would be awesome if I could still make use of this sucker if the difficulty skyrockets. Has anybody tried using monte carlo for forcasting? (This also might not be the right forum)
I'm not sure that it actually comes with it, have you contacted them about the specifics? It would need a firmware other than the usual bitcoin mining firmware.

probably easier to do on the icarus miners.
newbie
Activity: 39
Merit: 0
April 04, 2012, 10:45:29 AM
#5
I just ordered my first BFL single Smiley I'm excited by the other real world applications this might have, while I am a programmer and could learn to do my own FPGA algos, the BFL Single comes with some sort of monte carlo application.

I'm hoping I could use this in my custom trading software. Anybody else think this could work? It would be awesome if I could still make use of this sucker if the difficulty skyrockets. Has anybody tried using monte carlo for forcasting? (This also might not be the right forum)
I'm not sure that it actually comes with it, have you contacted them about the specifics? It would need a firmware other than the usual bitcoin mining firmware.


It says in the product specs that it comes with some application for it (probably so they can market it as something else, possibly for us encryption compliance).

(from the site)
Packet integrity verification
2 step SHA256 block hash
MonteCarlo Calculation support

But yea, Im really curious what that actually means... I'm hoping there is some sort of API, but it might also just be some lazy windows app. I'm gonna email Sonny from BFL and see if i can get any more info.

Assuming I don't just want to keep the thing mining all day when I get it Wink



I have used Monte Carlo simulation in my actuarial studies.  No real world applications as I don't work with investments (assuming that would have an application there).

I'm interested in this idea.  If the difficulty goes way up, it'd be nice to know I can use my BFL single for something other than a paper weight.  Smiley

Yeah, it'd be nice to try and use it for trending. And i don't need another expensive paperweight Wink I find myself with enough of those dealing with computers.
legendary
Activity: 2114
Merit: 1031
April 04, 2012, 10:42:41 AM
#4
I have used Monte Carlo simulation in my actuarial studies.  No real world applications as I don't work with investments (assuming that would have an application there).

I'm interested in this idea.  If the difficulty goes way up, it'd be nice to know I can use my BFL single for something other than a paper weight.  Smiley
donator
Activity: 1218
Merit: 1079
Gerald Davis
April 04, 2012, 10:34:13 AM
#3
The BFL singles use an unknown chip with encrypted single purpose bitstream.  BFL indicated probing via JTAG wouldn't work which would indicate some other form of card level obfuscation/encryption/security.  Exactly how hard it is to bypass is unknown.  Obviously once someone gets a BFL Single there is significant time value.  Tinkering with it means no revenue so it isn't surprising that we haven't seen any users reporting details. 

So right now you can mine with it and nothing else unless either BFL provides programming information and/or someone does some reverse engineering/hacking.

rjk
sr. member
Activity: 448
Merit: 250
1ngldh
April 04, 2012, 10:34:02 AM
#2
I just ordered my first BFL single Smiley I'm excited by the other real world applications this might have, while I am a programmer and could learn to do my own FPGA algos, the BFL Single comes with some sort of monte carlo application.

I'm hoping I could use this in my custom trading software. Anybody else think this could work? It would be awesome if I could still make use of this sucker if the difficulty skyrockets. Has anybody tried using monte carlo for forcasting? (This also might not be the right forum)
I'm not sure that it actually comes with it, have you contacted them about the specifics? It would need a firmware other than the usual bitcoin mining firmware.
newbie
Activity: 39
Merit: 0
April 04, 2012, 10:30:12 AM
#1
I just ordered my first BFL single Smiley I'm excited by the other real world applications this might have, while I am a programmer and could learn to do my own FPGA algos, the BFL Single comes with some sort of monte carlo application.

I'm hoping I could use this in my custom trading software. Anybody else think this could work? It would be awesome if I could still make use of this sucker if the difficulty skyrockets. Has anybody tried using monte carlo for forcasting? (This also might not be the right forum)
Jump to: