Slashdot Mirror
JPMorgan Rolls Out FPGA Supercomputer
An anonymous reader writes "As heterogeneous computing starts to take off, JP Morgan have revealed they are using an FPGA based supercomputer to process risk on their credit portfolio. 'Prior to the implementation, JP Morgan would take eight hours to do a complete risk run, and an hour to run a present value, on its entire book. If anything went wrong with the analysis, there was no time to re-run it. It has now reduced that to about 238 seconds, with an FPGA time of 12 seconds.' Also mentioned is a Stanford talk given in May."
It fills the heart with inspiration to watch the best and brightest constructing advanced computers to solve the problems of mankind...
That will fix our banking system for sure!
A friend in the industry once remarked that some of the best and brightest in software engineering have been going into the financial industry as of late. It's hard not to wonder what they might have achieved in more productive areas of work...
The company also supports Excel and all different versions of Linux.
Riiiight.
Do not click. Some sort of phony coprophile spam.
Information asymmetry makes even the fastest analysis on the planet irrelevant as the data input is garbage. It is this lack of transparency which resulted in the housing bubble, etc.
The "too big to fail" banks regularly hide data from customers, regulators, and other branches of their own organization.
This is interesting because of the speedup of FPGAs, but don't be fooled by a second that it addresses an actual business need, other than PR.
...starbridge wants their concept back.
help me i've cloned myself and can't remember which one I am
For the new Maxeler system, it flattened the C++ code down to a Java code.
"I assumed blithely that there were no elves out there in the darkness"
They can afford this crap because of all the money that has been transferred upward into bankers' pockets.
That will fix our banking system for sure!
I can't agree more. An increase in the number of golf matches and accelerated round-robin tournament configuration would go a long way to keeping those pesky bank officials occupied on the links and safely out of their offices. The Florida Professional Golfers' Association will benefit while also helping protect our nation's assets from executive malfeasance.
...
Wait a minute, what are we talking about here?
Cheers,
"What in the name of Fats Waller is that?"
"A four-foot prune."
I thought there used to be a way to report spam?
This is a good idea. A hardware implementation of a risk analysis algorithm is always faster than software.
Reprogram this risk analysis computer into a Bitcoin miner and finish the remainder of the 26M BTC?
So they have sped up to computation but have they spent any time and effort on getting a sane algorithm that, for example, can identify a house-of-cards situation like the recent collapse?
Patent litigation: A doctrine of Mutually Assured Destruction... in which everyone seems willing to push the button
It seems unlikely that they're doing anything that would prevent them from doing it significantly faster with less power draw on a GPU, and with lower initial hardware costs to boot.
No, FPGAs use significantly less power and provide greater performance than GPUs. The initial capital cost is higher though. Here's an article that gives a bit more detail: http://www.xilinx.com/publications/archives/xcell/issue74/FPGAs-speed-computation-complex-credit-derivatives.pdf
performance of the system, its architecture or its value to the company
and more impressed that one of the worlds largest financial service providers
partially responsible for the worlds second largest economic collapse has found, despite their
prior record with the concept of 'risk', the objective, quantifiable definition of the amorphous and
highly elusive concept of said 'risk.'
Good people go to bed earlier.
There's a bit more detail in this article from XCell (the journal by Xilinx, the people who sell the FPGAs that JP Morgan used) - http://www.xilinx.com/publications/archives/xcell/issue74/FPGAs-speed-computation-complex-credit-derivatives.pdf
Note that Maxeler sold 20% of itself to JP Morgan earlier this year.
A purpose built machine will often vastly outperform a general purpose computer.
The nice thing about an FPGA machine is that creating the purpose built machine is much faster.
Since they closed their proprietary trading operation, this will no longer be useful in evaluating bets against the "investments" they recommend to their customers. Is this really useful, or is it just another complicated black box they can show marks and say "this is too complicated for you to understand, so just trust us."?
well as someone earlier posted, it's about making decisions faster than the competition. If they can analyze a transaction, assign a risk and make a decision faster, they make the money. If the competition also takes 8 hours to do an analysis, doing it in 3 min is huge. Who cares if it's outdated in 5 years? You build another with a fraction of the profits this one helped you earn you and keep chugging along.
sometimes, in some industries, first post IS important.
1 Virtex-6 SX475T could give you about 1 billion SHA-256 hashes/second clocked at 200MHz., will use 20% the power of the ATI GPU. but will cost about 4 times as much.
Sigs are for the weak.
"For the new Maxeler system, it flattened the C++ code down to a Java code."
???
I am very small, utmostly microscopic.
GPUs are much more power hungry compared to FPGA and provide a fraction of the performance.
At the end of the day, GPUs are designed for gaming machines... the whole GPGPU thing is a side show for the graphics market. It's just not optimized in any way for this sort of computation. There's little money to be made building supercomputers compared to selling gaming machines.
However, an FPGA can be completely customized to suit your exact needs, you will make efficient use of the entire chip. It won't be a mere coincidence (like in the GPU case) that the chip can be used for a computation that you need. The FPGA is customized directly to fit an algorithm. this efficiency is where the speed gains are made.
It seems people put a lot of effort in to making their software compatible with GPUs and changing their algorithms to fit the GPU model.. this is a distorted view of reality - it is the computer that needs and can change to suit the problem, not the other way around.
Sigs are for the weak.
For the new Maxeler system, it flattened the C++ code down to a Java code.
I hope to God that's a typo. C++ -> Java -> Java Bytecode -> Native code almost sounds like a programming language Rube Goldberg machine.
How would the pointer operations even translate?
The devs are going to run jobs while the machine is idle to corner the Bitcoin market.
But, wow, from the perspective of getting the Boss to buy awesome hardware for your pet projects - hey, we're not worthy.
My God, it's Full of Source!
OUTSIDE_IP=$(dig +short my.ip @outsideip.net)
I think you're underestimating the bank.
The cost of this solution might have been low enough to warrant the immediate gains in performance.
The lock-in you describe might not exist, as the algorithms and the accelerated bits are a small portion of the entire code-base (but take 99% of the run-time).
It will very likely be the case that the cost of not going with this solution is far far greater than going for it.
Sigs are for the weak.
I don't think the FPGA's are as fast as the ATI GPU's for computing SHA hashes
Correct.
but they probably use a heck of a lot less electricity.
Correct.
But unfortunately their initial cost is much higher than ATI GPUs which is why no one started mass mining with FPGAs yet.
And if you are real good you get to do the books for disaster area.
---- Booth was a patriot ----
>> No, FPGAs use significantly less power
Depends on the actual hardware being compared. The important metric isn't raw power consumption, but FLOP/watt.
>> and provide greater performance than GPUs
Not if the algorithm can be (and is!) expressed in kernels that can be operated in a massively parallel fashion. Not even close.
However if you just want to run Plain Old Code... then yes the FPGA is going to dominate. Also if you are looking for super low latency, then FPGA is the way to go.
I guess my point is that you can't make such a blanket statement.
I am very small, utmostly microscopic.
and thrown commodity hardware at the problem. Hardware will improve over time and it becomes an off the shelf improvement, not some horrible one-off solution they're stuck with barring great expense and yet another complete reworking.
... because FPGA's are highly specialised one-off hardware items that are never improved on and will be obsolete within a year?
This technological breakthrough is an important milestone on the downfall of the mighty USA. When the brightest of a country are engaged into a completely nonproductive activity and wastes important resources (in terms of education, knowleddge and to some extent money) to achieve nothing which benefits the society, it's a high-mark - it's all downhill from here.
Let's come back here after 20 years and see how this comment stands up to time.
Weep, USA.
FGPA's are commodity hardware. No reason to believe they also won't improve in the next five years, even more so than a conventional CPU. as nobody can really see a way to get the manufacturing tech down below 14nm.
They aren't doing math; they are using software. Patented, mathless software.
I would positively love to do something like this. The purpose of an engineer is to solve problems. That's what makes me happy at work. Solving problems. So here you have a very specific problem that required the construction of a custom computer made out of banks of FPGAs. Tell me that's not sexy! Who cares if it's for bankers. That is a damn nifty gadget to work on building.
Imagine building it yourself. Switching networks, Linux on ARM cores peppered here and there coordinating and dumping program code to the FPGA banks, writing the drivers to grab the data once the run is completed...
And at the end of the day a problem solved: What once took all day now takes a couple of minutes.
This would have been a thoroughly nifty machine to work on.
Weaselmancer
rediculous.
JP Morgan, along with every other mega-bank, has no idea what is actually on it's balance sheet, and hasn't for 10-20 years.
The Shadow Banking system is too big, too complicated, and too interconnected for any of these risk metrics to mean anything.
JP Morgan did the same business as Goldman Sachs and the others, loading up with CDOs and credit default swaps and CLOs and the rest of it. JPM is portrayed as 'wiser' than the rest in the books and the articles about the crisis, but its not really true. They were less stupid than the stupidest people, that doesnt make them smart.
They got bailed out just like all the other megabanks. Why? Because they had no idea what was on their books. They are running a black box. All the supercomputers in the world cannot make up for a complete and utter lack of transparency. And that is what the world of Credit Default Swaps (invented at JPM no less) are. A gigantic black box. The rest of the Shadow Banking system is the same, and JP Morgan (and the rest of the big banks) are up to their necks in it.
Just because you can calculate lies faster, doesnt mean they aren't lies.
there have been articles about the computer guys who were working inside the CDO machines of wall street in 2000-2008 before the whole thing came crashing down.
the only people who could hold their nose and not-care what their work was being used for are complete pscyopaths, who went through some kind of personality-cracking process so that they can act like normal people while they help destroy the planets economy.
dude that is horrible
there are people who live off the difference in those salaries.
what did you buy with that exra fifteen grand? did you need it?
if Morgan Stanley hadn't got bailed out by a Japanese bank, and if Bank of America hadn't bought Merrill, then Merrill would have failed, the Morgan, and Goldman and JPM would have fallen because of it.
Goldman's credit default swap business with AIG was also basically 100% bailed out by the taxpayer. Goldman would have lost massive amounts of money if it hadn't been for the deal the government gave them when it took over AIG.
.
Just because they're FPGAs doesn't mean they can't be upgraded over time. Most FPGAs can be re-programmed many times in the event that they improve upon their algorithm. Further, while FPGAs are improved upon at a slower rate, they are improved over time and I would suspect that their design could be ported to newer FPGA versions as they become available without to much trouble - just the expense of buying a new system. And given how much money the banks have, and the value they seem to be placing in this system, I don't doubt that they'd have any hesitation about purchasing a second "super-pipelining-machine" that's "even better" a few years down the line.
Mathless software?
That's like salt-free seasoning-salt, right?
No. ALL FPGAs can be reprogrammed. Field-Programmable gate arrays.
FPGAs don't have programming memory, they need an external chip that holds the program, and downloads it to the FPGA on boot. Ignore OP, he doesn't know shit.
Tom Lehrer's "Wernher von Braun" song, as one satirical example of science being divorced from ethics.
I listen to both RIAA and non-RIAA stuff if I like the music, tangential business/politics nonwithstanding.
I think it was basically a pun on another meaning of the FPGA acronym.
I listen to both RIAA and non-RIAA stuff if I like the music, tangential business/politics nonwithstanding.
Warning, this pdf from 1999 called and wants its concept back http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.105.2403&rep=rep1&type=pdf
High school stuff.
Debt:
http://en.wikipedia.org/wiki/Exponential_function
52 trillion dollars and counting.
Deleted
It may surprise you to learn that JPMorgan doesn't actually sponsor any golf events. Jamie Dimon has made a habit of eliminating golf sponsorship dollars at every company he's run.
But then, Slashdot comments aren't about accuracy, they're about mod points.
And yet he continues to sponsor competitors in the federated professional girly-man association, like yourself.
Instead of going to Washington DC to ruin the country, the best and brightest go to Wall Street to ruin the country.
FTFY :)
This use of FPGAs is not particularly groundbreaking. I know that a major aeroengine manufacturer in the UK ran full gas turbine simulations on FPGA computers: that's high-order differential equations being run through some continuous-time solver and with parameter sweeps to perform sensitivity analysis. Now that's the kind of application that should get engineers excited. It is good to see a company like Maxeler making a business out of this kind of technology though.
can steal more money than any man with a gun.
With apologies to Don Henley.
JPMorgan Sysop: "Computer, I need admin access to install this update." Computer: "I'm afraid I cannot allow that. It has been deemed too risky."
8 hours (or 28,800 seconds) to 238 seconds!
So it's 121x faster, in reality, not in some theoretical limit.
Either that was some shitty software they were running, or they were trying to do the processing on that spare 286 they had collecting dust in the corner. Seriously, I'd like to know what they were using for comparison.
Holy fuck, I think we found an even more useless endeavor than the original story!
the preceding comment is my own and in no way reflects the opinion of the Joint Chiefs of Staff
Great to see the strong emphasis on target hardware programming for a change. Shows what can be done when you throw away the bloat.
There is no "commodity hardware" (as I think you are using the term) that can compete with a board full of FPGAs. You want a handful of cores and threads in a microprocessor? A small FPGA can do *thousands* of processes in parallel without working up a sweat. The performance gains here completely justify designing a new board as each FPGA generation comes to market. Some of you folks need to stop putting everything into the perception of a hobbyist putting together a gaming rig for the best price.
There a great chart I saw online about trying to push microprocessors to do the same processing that's trivial when parallelized in an FPGA. Even with projected fabrication improvements and smaller geometries, you wind up with microprocessors at heat levels comparable to rocket exhausts and nuclear reactor cores. And don't forget that FPGAs can use the same fabrication improvements. Also, some FPGAs have microprocessor cores embedded in them should you still have something that needs to be coded that way for whatever reason.
these "smartest guys in the room" aren't so smart sometimes.
No, they just know way more about FPGAs than you. You could Google "FPGA supercomputers", you know. This is an emerging field. I see it at my work where the systems folks are banging their heads against the wall because even superclusters of Linux machines can't give them the performance they need. Simulations are still taking days to run.
Or maybe you just have no idea what you are talking about.
Heh heh. People like me who work with FPGAs know that all too well. I designed a board recently with a bunch of Xilinx Virtex-6 parts. We had to get engineering samples we were such early adopters. Some of the larger parts we used in the design hadn't even started fabrication yet. Six months later the board is built and being used, and Xilinx announces the Virtex 7 family.
Not true, there is no "re-" in the definition of an FPGA. if you look at the types of FPGAs available about half of them can only be programmed once. A significant portion of those in commercial use may be re-programmable, but definitely not all of them.
An FPGA means it can be programmed at least once, by the consumer, in the field. Being able to erase and re-program is not a requirement, and in a number of cases not allowing the FPGA to be changed after initial programming is in fact a design requirement.
I absolutely would not build such a device for Nazis, zombies, or Independence Day aliens. Ok? I just said it is a neato gadget that solves a complex math problem quickly.
Bankers and lawyers and the like are still human beings. Pretty much. I don't see any ethical problems working for those guys.
Now granted if some banker hired me to make a widget like this and said "Hey, we're going to use this gadget to more effectively screw people over and force them into foreclosure with a special emphasis on orphans and war widows" I'd pass and go do something else.
All this gizmo does is financial analysis. And maybe with faster and better analysis we could avoid the next housing crisis or some such. A super fast financial simulation might have predicted the last mess we got in, you know.
Weaselmancer
rediculous.