11–20 of 83 posts
Anyone knows how much something like this costs? My guess is around 20K.
To give an idea of scale / endpoint, a top-tier HFT infrastructure costs on the order of $100k/day. There are probably only a half dozen total of those in the world, though.
I wonder what exactly the FPGA does, and how they can make it practical and flexible. Of course they could always just test with C and then compile that to HDL.
A month or two ago I spoke to someone working in these shops. Apparently, they're using the FPGAs to run many of the high-bandwidth algorithms, because going from NIC to motherboard to CPU and back again added too much latency. So they program the algorithms into the FPGA (which has its own Ethernet port) so it can receive, process, and send the data without ever hitting circuitry not dedicated to the task.
FPGAs canonical use case is still as very fast codecs. There are exceptions, but they're really only the very simplest of arbitrage strategies. And in order to make those work you need top-of-the-line network gear and engineers as well.
Earlier quoted context omitted.
This seems a bit surprising: what is the latency coming from the NIC to the motherboard ? I somehow doubt the latency due to distance within one piece of hardware matters (current CPU cannot do much more than a few 1000s cycles for the electricity to go around a few times in a motherboard, which is nothing to do any kind of advanced processing).
The problem is that a high-bandwidth link overruns the CPU with thousands of interrupts per second. It can paralyse the system, where the CPU is so busy handling I/O that it doesn't get around to actually scheduling processes. Disclaimer: I sometimes consult for HFTs but we don't use FPGAs (yet).
Don't things like real time linux or soft real time patches for linux help, though ?
Why is this interesting?
They're also very vague on the software architecture or framework "kit". They don't say which feeds they support or if they provide any domain-specific data structures/standard trading algorithms like VWAP (slicing an order up over the day to give you "better" fills). I do see mention of an optional Feed Parser reference design, which I can't ascertain much value from. I am guessing it is mostly a generic kit for writing on their specialized hardware, and not specific to trading. There will be a lot of pain, maybe impossible to take vendor or open source libraries for market feed handling and compile it down using their embedded libraries. You'll be stuck doing tons of TCP/IP work, fun for hacking, but relative to the latency issues in your market data/order routing, it won't help you profit. Most people buy these appliances with included feed support.
With a big budget and co-location, this may be useful in bootstrapping some really low-level processing of market data, although there are a variety of hardware and software messaging solutions that would be a good alternative in driving latency lower if you have some money, Solace (HW), Tibco, 29West, etc to name a few.
I wish there was a hobbyist solution out there, but sorry, my initial impression says this isn't it. I would like more information, but maybe there are better hobbyist efforts out there. Interactive Brokers API isn't bad, and here's a full-blown real open-source trading kit that has a similar business model to MySQL. http://www.marketcetera.com/ Note: I feel that there are a few issues with this as well in some of the technologies they chose (JMS, Eclipse RCP).
I am waiting for one of these shops to open up the opportunity to hobbyists/devs. I would love to try and create my own decision/trade algorithms but dont have the kind of money it would take to mitigate the cost effect of the trade volume (ie, at $8 per trade and making 1000 trades per day with an average return of 2% per day you would need to be working with more than 400K to even cover your tx cost). Strikes me th…
Why is this interesting?
It looks like you got a few downvotes because people assumed you're being snarky (maybe you are?) -- but I think that most people (even hackers) who work outside of the financial sector are going to wonder why this is on the front page. In short though, it's a 1U blade server with a full-size FPGA in it; implementing your algorithms in hardware makes them extremely fast (thus the mention of the C-to-HDL compiler), wh…
I am waiting for one of these shops to open up the opportunity to hobbyists/devs. I would love to try and create my own decision/trade algorithms but dont have the kind of money it would take to mitigate the cost effect of the trade volume (ie, at $8 per trade and making 1000 trades per day with an average return of 2% per day you would need to be working with more than 400K to even cover your tx cost). Strikes me th…
Set it up, trade 'virtually' (i.e., just simulate), prove your algos, get bankrolled by someone who is convinced of the algos worth. Quite straightforward. I'm highly skeptical of a hobbyist beating the competition though (maybe in micro niches, where one would add domain specific knowledge to get an edge - that's my pet idea on this topic :) )