Flash-crashing Wall Street is so much more fun now than in the old days, when you had to cause an actual panic over a number of days to cause a crash. And if you're a hedge fund that wants to do super-fast trades to get ahead of the about-to-crash competition, Appro International has just the server for you. It's called the HF1 …
Something that might play Crysis
However it's more likely to cause one.
Geez, you can get to 5GHz with liquid nitrogen.
The key in low latency trading is exactly that - low latency. Approaches like this at most provide a minimal incremental decrease in latency. There is no "quantum leap" here. If you look into what is causing the latency in the first place and fix that you may get a much better bang for your buck.
An Intel system running Linux has plenty of places to be optimised, same goes for the network. Every and each of these can yield more than you get from overclocking your trading engine.
They're probably doing everything they can think of. This article is just about the server.
one has to wonder...
what sort of code they are actually running. Anyone has any idea? 3.33 -> 4.4 gigs versus having more cores...
Surely they'd want either PCIe SSD or at least SATA SSDrather than a feeble 7200rpm spinny disk....
I like the ports on the back
Never seen a server with 7.1 audio on the back. I guess they couldn't get supermicro to cook them up a board with overclocking options, so went with something COTS from Asus.
That'll work well
No graphics card...
^ Says it all.
Surely these finance types should be spending their ill gotten gains on optimising the crap unthreaded algorithms, or porting them into CUDA.... rather than on overclocked boxes?
Why don't they get some old P4s?
This was the one thing that the NetBurst architecture was actually good at. They had 3.8GHz processors in 2004, and they would overclock very nicely.
They don't go for old P4s, because the actual number of instructions executed per clock cycle is far lower than for the Nehalem micro architecture.
The goal is really the highest single threaded performance possible, rather than the highest possible clock speed.
> For many of their algorithms, having extra cores does not help their applications run faster — having faster clocks does.
Like 99% of apps. The multi-core con finally exposed.