My first look at the Chemnitz University of Technology's “Coffee Table of Doom” entry into the recently concluded ISC’13 student cluster-building challenge really gave me pause for thought. Four very large workstations crammed to the brim with 16 accelerators (eight Intel Phi plus eight NVIDIA K20) was something to behold. Two …
While this press release makes some good points, there's a serious red flag raised when someone suggests overclocking to get mild improvements from their HPC kit. Overclocking is great for enthusiasts and playing around, but I really hope there's no data centre or laboratory relying on overclocked parts.
HFT and overclocking
High-frequency traders use servers with overclocked processors as one of many ways to gain an edge in trading. Granted, they don't depend on them for long term usage, considering that many vendors don't warranty such servers as long as non-overclocked, standard servers.
If I were doing HFT systems, I'd be better off using FPGA/ASIC systems or even GPUs. x86 is garbage for heavy loads (or anything else, we should be using RISC) and in fact most HFTs are now doing FPGA these days. Indeed, overclocking an x86 craptel is kinda like putting aviation fuel in a VW Beetle. Sure it might reach 200 km/h in 3 seconds, but the engine's going to blow up in 15 seconds!