Maxeler sells embedded FPGA appliances, which is only slightly what HPC means. Yes, there are some CDO types who love their new Maxeler risk appliance, but that means nothing to the hundreds of thousands of people doing HPC research and programming, or the millions who use the resulting programs.
FPGAs once promised greater accessibility in the form of languages and libraries, but that was pretty much 10 years ago. Now all the focus is on big-idea platforms like AMD Fusion, which promise wide relevance, not a trivial niche market. Even if that niche, risk trading, is populated by still-rich idiots who should have been prosecuted for the damage they did to the world economy in 2008...
Fpga has been around for a long time, but never caught on for hpc. The reason? Too hard to program. I believe the gpgpu will get the same fate unless the vendors come up with a solid, simple and efficient new directive based programming model, that can propel dusty deck code towards more flops without too much hassle. Most people in hpc, are paid to do science. Porting exercises are a burden.
This might be true for the big top 500 cluster systems. But then for all the other guys I have met don't have the man power to do all the programing when you change from one architecture to the other. So this will be very welcomed by many customers I have talking to. Those guys are engineers/scientists. HPC and software coding is not their core business.
Maxeler sells embedded FPGA appliances, which is only slightly what HPC means. Yes, there are some CDO types who love their new Maxeler risk appliance, but that means nothing to the hundreds of thousands of people doing HPC research and programming, or the millions who use the resulting programs.
FPGAs once promised greater accessibility in the form of languages and libraries, but that was pretty much 10 years ago. Now all the focus is on big-idea platforms like AMD Fusion, which promise wide relevance, not a trivial niche market. Even if that niche, risk trading, is populated by still-rich idiots who should have been prosecuted for the damage they did to the world economy in 2008...
Fpga has been around for a long time, but never caught on for hpc. The reason? Too hard to program. I believe the gpgpu will get the same fate unless the vendors come up with a solid, simple and efficient new directive based programming model, that can propel dusty deck code towards more flops without too much hassle. Most people in hpc, are paid to do science. Porting exercises are a burden.
This might be true for the big top 500 cluster systems. But then for all the other guys I have met don't have the man power to do all the programing when you change from one architecture to the other. So this will be very welcomed by many customers I have talking to. Those guys are engineers/scientists. HPC and software coding is not their core business.