>>I wonder what they paid for it.
That is just like saying 'A Ferrari 360 consists 1100Kg of metal so I wonder what they paid for that 1000Kg of metal'.
FPGAs and GPUs can't be compared in that way.
FPGAs are slower in flops performance but that can do everything in a massive parallel setup if that is needed for the
data processing. If a certain processing step can be done with 2000 independent multiplications then you can just create such logic in FPGAs and perform that step in just a few clock cycles while the GPU must revert back to ordinary loops because no GPU has that many ALU's.
And don't have too much admiration of financial companies and their contractors.
They are generally not into this stuff. A lot of senior developers have tons of certificates but can't handle new technologies because they are so caught up in their fixed paradigms. One of the so-called senior architect engineers at a financial software company told that they turned of a brand new GPU cluster because no-one could get it to work.... Well, I guess that was because there was no Java/Microsoft certificate for GPU supercomputing yet...
They already exist but they are pretty expensive.
A single Intel's XD2000 inSocket FPGA accelerator costs more than a really high end user workstation.
I worked with one of such devices for a while and it is amazing what you can do (besides bitcoin mining *laugh*).
As long as what you want to calculate or process can be done in parallel, the speed ups are phenomenal.
With devices like the XD2000, the limit was the size of the FPGA but with the next generation FPGA's from Xilinx or Altera,
we will definitely see a big leap forward in certain areas.
When you can let go of the old paradigms then the sky is the limit.
More interesting however are hybrid devices like the new Zynq which is basically a (dualcore) CPU plus FPGA.
In this case you never have to leave the silicon to use customizable logic.
Some mainframes back 30 years ago already had reprogrammable CPU's where custom logic could be added but not to the extend of these device.
Slashdot Mirror
>>I wonder what they paid for it.
That is just like saying 'A Ferrari 360 consists 1100Kg of metal so I wonder what they paid for that 1000Kg of metal'.
FPGAs and GPUs can't be compared in that way.
FPGAs are slower in flops performance but that can do everything in a massive parallel setup if that is needed for the data processing. If a certain processing step can be done with 2000 independent multiplications then you can just create such logic in FPGAs and perform that step in just a few clock cycles while the GPU must revert back to ordinary loops because no GPU has that many ALU's.
And don't have too much admiration of financial companies and their contractors.
They are generally not into this stuff. A lot of senior developers have tons of certificates but can't handle new technologies because they are so caught up in their fixed paradigms. One of the so-called senior architect engineers at a financial software company told that they turned of a brand new GPU cluster because no-one could get it to work.... Well, I guess that was because there was no Java/Microsoft certificate for GPU supercomputing yet...
They already exist but they are pretty expensive. A single Intel's XD2000 inSocket FPGA accelerator costs more than a really high end user workstation. I worked with one of such devices for a while and it is amazing what you can do (besides bitcoin mining *laugh*). As long as what you want to calculate or process can be done in parallel, the speed ups are phenomenal. With devices like the XD2000, the limit was the size of the FPGA but with the next generation FPGA's from Xilinx or Altera, we will definitely see a big leap forward in certain areas. When you can let go of the old paradigms then the sky is the limit. More interesting however are hybrid devices like the new Zynq which is basically a (dualcore) CPU plus FPGA. In this case you never have to leave the silicon to use customizable logic. Some mainframes back 30 years ago already had reprogrammable CPU's where custom logic could be added but not to the extend of these device.