With Intel in the process of buying venerable FPGA-maker (field programmable gate arrays) Altera and adding FPGA-like customisability to some Xeon silicon, the industry has been anticipating a response from AMD. And perhaps the first fruit of that response are now emerging.
A report from Italy's Bits and Chips says the company wants to stack its x86 Opteron chips with FPGAs, with its HBM2 (high bandwidth memory) technology providing the high-speed communications.
The Italian report is based on this patent application filed by AMD, titled Die-stacked memory device with reconfigurable logic.
In the patent filing, AMD says it wants to give designers the reconfigurability of FPGAs, but without the on-board real estate penalty of having to design in an extra chip.
As the patent application states, tight integration between the FPGA, the CPU and memory dies would mean “the reconfigurable logic device can rapidly access the stacked memory dies for store or load purposes, and thus perform related data manipulation operations with higher bandwidth and lower latency and power consumption compared to the performance of the same data manipulation operations by devices external to the die-stacked memory device”.
The decades-old FPGA is attracting attention in the data centre, and not only from Intel. Microsoft last year described how it's using the devices in Project Catapult to boost Bing searches. Facebook is also flirting with the technology.
Once regarded as a niche technology, FPGAs came to more mainstream awareness as the technology of choice for serious Bitcoin miners. ®