3 ms·
Now, this gets interesting. FPGA's usefulness as an accelerator in DCs was hugely hampered by IO bottlenecks. Even PCI-E can be said to be limiting the theoret
by baybal2 9y ago
Now, this gets interesting.
FPGA's usefulness as an accelerator in DCs was hugely hampered by IO bottlenecks. Even PCI-E can be said to be limiting the theoretical peak throughput of top FPGAs. With more IO being done by physical hardware, this should offset IO bottleneck somewhat.
I'm still waiting for FPGA on package Xeons that will have raw register to register linkage connection. They were once announced in 2016, some samples were shown, but then they went silent.
- throwaway2048 9y agokeep in mind that raw speed (in terms of say, ram access or PCI-E access (which is already faster than this implementation)) is rarely the obsticle these days, the issue is fetch/rtt latency, especially with trying to integrate FPGA acceleration into a codebase for offloading some operations, the penalty for for which can often be tens of thousands of cycles, or more.